Hydrodynamics; a study in logic, fact, and similitude.

Birkhoff, Garrett, 1911-1996.

New York, Dover Publications [1955, c1950]

[Link]

Public Domain, Google-digitized

[Link]

We have determined this work to be in the public domain,

meaning that it is not subject to copyright. Users are
free to copy, use, and redistribute the work in part or
in whole. It is possible that current copyright holders,
heirs or the estate of the authors of individual portions
of the work, such as illustrations or photographs, assert
copyrights over these portions. Depending on the nature
of subsequent use that is made, additional rights may
need to be obtained independently of anything we can
address. The digital images and OCR of this work were
produced by Google, Inc. (indicated by a watermark
on each page in the PageTurner). Google requests that
the images and OCR not be re-hosted, redistributed
or used commercially. The images are provided for
educational, scholarly, non-commercial purposes.
 Garrett Birkhoff
Mathematics t9SS
' -''•'

Sf1
HYDRODYNAMi

^_^
1

«h,
1
1

kAc
1

^ **
1

kXl
s
 HYDRODYNAMICS
A STUDY IN LOGIC, FACT,
AND SIMILITUDE
Common sense contradicted (courtesy of Ballistics Research Laboratories,
Aberdeen Proving Ground.)
HYDRODYNAMICS
A STUDY IN LOGIC, FACT,

AND SIMILITUDE

BY GARRETT BIRKHOFF

PUBLISHED WITH THE SPONSORSHIP OF THE TAFT FUND

OF THE UNIVERSITY OF CINCINNATI

DOVER PUBLICATIONS, INC. NEW YORK

Wath.-Econ.
Library

QA
9 II

/9SS

Copyright 1950 by Princeton University Press

This new Dover edition first published in 1955

is an unabridged and unaltered republication of
the first edition. It is published through
special arrangement with Princeton University Press.

Manufactured in the United States of America

mi u^-o•vn „

51-50/^

TO RUTH
 PREFACE
The present book is largely devoted to two special aspects
of fluid mechanics: the complicated logical relation between
theory and experiment, and applications of symmetry con
cepts. The latter constitute "group theory" in the mathe
matical sense.
The relation between theory and experiment is introduced,
in Chapter I, by numerous "paradoxes," in which plausible
reasoning has led to incorrect results. In Chapter II, this
relation is studied more closely in the special case of flows
with "free boundaries."
Chapter III is devoted to analysis of "modeling" and its
rational justification. Theory and practice are compared (or
contrasted!), and the origins of modeling in symmetry con
cepts are described as well, thus providing a transition between
the two main aspects of fluid mechanics studied in this book.
The rest of the book centers around applications of group-
theoretic ideas. These are shown in Chapter IV to motivate
a large proportion of the known exact solutions to problems
involving "compressible" and "viscous" flows, and (in Chap
ter V) to yield the classical force equations for a solid moving
through a perfect fluid.
Perhaps Chapters II, III, and V contain the largest pro
portion of original contributions. I know of no other treat
ment of "free boundary" theory which penetrates as far into
reality as the discussion of Chapter II. Again, Chapter III
provides the only study of modeling that I have seen, which
avoids the mathematical oversimplifications of "dimensional
analysis" (Bridgman's well-known book avoids the physical
oversimplifications), and at the same time gives a realistic
picture of actual laboratory practice. Chapter V deduces
much of the classical theory of "virtual mass" from the
modern geometrical theory of "homogeneous spaces," thereby
making possible very extensive generalizations.
ix
 PREFACE

Most of the material was originally presented in lecture form

at the University of Cincinnati, in January 1947. The book
owes much to a John Simon Guggenheim Memorial Fellow
ship, which made possible visits to many European research
laboratories, and gave me leisure to rewrite it. In addition,
Chapter I has been considerably clarified through discussions
with Prof. Kampe1 de Fenet, while Chapter III owes much to
stimulating talks with Prof. Percy Bridgman. Various other
improvements were suggested by Prof. John von Neumann
and Dr. Eduardo Zarantonello.
Even after the most careful attempts at simplification, the
material treated still involves a heterogeneous combination of
destructive logical criticism, detailed experimental fact, and
occasional mathematical sophistication, which may at first
seem unpalatable to many readers. I ask such readers to
reflect on the correctness of the somewhat unorthodox views
presented, in those cases of greatest interest to them. I hope
that after this reflection, they will become convinced of the
difficulty of presenting the true facts more simply, and of the
need for more studies such as the present one, to clarify the
foundations of fluid mechanics.
Garrett Birkhoff
Harvard University
December, 1949
 CONTENTS
Preface vii

Chapter I. Hydrodynamical Paradoxes 3

1. The Practical Value of Paradoxes
2. Their Mathematical and Philosophical Interest
3. Fundamental Approximations
4. Non-viscous Fluids
5. Mathematical Properties
6. The D'Alembert Paradox
7. The Reversibility Paradox
8. Paradoxes of Airfoil Theory
9. Discussion of Preceding Paradoxes
10. The Rayleigh Paradox
11. Von Neumann's "Triple Shock" Paradox
12. Kopal's Paradoxes of Linearized Projectile Theory
13. Viscous Incompressible Fluids
14. The Paradox of Turbulence in Pipes
15. Other Symmetry Paradoxes
16. The Eiffel Paradox
17. Stokes' Paradox
18. The Rising Bubble Paradox
19. "Drift"; Magnus Effect Paradox
20. A Role for the Mathematician
21. Mathematics as Reliable as "Common Sense"; Con
clusion

Chapter II. Recent Progress in Free Boundary

Theory 40
1. Wakes
2. Mathematical Model
3. Borda Mouthpiece
4. Further Generalizations
5. Indeterminacy of Separation Point
6. Physical Unreality
7. Dimensionless Parameters p'/p and Q
8. Flows with Q ^ 0
9. Case p'/p Near Zero: Cavity Flows
10. Transient Scale Effects
xi
 CONTENTS
11. Wall Corrections; Stability of Pressure Coefficient
12. Lined Cavity Charges
13. Penetration by Liquid Jets
14. Stationary Cavities in Accelerated Motion
15. Axially Symmetric Flows
16. Summary of Recent Tendencies

Chapter III. Modeling and Dimensional Analysis . 77

1. Introduction
2. Dimensional Analysis
3. Statement of Assumptions
4. Pi Theorem
5. Dimensional Homogeniety
6. Validity of Assumption II
7. Validity of Assumptions I and IV
8. Inspectional Analysis
9. Relation to Groups
10. Theory of Modeling, I
11. Theory of Modeling, II
12. Inertial Modeling
13. Reynolds Modeling
14. Froude Models
15. River and Harbor Models; Nearly Horizontal Flows
16. Modeling of Cavitation
17. Modeling for Entry into Water
18. Mach Modeling
19. Linearized Mach Modeling

Chapter IV. Group Theory and Fluid Mechanics . 116

1. Introduction
2. Symmetric Solutions of the Heat Equation
3. Prandtl-Meyer Expansion Wedges
4. Taylor-Maccoll Conical Flows
5. Non-stationary One-dimensional Compressible Flows
6. Spiral Viscous Flows
7. Laminar Boundary Layer
8. Other Symmetric Flows
9. General Local Existence Theorems
•10. Group Theory and Separation of Variables
11. Application of the Special Group of Mach Modeling
12. Separation of the Time Variable: Viscous Flows
13. "Inverse Methods" and Group Theory
14. Hodograph Method and Groups
15. Nonlinear Theory of Inertial Scaling
16. Generalization
17. Summary
xii
 CONTENTS
Chapter V. Virtual Mass and Groups 151
1. Introduction: Virtual Mass of a Sphere
2. Physical Validity of Conclusions
3. Induced Mass Tensor
4. Reduction to Canonical Form
5. Induced Mass and Momentum
6. The Case of Rankine Bodies
7. More Applications of the Induced Mass Concept
8. Inertial Lagrangian Systems
9. Geometrical Interpretation
10. Group-theoretic Interpretations
11. Concepts from Lie Theory
12. Curvature of One-parameter Subgroups
13. Classical Applications
14. Generalizations
15. Proof that System is Lagrangian
16. General Conclusions

Bibliography 181

Index 185

xin
Hydrodynamics
 CHAPTER I

Hydro dynamical Paradoxes

1. The Practical Value of Paradoxes

As fluid mechanics comprehends the physics of two of the

three most common states of matter — the liquid and the
gaseous — it is superfluous to explain why one or even many
books should be devoted to the subject. It is less obvious
why I have thought it suitable to begin this book with a dis
cussion of "hydrodynamical paradoxes." It might seem
that this unduly emphasizes negative aspects of the subject.
Therefore I should like to begin by asserting the importance
of such paradoxes from the practical, the mathematical,
and the philosophical points of view. By a Varadox I shall
mean a plausible argument which yields conclusions at var
iance with physical observation, as it is in this sense of an
apparent inconsistency that the term is most commonly used
in hydrodynamics.
From a practical point of view, perhaps the most important
art to acquire in fluid mechanics consists in knowing when a
given theory is "applicable," to use the current phrase. The
fundamental importance of acquiring this art has been stressed
in various recent books,1 which have tried to bridge the
wide gap between theory and experiment which exists in
hydrodynamics.
Such books tend to attribute this gap to the difference
between "real" fluids of small but finite viscosity, and "ideal"
fluids having zero viscosity.2 They tend to minimize the
importance of logical and mathematical rigor, and to trace
all discrepancies between "theory" and "experiment" to
the single "unjustified" approximation of zero viscosity. By
familiarizing the reader at the outset with the physical facts
of separation of the boundary layer and turbulence, they
1 See for example the prefaces of [12] and [14], or p. 2 of [6]. Numbers
in square brackets refer to the bibliography on p. 181.
1
See [6], pp. 2, 48; [12], Vol. 1, p. 3, and the introduction to Vol. 2;
[14], p. x.
3
 HYDRODYNAMICAL PARADOXES [CH. I

put him on his guard against some of the oversimplifications

of classical hydrodynamics.
The art of knowing "how to apply" hydrodynamical
theories can be learned even more effectively, in my opinion,
by studying the paradoxes which I shall describe. Moreover,
I think that to attribute them all to the neglect of viscosity
is an unwarranted oversimplification. The root lies deeper,
in lack of precisely that deductive rigor whose importance
is so commonly minimized by physicists and engineers. They
can be traced to the use of plausible arguments, which have
often proved most fruitful in physical reasoning, but which
are nevertheless fallible, as simple examples show. Among
these are the arguments that "small causes produce small
effects" and that "symmetric causes produce symmetric
effects."
A clear understanding of the fallibility of these arguments
in specific hydrodynamical applications, should be useful to
any research worker who wishes to make reliable theoretical
predictions.

2. Their Mathematical and Philosophical Interest

The possibility that arbitrarily small causes can produce

finite effects in fluid mechanics has been clearly recognized by
Oseen ([11], pp. 211-213). Oseen has pointed out that in
systems of differential equations, the presence of arbitrarily
small terms of higher order can change entirely the behavior of
the solutions. It need not be true that as the coefficient of a
certain term tends to zero, the solution tends to the solution
obtained by annihilating that coefficient. Paradoxes due to
this source may be called asymptotic paradoxes.
The possibility of asymptotic paradoxes cannot be admitted
If it
is,

as a universal principle.3 since every physical experi

ment actually affected by countless minute influences,4 we
is

have been informed by J. R. Porti6 and L. L. Whyte that asymptotic

1
1

paradoxes are common in biology (mutations of genes, etc.).

Heating and quantum reactions from cosmic rays, variations in
4

;ravity due to distant stars and the Coriolis force, seismic vibrations,
mpurities in the materials used, etc.
4
 §2] MATHEMATICAL INTEREST

should have no chance of ever predicting correctly the result

of any specific experiment. Since the results of many experi
ments can be predicted with practical certainty, it is not a
universal principle; but is presumably restricted to certain
types of partial differential equations.
Hence asymptotic paradoxes should interest the pure
mathematician by providing suggestive evidence regarding
the facts about partial differential equations — a complex
subject concerning which our knowledge is still very rudi
mentary. Several of the other paradoxes which I shall
describe will provide physical evidence of further properties
of partial differential equations, and should therefore also
interest the pure mathematician. Indeed, I am convinced
that in order to make real theoretical progress, we must study
a wide variety of physically and mechanically constructed
solutions of partial differential equations of specific types.6
Thus there are several interesting cases in hydrodynamics
where an apparent symmetry of causes is not preserved in the
effects. Such paradoxes may be called symmetry paradoxes,*
and they have an evident philosophical interest for the logic
of physics.
In general, hydrodynamical paradoxes have a philosophical
interest, as illustrating the profound interweaving of deductive
and inductive logic in scientific research. I believe that the
resourcefulness and ingenuity of the human mind in combining
these methods is vastly underestimated in all existing formal
•theories of mathematical and physical logic.7
They also have an interest for the pure mathematician, by
illustrating a wealth of phenomena not anticipated by classical
mathematical theories. Even a superficial attempt at mathe
matical classification reveals what may be called paradoxes
of overdeterminism, singular point paradoxes, paradoxes of
approximation, and paradoxes of topological oversimplifica-
• This idea, with especial reference to high-speed computing machines,
has been stressed repeatedly by J. von Neumann.
•They seem to contradict the logical axiom that "causes invariant
under a group G induce effects invariant under the same group," G.
Bouligand, Thiwie gknirale des groupes, Paris, 1935, p. 3.
7
In particular, mathematical logic seems to me to suffer from too
exclusive preoccupation with the foundations of set theory.
5
 HYDRODYNAMICAL PARADOXES [CH. I

tion. I shall now try to support the statements made above

by a discussion of precise mathematical details.

3. Fundamental Approximations

The role of postulates in theoretical fluid mechanics is

played by an "equation of state," and certain partial differ
ential equations and boundary conditions.
One assumes that the density p and viscosity /* of any fluid
depend on the pressure p and temperature T, so that

(1) p = f(p,T), (equation of state)

(1') n=fi(p,T).

In mathematical discussions, fi is commonly treated as a

constant, which is nearly fulfilled in most practical applica
tions, provided temperature variations are moderate.
It is then convenient to take the "Eulerian" independent
variables of vector position x = (xi,x2,x3) and time t. Assum
ing that all functions involved are sufficiently differentiable8
(even the assumption of continuity is non-trivial, as we shall
see!), we can express the conservation of mass as the so-called
Equation of Continuity

div (pu) + = 0,
(2)
^

where u = (wi,M2,m3) denotes vector velocity.

It convenient to use instead "Lagrangian"
is sometimes
independent variables a = (ai,a2,a3) associated with material
points moving with the fluid, and t. If a is held fixed, the
time derivative becomes the "material" derivative

-rA
= — + ) Uk ^
— (in Eulerian variables).
dt dt Li dxk

8 Careful discussions of possible discontinuities have been given by

Paul Duhem, "Recherches sur l'hydrodynamique," Annates de la
Faculti des Sciences de Toulouse, Vols. 3-5 (1901-1903) and by L. Lichten-
stein, Hydrodynamik, Berlin, 1933. I am indebted for these references,
and for other helpful comments on Chapter I, to Prof. J. Kampe de
Feriet.
6
§4] NON-VISCOUS FLUIDS

Hence Ui = dxi/dt and (2) is equivalent to

(2') g + divu-0.
The other basic differential equations of fluid mechanics
express Newton's momentum law for continua, and are called
the Equations of Motion; they involve the vector acceleration
of gravity g = g(x).
In stating the Equations of Motion, one must make one of
two approximations. Either one must neglect viscosity,
which gives the approximation

(3) -ST + - grad p = g (p variable)

at p

for a so-called non-viscous fluid; or one must neglect compressi

bility, setting p = const., and use the Navier-Stokes equations

(3*} + igradp = „V'u + g(

^
for an incompressible fluid. Here the ratio v = n/p is called
the kinematic viscosity of the fluid.
These approximations are made partly for simplicity — even
with them, one cannot in most cases solve the mathematical
problems involved. They are also made because we do not
actually know how viscosity acts in a fluid under rapid com
pression (see [12], Vol. 1, p. 260)!
Hence if we are too impressed by the dangers of approxima
tion, we must despair at the outset of attaining any rational
understanding of fluid mechanics. However, science was not
built by timid or faltering hands. Let us move boldly forward,
seeing what kind of edifice one can construct on these some
what insecure foundations.

4• Non-viscous Fluids

I shall begin (Sections 4-12) by making the approximation

of zero viscosity. The a priori plausibility of this approxima
tion would be hard to exaggerate. Thus ([6], p. 4) a change
in velocity of 100 m.p.h. across a layer of air .01 in. thick,
7
 HVDRODYNAMICAL PARADOXES [CH. I

exerts a shear force of only one ounce weight per square foot.
Its plausibility can also be expressed in terms of the Rey
nolds number, denned as the dimensionless ratio

(4)
a-fi™-™,
H v

where V velocity and L a representative linear

is a typical
dimension. Since R expresses the ratio (inertial forces)/
(viscous forces), it seems reasonable to suppose that the
approximation v = 0 should become asymptotically valid for
sufficiently large R, say, over one million. This supposition
becomes even more reasonable if we recall (Section 2) the
many other minute forces which are always necessarily
neglected in making physical predictions. Without further
discussion, let us therefore introduce the fiction of a "non-
viscous" fluid, in which (3) is valid.
In studying the flows of "non-viscous" fluids from homo
geneous reservoirs, it is usually also assumed that the fluid
expands and contracts adiabatically, so that

(1') p = f(p). (In perfect gases, p = kpy.)

This assumption is reasonable if viscous forces are negligible,

because conduction and viscosity are both manifesta
heat
tions of molecular diffusion. It can also be justified by
separate computations, if we admit the principle that small
variations in density exert a small influence in fluid flow. This
principle, unlike the corresponding principle for viscous forces,
appears to be entirely supported by the physical evidence.

5. Mathematical Properties

From the preceding approximate assumptions, several

fundamental mathematical consequences can be derived, and
I should like to emphasize the fact that these consequences
do in fact explain observed resistance to acceleration from
rest (see Chapter V) and many other physical phemomena.
Thus, in the case of "barotropic" fluids satisfying (1'), one
can prove ([12], Vol. 1, p. 191) that the "circulation" fLukdxk
around any closed "material" curve moving with the fluid,
§5] MATHEMATICAL PROPERTIES

is constant. Hence if the fluid comes from a reservoir without

circulation (i.e. at rest in some non-rotating reference frame),
and if the curve remains closed at all times, the circulation
should remain zero. This is mathematically equivalent to
the statement that there should exist a locally single-valued
scalar velocity potential U(x;t), such that

(5) u(x;0 = grad U, or w, -^ [i - 1, 2, 3].

In a simply connected domain, such as the exterior of a

solid or half the symmetric exterior of a circular cylinder,
U must therefore be single-valued in all space. In any case,
this is the simplest possible assumption; it was universally
accepted for over a century; and we shall make it until Sec
tion 8, where the possible exceptions will be discussed.
Further consequences of the basic hypotheses (1)-(3)-(4)
can be deduced in the case U = U(x) of "steady" fluid
motion. Consider the case of the flow of a stream of fluid
(such as air or water) moving at constant speed past a fixed
obstacle — to which the case of a rigid body moving at nearly
constant speed through an infinite fluid initially at rest can be
reduced, by the use of moving axes.
Since the boundary conditions and differential equations
are independent of time, one then naturally writes U(x). One
can prove that, under these circumstances, the equations of
motion (3) are equivalent after one space integration to the
Bernoulli equation

(6)
tvTTVTT-L.
VUVU + J (*l
^ +G = po, or
^

I Ujdui H
—-
P
+ dG = po.

Here G denotes the gravitational potential, so that g = — VG.

An important special case is that of (nearly) incompressible
flow, in which case we set p = const. In this case, (2) is
simply div u = 0, which is equivalent by (5) to

(7) V2C/ = 0,
9
 HYDRODYNAMICAL PARADOXES [CH. I

[Link] the existence of a harmonic velocity potential. For

such flows, (6) assumes the familiar simple form

(6') p = po - VipVUVU - pG.

In fact, one can prove more generally, in the case of

"unsteady" or "accelerated" motion of an incompressible
non-viscous fluid, initially without circulation, that the pres
sure must satisfy

(6") p = Po-p +
^+g},
{IwvU
which is equivalent to the equations of motion, just like (6).
So much for the internal conditions satisfied by flows of an
incompressible, non-viscous fluid from a still reservoir: we have
the basic kinematic condition (7), from which pressure can be
determined by (6') in the case of steady motion. In this case,
the equation of continuity clearly becomes

(8)
—
HIT
= 0 on the boundary,

where d/dn denotes the outward normal derivative. Simi

larly, the condition that a is the "free stream" velocity is
clearly
(9) limx->„ grad U = a.
Now by potential theory pp. 211, 311), with bodies
([7],
of reasonable shape, equations (7) to (9) have one and only
one solution. This rather deep mathematical result is in
perfect accord with the intuitive idea that there should be
exactly one flow around a given fixed obstacle, having a given
relative velocity a. The flows obtained in this way are called
Dirichlet flows.
The theory of Dirichlet flows is completely satisfactory
from a logical and mathematical point of view. Let us now
apply it to the study of fluid resistance.

6. The D'Alembert Paradox

One of the simplest examples of a Dirichlet flow is the steady

flow of an incompressible non-viscous fluid past a sphere of
radius L centered at the origin. We may suppose that the
10
§6] THE D'ALEMBERT PARADOX

free stream velocity has magnitude V, and is parallel to the

x-axis. The solution is found by superposing a "dipole
field " on a uniform flow, so that

(10) U = Vxl + ^ [«-WH-^J.

It is easy to compute by (6') the theoretical pressure dis
tribution around under the flow (10).
a sphere One finds
that, apart from the hydrostatic buoyancy (cf. Chapter III,
Section 12), the pressure distribution has complete fore-and-
aft symmetry, and hence gives a pressure thrust zero.
This theoretical conclusion, called the d'Alembert Paradox,
is not limited to spheres; one can prove generally9
Theorem 1. The hydrodynamic force system on any finite
obstacle B, under steady Dirichlet flow in an infinite stream,
reduces to zero or a pure couple.
Proof. The assertion is that the integral FB of the pressure
differentials over the surface of B is zero. To prove this,
we choose the Xi-axis in the direction of flow, and let V be
the free stream velocity. By hypothesis, V(J7 — Vxi) — > 0
asx-> oo ; moreover ff(dU/dn)dS = 0 on any closed surface
containing the obstacle. It follows by potential theory ([7],
p. 270) that U — Vxi = 0(l/r2), meaning that it tends to
zero as rapidly as 1/r2, where r = \x\. Further

(10') u = (7,0,0) +
•G>
whence VC/V U = F2 + 0(1 /V3).
Neglecting hydrostatic buoy
ancy as before, we find that the hydrodynamic pressure satis
fies, by Bernoulli's equation (6'),

(11) P =
P.-^-o(l) =
P/-o(i).
Here p0 is the "stagnation pressure" at zero velocity; the
constant p/ = p0 — UpV2 = lim^. p(x) is called the "free
stream pressure. "
For a recent discussion, see M. Manarini, "Sur paradossi d'Alembert
9

e di Brillouin nella dinamica dei fluidi," Atti Accad. Naz. Lincei, Rend. 4
(1948), 427-433, reviewed in Math. Reviews, 10 (1949), p. 335.
11
 HYDRODYNAMICAL PARADOXES [CH. I

a.

'8 cc a

i
i
M
c3

o o o o o o o — (O
o o o o —
<o CM
6 o o o 9
tP do
o o cm —
<o CN
6
6
o o o o
do
CO •* oo
o o <0 •*
00 ><t
9
o o

12
 §7] THE REVERSIBILITY PARADOX

Now consider the momentum equation Vol.

Chapter
([12], 1,

XIV) for the fluid occupying, at variable time t, a fixed large

sphere jS of radius a. Clearly, for any a large enough so that
S includes B, FB is equal to the difference E(a) — Fs(a)
between the rate E(a) at which momentum flows outwards
through S, and the pressure thrust Fs(a) exerted inwards
across the surface of S. Hence

FB =
lim0_.{E(a) Fs(a)}.
-
But by equations (lO')-(H), the pressure on the surface of S
differs by 0(l/a3) from the constant p/; therefore, the area
of S being 4ta2, Fs(a) differs by 0(l/a) from zero: Fs(a) =
0(l/a). Similarly, the flux of mass per unit area through £
is Sutxk/a, and so by (10') differs by 0(1 /a') from that of a
"uniform flow" with constant velocity (F,0,0). Again, the
vector momentum transported through a unit area per unit
time, is u2M*x*/a; this also differs from that of a uniform flow
by 0(l/a3), using (10') again. But for a uniform flow, clearly
E(a) =0; hence actually, multiplying once more by 4xa2,
E(a) = 0(l/a). Substituting above, we conclude FB = 0, as
desired.
The preceding theoretical prediction, of zero resistance or
"drag" D and lift L, is obviously contradicted by everyday
experience; thus in Figure 1 one sees the drag coefficient
CD = D/}4pV2A (A = cross-section area) of a sphere, plotted
as an empirical function of the Reynolds number R. I shall
discuss shortly (in Section 9) the rationalizations of this con
tradiction between theory and experiment, but first I should
like to describe some closely related paradoxes.

7. The Reversibility Paradox

Although it is probable that the d'Alembert paradox is also

valid for subsonic, compressible flow of a non-viscous fluid,
this seems hard to prove rigorously.10 However, a very
simple argument shows that fluid resistance can never be
10 A "proof"
has been given by U. Cisotti, Comptes Rendus, 178 (1924),
p. 1792. But as there is no rigorous analog applying to nonlinear partial
differential equations of the result proved on p. 270 of [7], much stronger
hypotheses regarding the behavior at infinity must be made.
13
 HYDRODYNAMICAL PARADOXES [CH. I

explained from the hypothesis of a non-viscous fluid as

formulated above (Sections 4-5).
Definition 1. By the reverse of a given flow, described
in Lagrangian coordinates, we mean that flow obtained from
the given flow by the substitution t—> — t. In Eulerian
coordinates, this corresponds to the reversal w< — > — w< of
velocity direction as well as reversal t — > — t of time, but
pressure p and density p are still preserved at corresponding
points (xi, • • • ,xn;t) of space time.
Definition 2. A condition on flows is reversible when

if,
ever holds for a flow, holds for the reverse of that flow.
it

it
theory of fluid dynamics reversible theory all the con
A

is

if
a
ditions which imposes on flows are reversible.
it

Actually, all the familiar conditions on flows are reversible,

except those which involve viscosity (friction), thermal con
duction, diffusion, or shock wave fronts.
Definition theory of fluid dynamics will be called
A
3.

incomplete its conditions do not determine the steady flow

if

around body uniquely; overdetermined no flow com

is
if
a

patible with it; incorrect its predictions do not agree closely

if

with experimental fact.

Theorem (Reversibility Paradox.11)
2. Any reversible
theory fluid dynamics is either incomplete, overdetermined, or
of

grossly incorrect, so far as its predictions steady state lift and

of

drag are concerned.

Proof. Such theory will predict by (6') that steady
a
a

flow and its reverse will give the same pressure thrust on an
obstacle, whereas matter of• common experience that
is
it

a
a

flow and its reverse ordinarily give pressure thrusts in approxi

mately opposite directions.12
11 In general, physicists only philosophical interest to
seem to ascribe
the fact that the concept of non-viscous fluid motion reversible. Thus
is

the fact not mentioned at all in [9], [10], or [12].

is

11 The
preceding discussion borrowed from [1]. Recently, M. Munk
is

has extended the reversibility principle to linearized supersonic airfoil

theory, in Naval Ordnance Lab. Memo. 9624 (1948). reversibility
A

principle for theoretical ship wave resistance has been derived by T. H.

Havelock, Proc. Roy. Soc., A149 (1935), p. 423; the discrepancy with
experiment analyzed by W. C. S. Wigley, Trans. Inst. Nav. Arch., 86
is

(1944), 1-20. Similar considerations have recently been used by

A. Busemann to show that no existence and uniqueness theorem can be
hoped for in the case of transsonic flow.
14
§8] PARADOXES OF AIRFOIL THEORY

Now the approximations of Sections 4-5 are certainly com

patible, because they permit Dirichlet flows around objects
of (almost) arbitrary shape. Hence any discussion of fluid
resistance which is based on them must either be incorrect,
or else involve further assumptions. It is generally believed
that both alternatives are possible.
Thus for subsonic flow,18 it is often stated that no physically
correct theory of resistance is possible, which is based on the
differential equations of Sections 4-5. Whereas for supersonic
flow it is known that these equations are compatible with more
than one flow, and hoped that resistance can be explained by
using these equations and qualitative physical considerations.
I shall now discuss in more detail the incompressible case,
which is generally supposed to be typical of subsonic flow,
because the partial differential equations involved are of
elliptic type in both cases.

8. Paradoxes of Airfoil Theory

It is well-knownthat, if one is willing to admit discon

tinuous functions — that
is,

to abandon the assumption that

all "reasonable" functions are differentiable — the differential
equations (1)-(3)-(4) permit many steady flows around an
obstacle. These flows are of various topological types, and
most of them avoid the d'Alembert paradox. As such "dis
continuous fluid motions" form the main theme of Chapter II,
shall not discuss them here, except to mention that they,
I

too, will be found to be subject to paradoxes.14

Instead, shall discuss two-dimensional airfoil theory, which
I

gives the best known theory of lift — although does not give
it

basis for estimating resistance. In this theory, the multiple-

a

valued velocity potentials, excluded in Section are con

5,

sidered. Mathematically, this amounts to replacing (10') by

the multiple-valued potential

IISee Th. von Karman, "The problem of resistance in compressible

fluids," Reale Accad. d'Italia (1936); U. Cisotti, op. cit. supra; and Sec
tion 12 infra.
14 Attention called to the Hadamard paradox, and to the paradoxes
is

of wake underpressure, of indeterminacy, of surface seal, and of down-

nosing, discussed in Chapter II.
15
 HYDRODYNAMICAL PARADOXES [CH. I

(12) u - r^ +^ <*»*» + Ui,

where V*Ui = 0, £/i = 0 ( ^Y

Here r = fZuaixk = <fdU is called the circulation, the
integral being taken along a contour going once around the
"airfoil" (or obstacle) B in question. Since any real har
monic function U(x,y) is the real part of a complex analytic
function W(z) of z = x + *#' we can write equivalently
(c = T/2x being real and c' complex),

(12') U =
fle{ -ic In z.+ c'z + Wi),

Wt = 0
Y
where
f^2

We shall find convenient to use such "complex potentials"

it

W below, and again in Chapter II.

The introduction of the term Td/2v in (12) sacrifices the
•determinism principle established at the end of Section 5:
there solution of (12) for each choice of r. One can rescue
is
a

determinism in the case that the airfoil has sharp "trailing

R

"
edge" (but not otherwise), by assuming the so-called Joukow-
sky Condition," that "velocity finite at the trailing edge."
is

By assuming that the sharp edge trailing, one avoids the

is

Reversibility Paradox. Dirichlet flow with circulation

A

which satisfies the Joukowsky Condition may be called a

" Joukowsky flow." Thus the Dirichlet flow without circula
tion past flat plate inclined at 15° to the main stream
is
a

sketched in Figure 2a; the Joukowsky flow past the same plate
sketched in Figure 2b; the "cavity flow" (cf. Chapter II)
is

past shown in Figure 2c.

is
it

U,

Knowing one can then get theoretical estimate of the

a

thrust on an "airfoil" of arbitrary cross-section, using the

following " Kutta-Joukowsky Theorem."
Theorem In any Dirichlet flow with circulation, the
3.

drag1t' is zero, and the lift is the product = pvfT,

L

the fluid
of

density, the free stream velocity, and the circulation.

l* The "drag" as usual the component of force parallel to the flow;
is

the component perpendicular to the flow. In Theorem D =• = 0.

L
is
L

1
,

16
§8] PARADOXES OF AIRFOIL THEORY

This theorem can be proved in the same way as Theorem 1,

without assuming the Joukowsky condition. It is a corollary
that the lift is

(13) L = Y2pv) cC* sin a,

where c is the wing chord, a is the "effective angle of attack,"

and C* depends only on the shape of the wing cross-section.

Fig. 2. Three ideal flows past inclined plate.

A similar asymptotic argument determines the moment in

any Joukowsky flow,16 and this gives the best known theo
retical estimate for the variation in airfoil forces with wing
cross-section. Nevertheless, the theory outlined above is
subject to three paradoxes. The first of these is quite amusing

" The complete theory was given by R. von Mises, Zeits. Flugtechnik
und Motorluftschifffahrt, 1917, pp. 157-163, and 1920, pp. 67-73 and
87-89. For Theorem 3, see [12], Vol. 2, p. 163; shorter proofs are known.
17
 HYDRODYNAMICAL PARADOXES [CH. I

and purely mathematical in nature; it may be called the

Cisotti Paradox (references for the full discussion may be
found in [2]).
Consider the Joukowsky flow past an inclined flat plate in a
horizontal wind (Fig. 2b). By the Kutta-Joukowsky Theo
rem, the resultant force should be vertical (i.e. pure lift). On
the other hand, as all pressures are normal to the plate,
the resultant force should be normal to the plate, giving a clear
contradiction. The latter conclusion is confirmed by an inte
gration along the plate of the pressure difference between the
two sides; the detailed formulas follow.
The most general Joukowsky flow around the unit circle
z = eu has the velocity potential proportional to
L5?
U = Re + ln

ic
e-£-

4
where a and Since 2f =
are real. 1/z maps = eie onto
+
z

z
c

—
l^|^l[f = + irt), the most general Joukowsky flow
£

around the segment — = cos 0^1 has velocity poten

^

a
1

tial on this segment conveniently written as

= +
+
U

Re[Y2(ei^+^ e-«**»>) cd]

= cos +
+
(a

cd.
6)

Using as parameter, the velocity on the plate

is
a
6

dU _ -dU sin
+

c_
(a

6)

=
dt sin Odd sin sin
6

= cos a sin a cot —

+

esc
6.
6

Hence the difference in pressure between similarly situated

points on opposite sides —6 given by Bernoulli's Theorem
is
6,

as
Ap = 2p cos a[sin a cot —
6];

esc
6

'
the integral Apds = — Ap sin clearly convergent.
is

Odd

The resolution of the Cisotti Paradox leads one to the

astonishing conclusion that there non-zero force acting on
is
a

the single point at the front end! This clearly a very pure
is

example of Singular Point Paradox, arising from neglect

a
a

18
§9] DISCUSSION OF PARADOXES

to study rigorously the behavior of the flow at all singular

points.
A second mathematical paradox arises from the fact that,
in space, the exterior of wing of finite span is simply con
a
nected. It follows that a multiple-valued potential with
zero vorticity (i.e. curl u = 0) at all points is impossible;16
hence no escape from the d'Alembert Paradox along the lines
of two-dimensional airfoil theory is mathematically possible
in three-dimensional space.
Even the two-dimensional theory, however plausible its
assumptions, agrees doubtfully with experiment. It involves
the following Fatness Paradox (see [1], p. 252). Theoretically,
the quantity C* in (13) increases as the airfoil thickness increases;
experimentally, it decreases as the airfoil thickness increases.

9. Discussion of Preceding Paradoxes

In spite of the preceding the theory of an incom

paradoxes,
pressible non-viscous fluid, or "ideal fluid," is extremely useful
to the research engineer in dealing with fluids of small vis
cosity, at speeds up to one-fourth the speed of sound. Out
side of a rather wide belt behind a given obstacle or airfoil B,
called the "wake," and a very thin "boundary layer" near
the surface of B (cf. Figs. 3a-3b), its predictions are reliable.
By accepting the wake and boundary layer as empirical
facts, one can rationalize most of the preceding paradoxes.
More specifically, one must recognize that, as the viscosity
tends to zero, the width of the wake remains finite, and the
"wake underpressure coefficient" (pt — p^/^ipv2 (p/ = free
stream pressure, pw = wake pressure) does not tend to zero.
This leads to a "drag coefficient" CD = D/^pv2A which does
not tend to zero with the viscosity (see Fig. 1); and it gives a
flow pattern which is clearly irreversible. Finally, the Fatness
Paradox may be rationalized as the natural tendency of the
16This remark may be found in K. Friedrichs and R. von Mises, Fluid
Dynamics, Brown University Lecture Notes, 1941. Nevertheless
attempts have been made to compute lift forces on airships within the
framework of incompressible non-viscous flow! In certain pumps, the
propellers extend to the pipe walls, and so the fluid is multiply-connected;
this topological fact has been said to give such pumps a superior efficiency.
19
 HYDRODYNAMICAL PARADOXES' [CH. I

wake to become "fatter" with the airfoil, a tendency which

diminishes the circulation.
The preceding rationalizations are strongly confirmed by the
fact that ([6], Chapter XII, or [12J, Vol. 2, p. 81), by sucking
away a very thin boundary layer, with a relatively minor
expenditure of energy, Dirichlet and Joukowsky flows can be
very closely approximated. Correspondingly, the d'Alembert
and Fatness Paradoxes can be largely avoided. This suggests

Dividing .Streamline

Boundary Layer

Dividing Boundary Layer

Streamline

Fig. 3. Schematic diagrams of "real" flows.

that the prevailing view already mentioned (Sections 1 and 7),
that the paradoxes of fluid mechanics are due to the neglect
of viscosity, may be incorrect. It is conceivable that, when
the Reynolds number R is very large, the concept of an ideal
fluid is not incorrect, but merely incomplete17 (cf. Theorem 2).
In support of this idea, it may be mentioned that the deduc
tion of Dirichlet flows from the hypothesis of an "ideal fluid"
involves implicitly two arbitrary topological assumptions:18
that the streamlines from the upstream side cover the
(i)

whole plane, and (ii) that locally single-valued velocity

a

17 The sense in which the fundamental differential equations, etc., for

an ideal fluid are "complete" discussed in Chapter III, Section 11.
is

18
Much of modern clarification of the foundations of geometry due
is

to the explicit consideration of Euclid's implicit topological assump

tions — as in D. Hubert's classic Grundlagen der Geometrie, Berlin, 1930.
20
§9] DISCUSSION OF PARADOXES

potential U is single-valued in the large. There is no valid

mathematical reason for excluding from consideration Jou-
kowsky flows, flows with wake (see Chapter II), and many
other topological types of flows.19 Hence the paradoxes of
ideal fluid theory may be, in part, paradoxes of topological
oversimplification .
In any case the preceding examples make it clear that the
theory of non-viscous flows is incomplete. Indeed, the reasoning
leading to the concept of a "steady flow" is inconclusive;
there is no rigorous justification for the elimination of time as
an independent variable.
Thus though Dirichlet flows and other steady flows are
mathematically possible, there is no reason to suppose that
any steady flow is stable. It is perfectly conceivable that, in
an "ideal" fluid, initially departing slightly from Dirichlet
flow, irregularly varying turbulent "eddies" are built up
mathematically in the "wake" of an obstacle — reproducing
mathematically what is observed physically at large Reynolds'
numbers R. Indeed, such a model has already been con
structed by von Karman (loc. cit. supra), for the alternating
vortices shed by cylindrical obstacles at intermediate Reynolds
numbers, 50 < R < 1000.
To admit this possibility, we must reject the idea that there
is a necessary tendency towards symmetry in natural phe
nomena,20 and admit the possibility that a symmetrically
stated problem may not have any stable symmetric solution.
Extensive hydrodynamical evidence of this possibility is
presented in Section 14-15. Admitting it provisionally, we
" For example: The "multiple solutions" of H. Villat (see Lecons sur
I'hydrodynamique," Paris, 1929, Chapter VIII); the solutions with two
symmetric vortices in a finite wake, introduced by F8ppl (see [9], p. 223,
or H. Bateman, Partial Differential Equations, 1932, p. 251, and refs.
given there) ; Karman's flows with infinite vortex trails ([9], p. 225; or [10],
p. 341); and the reverses of the preceding flows (see E. Picard, Comptes
Rendus, 159 (1914) 639). It should be noted, however, that circulation
or a wake of infinite length and breadth is needed to avoid the d'Alembert
Paradox.
10 This idea has been advanced by various philosophers of
science; see
L. L. Whyte, "Tendency towards symmetry in fundamental physical
structures," Nature, 163 (1949), p. 762. For a general discussion, see
P. Duhem, La thiorie physique, son objet et sa structure, Paris, 1907,
Part II, Chapter III.
21
 HYDRODYNAMIGAL PARADOXES [cH. I

are led to the conclusion that the paradoxes of ideal fluid

theory are, in part at least, paradoxes of topological over
simplification and symmetry paradoxes, in which "apparently
symmetric causes do not have symmetric effects." (The latter
is especially evident in the case of the Reversibility Paradox,
if one considers wake theory.)
At all events, the most obvious paradoxes can be attributed
as much to lack of deductive rigor as to faulty "physical"
assumptions or approximations. In further support of this
view, I shall now show that the theories of compressible fluids
and of viscous fluids (so-called "real" fluids) have their full
share of paradoxes.

10. The Earnshaw Paradox

I shall first treat the paradoxes which specifically concern

compressibility effects arising in the high-speed flow of non-
viscous fluids, beginning with a paradox due to Earnshaw.21
Earnshaw Paradox. Plane sound waves of finite amplitude
and stationary form are mathematically impossible, in a gas
which vibrates adiabatically.
Proof. We assume that the form of the sound waves is
stationary, and that the train of waves moves with constant
velocity normal to the wave fronts. Hence if we change to
axes moving with the waves, the fluid motion will be not only
one-dimensional but steady, and we can write p = p(x),
u = u(x), etc. Neglecting gravity, Bernoulli's equation (6)
reduces to udu + dp/p = 0. Moreover the equation of con
tinuity (1) reduces to pu = const. = C, or u = C/p. Sub

-_
stituting back we get

,,,. CHp .dp . , CHp

(14) + -£ = 0, or dp = —r*
Hence such a wave motion is possible only if the fluid satisfies
an equation of state (2') of the special form

"8. Earnshaw, Phil. Trans., ISO (I860), 133-148. See also Stokes,
Phil. Mag., 33 (1848), 349; Rankine, Phil. Trans., 160 (1870), 277; [13],
Vol. V, p. 573, or Proc. Roy. Soc., A84 (1910), 247-284; [9], §283; [60],
§51. Of course, the paradoxes of Sections 6-8 also apply to compressi
ble non-viscous fluids.
22
 §10] THE EARNSHAW PARADOX

C2
(15) p = p0
P

There is no known gas whose adiabatic equation of state has

the form (15).
We have here a typical paradox of overdeterminism; such
paradoxes arise typically when too many approximate hypoth
eses are made. It would be hard to say that any one approxi
mation was at fault; indeed, like many other paradoxes,22
the Earnshaw Paradox contains the germ of an important
truth. For waves of large finite amplitude, it leads to the
correct and remarkable conclusion that the dense portions
of such waves gain progressively on the more rarefied portions,
until eventually shock waves occur where there is a near-
discontinuity of density, pressure, and velocity.
Shock waves are now such a commonplace that it may be
of interest to recall that Riemann treated them incorrectly,
and that Lamb, even in the last edition of his classical Hydro
dynamics, casts doubt on the validity of the now accepted
Rankine-Hugoniot theory.23
From the Earnshaw Paradox one can easily derive a second
paradox, noted independently by Mr. H. D. Ursell and myself.
"Everybody knows" that there is a close analogy between the
theory of shallow surface waves, and compression waves in a
gas whose equation of state is p = kp2 + c, with •y = 2 (cf.
Chapter III, Section 15); this analogy has been effectively
exploited recently by J. J. Stoker.24 Yet it was proved by
D. J. Struik in 1925 that shallow surface waves of finite ampli
tude were possible; and we have just proved that waves of

" Thus the d'Alembert paradox contains the germ of the following
useful engineering principle: By "sucking away the boundary layer,"
in a fluid of sufficiently low viscosity, resistance can be reduced to nearly
zero, by an operation which involves very little energy.
"Lamb ([9], p. 486) states that "no physical evidence is adduced in
support of the proposed law," and that "in actual fluids, the equation
of energy cannot be satisfied consistently. ..." I am indebted to Mr.
R. H. Kent for this reference.
u See N.Y.U. Communications in Applied Mathematics, 1 (1948),
1-87 and refs. given there. The analogy was apparently first noted by
E. Jouguet, Jour, de math, pures appl., 3 (1920), p. 1. For Struik's
results, see Rendic. Lincei, 1 (1925), 522-527.
23
 HYDRODYNAMICAL PARADOXES [CH. I

finite amplitude are impossible for 7 ?■*

— 1. What is the
explanation? Nobody knows.

11. Von Neumann's "Triple Shock" Paradox

Since shock waves are now in the center of popular atten

tion, it may be of interest to call attention to a new paradox
regarding shock waves.
The mathematical theory of plane shock waves can be
derived simply from the laws of conservation of mass, momen
tum, and energy (heat plus kinetic energy), provided one

Reflected Incident Reflected Incident

Slipstream

Reflected , Incident Reflected Incident

A •,-.
JD
Fig. 4. Shock wave reflections.

assumes that on each side of the shock wave all relevant

physical variables (temperature, density, pressure) assume
limiting values. The predictions of the Rankine-Hugoniot
theory derived from these paws seem to be closely confirmed
experimentally, as regards single shock waves.
Furthermore, corresponding mathematical predictions are
possible in the case of double "regularly" reflected shock
waves (see Fig. 4a) and of "triple shocks" or "Mach Y"
waves (see Fig. 4b), provided it is assumed that in each rele
vant sector (1,2,3,4) physical variables assume limiting values
near the singular point at the shock "vertex." Again, many
of these predictions are confirmed experimentally, so that the
theory has a highly respectable status.
However, in the case of "weak" shocks, regular reflections
occur for angles of incidence somewhat greater than those
permitted by theory, and the predicted limits for triple shocks
seem to differ grossly from those observed. This discrepancy
24
§12] kopal's paradoxes

was apparently first discovered by J. von Neumann (1945);

in spite of various attempts, based on the idea that this is a
singular point paradox no satisfactory explanation of it has
yet been advanced.26

12. Kopal's Paradoxes of Linearized Projectile Theory

Time does not permit me to give the details here of the

interesting and useful "linearized theory" for forces on slender
projectiles moving at supersonic speeds, developed by von
Karman and Moore.26 The basic approximation is the
assumption that velocities are perturbed relatively little from
the free stream velocity. An analogous theory exists for wing
forces (see also Chapter III, Section 15).
Now in the case of a slightly yawed cone-headed projectile,
a comparison with a more rigorous calculation, based on the
exact Taylor-Maccoll theory (Chapter IV, Section 4) for
cones with zero yaw, is possible.
The linearized equations originally proposed by Karman,
Moore, and Tsien (op. cit.) underestimate by ten to fifty per
cent the head drag and pressure coefficients deduced from the
exact equations in case of moderate velocities (technically,
when the projectile head is much more slender than the "Mach
cone" starting from the projectile tip). Moreover in the
case of yawed cones, the cross-forces predicted by the per
turbation methods constituting the "linearized theory"
decrease with increasing Mach number, whereas those pre
dicted by perturbation of the Taylor-Maccoll exact solutions,
increase with increasing Mach number.
These discrepancies were revealed clearly by laborious
calculations made during the late war. They are purely
mathematical paradoxes due to lack of rigor in computing the
" A comprehensive review of the subject has been given by H. Polachek
and R. J. Seeger, "On shock-wave phenomena: interaction of shock waves
in gases," Proc. Symposia in Applied Math., Vol. I, pp. 119-144, Am.
Math. Soc, 1949. See especially p. 131 and the references given there.
See also W. Bleakney and A. H. Taub, Revs. Mod. Phys., 21 (1949),
584-605.
** Th. von Karman and N. B. Moore, "Resistance of slender bodies,"
Trans. Am. Soc. Mech. Eng. (1932), p. 303; H. S. Tsien, Jour. Aer. Sci.,
6 (1938), p. 480.
25
 HYDRODYNAMIQAL PARADOXES [CH. I
relative orders of magnitude in the asymptotic approximations
("perturbations") made, and from failure to determine the
order of magnitude of the effect on the resulting calculations.
They show up strikingly the insecure foundations of the
asymptotic theory of partial differential equations, alias
"perturbation theory." Recent new "linearizations" of the
equations of supersonic flow have attempted to rectify the
situation.27

13. Viscous Incompressible Fluids

I now come to paradoxes involving the theory of viscous

incompressible fluids. In this connection, I recall that the
Equations of Motion for a viscous compressible fluid have
never been proved correct, and that, according to Paul
Duhem,27* the supposition of an incompressible viscous fluid
is thermodynamically incompatible with the concept of a fluid.
However, these remarks need not concern us further.
In an "incompressible" viscous fluid, the equations of con
tinuity and state reduce to the single condition

(16) p = p0, or equivalently div u = N — - = 0.

These are combined with the Equations of Motion (3*). The

only kinematical limitation imposed by these equations is that
curl (grad p) = V X (Vp) = 0, which, if g = VG, reduces to

= vV2£
(17) + (frV)u. R = curl u]
^
The boundary conditions are

(18) u = 0 on any fixed rigid surface,

instead of merely that the normal velocity component (8) is zero.

It is commonly (but erroneously) believed that the theory
of viscous fluids is free from paradoxes. One reason for this
27
E. V. Laitone, Jour. Aer. Sci., 14 (1947), 631-642; also Brod-
See
erick, Quart. J. Meek. Appl. Math., 2 (1949), 98-120 and 121-128;
M. J. Lighthill, Jour. Aer. Sci., 16 (1949), p. 75. The perturbed Taylor-
Maccoll theory was first developed by A. H. Stone.
"• Traiti d'energitique, Vol. 2, p. 131, Paris, 1900.
26
§14] TURBULENCE I.N PIPES

is perhaps that the drag coefficient * an often be plotted as a

function CD(R) of the Reynolds number alone (cf. Chapter
III, Section 13, and Figure 1). Other experimental confirma
tions are found in the theory of lubrication, and in some parts
of the theory of turbulence.
In support of the view that the paradoxes of fluid mechanics
are due to an unjustified neglect of viscosity, the following two
mathematical remarks are often made. In non-viscous flow,
(18) cannot be satisfied, except in the trivial case of no motion;
hence the real boundary conditions are completely altered,
in the approximation of zero viscosity. Again, the term of
highest order in the Equations of Motion is suppressed, com
pletely altering the type of these differential equations.
One can easily see that the preceding criticisms have some
merit, by considering the solutions of the ordinary differential
equation ty" + y = 0, as « —> 0. If one uses the boundary
conditions y(0) = a, y(1) = b, one gets an interesting contrast
between the cases < f 0 and e J. 0.
Nevertheless, I think they are inconclusive. By "sucking
away the boundary layer" one can approximately realize
the flow patterns predicted by the theory of non-viscous fluids.
And clearly, with a sufficiently thin boundary layer,28 (18) is
approximately equivalent to the weak assertion that the
normal component of velocity vanishes. I therefore think
that the real question
is,

why does separation of the boundary

layer occur? Thus concerns the stability of nearly non-
it

viscous flows.
shall now show that similar problems of stability arise in
I

the theory of (incompressible) viscous fluids.

The Paradox Turbulence in Pipes

of

14-

Consider the steady flow of a viscous incompressible fluid

through an infinitely long straight circular pipe of radius c.
More precisely, suppose we use the following test, proposed by
28

Harold Jeffreys, for the validity of approximations: substitute the solu

tion of the approximate equations in the exact equations, and estimate
the size of the terms neglected. If small, we can introduce shear
is

a
v

flow in very thin boundary layer, by means of very small forces;

a

physically, this can be done by "sucking away the boundary layer." It

clear that the preceeding test gives no information about stability.
is

27
 HYDRODYNr\MICAL PARADOXES [CH. I

Let x denote distance allel to the pipe, and r distance from

pa?
the axis of the pipe. Let further «„ ur, and ue be the axial,
radial, and angular components of velocity, in cylindrical
coordinates. We have the following paradox.
Turbulence Paradox. The only flows possessing the sym
metry of the hypotheses of steady viscous flow in a circular
pipe are

(19) ux = a(c2
— r2),
ur = w» = 0

Such "Poiseuille flows" are observed if R < 1000, but not

usually if R > 10,000.
Proof. The hypothesis of "steady flow" means that u
should be a function of r, x, 6 alone, and not of t. The state
ment of the problem is invariant under reflection in every
plane through the pipe axis; the flow has this symmetry
only if ue = 0 and ux = f(x,r), ur = g(x,r). The statement
is also invariant under translations along the pipe axis; so
are conditions (16)— (18) since v is assumed independent of
pressure; the flow has this symmetry only if uz = f(r) and
ur = g(r). By (16), this implies wr = 0 (equivalently, the
stream function is a function of r alone).
We now use (3*). Since m = u3 = 0 identically, it gives
p = p(x). Now considering the x-component of (3*), we get

p'(x) «MV»«,
(20) -M|/"(r)+ffij.
The right-hand side of (19) is independent of r, since the
left-hand side is. Hence (rf)' = rf" + /' = kr for some
constant k, and rf = }^kr2 + C. This gives finite uz = f(r)
at r = 0 only if C = 0; hence/' = \ikr and/(r) = 3^Ax2 + b.
To satisfy condition (18) of adhesion to the pipe walls, we
must have ux = a(c2 — r2), completing the proof of (19).
Experimentally, though (19) is usually observed if R < 1000,
it commonly ceases abruptly near R = 2000. This is shown
in Fig. 17, p. 101; see also [12], p. 31; [14], p. 246. Above
R = 2000 the flow becomes turbulent (see Fig. 4b), and the
instantaneous velocity field has neither the temporal nor the
spatial symmetry of the hypotheses.
28
§15] OTHER SYMMETRY PARADOXES

It
may also be observed that the experimental facts just
noted seem to contradict the Principle of Least Action, since
Poiseuille flow involves much less expenditure of energy than
turbulent flow.

15. Other Symmetry Paradoxes

Itis most peculiar to have Nature's love of symmetry

suddenly cease to be verified physically near R = 2000. This
is reminiscent — though the limit is far less precise29 — of
Nature's abhorrence of a vacuum, which is limited to thirty-
five feet of water. Moreover there are other cases in hydro
dynamics where an apparent symmetry of causes fails to pro
duce symmetric effects.
One such case has been carefully studied by Sir Geoffrey
Taylor in a classic paper.30 Consider a viscous liquid between
two long coaxial cylinders, rotating with constant angular
velocities W and W in opposite directions. The problem
described has (approximate) symmetry under translation
along and rotation about the axis of the cylinders, and is
independent of time. There is only one solution of the
Navier-Stokes equation (3*) and (18) which also possesses
these symmetries; it is called "Couette flow." At low speeds
Couette flow is observed. But it is replaced at high Reynolds
numbers by an unsymmetrical flow which
is,

however, not
turbulent the symmetry in time preserved, through spatial
is
:

symmetry not.
is

Again, consider small air bubble, rising in water under its

a

own buoyancy. Under surface tension, will be nearly spher

it

—
ical and in any case there no cause which not symmetric
is

is

about a vertical axis through the centroid of the bubble.

Hence, by symmetry, the bubble should rise vertically. Yet
striking fact31 that > 50, the bubble ascends in
is

R
it

if
a

vertical spiral instead!

Nature's abhorrence of
*9
vacuum susceptible to some diurnal
is
a

variation; her desire for a symmetrical Poiseuille flow influenced by

is

wall roughness and initial turbulence; cf. infra.

,0
"Stability of viscous liquid contained between two rotating
a

cylinders," Phil. Tram., A$23 (1922), 289-343.

81
See Nisi and Porter, Phil. Mag., Jfi (1923), p. 754, also [50], pp.
23-24.
29
 HYDRODYNAMICAL PARADOXES [CH. I

An analogous phenomenon occurs in the wake behind a

cylinder pulled broadside through a stream. It is not sym
metrical if R > 50, but consists of alternating vortices
(B^nard-Karman vortex street).
The above examples might seem at first sight to contradict
Leibniz's metaphysical Principle of Sufficient Reason,32 that
a symmetry of causes must be preserved in the effects. A
deeper view is that, although symmetric causes must produce
symmetric effects, nearly symmetric causes need not produce
nearly symmetric effects: a symmetric problem need have no
stable symmetric solution. This possibility is the real source
of Symmetry Paradoxes.
In support of this view, it may be noted that even with
turbulent flow in circular pipes, the mean velocity u is found
to satisfy ur = He = 0, ux = F(r). Whereas in pipes not
having circular symmetry, ur ?* 0 ([14], p. 267); hence the
symmetry arguments of Section 14 are physically relevant.
But we do have to realize that, in order to be sure of verifying
this experimentally, we must first prove that the symmetry
is stable.
Now, wherever it is necessary to analyze the stability of a
deduction, since an analysis of the stability of a mathematical
deduction is far more complicated than the deduction itself,
it is almost certain that the deduction will be made long before
its instability can be tested. From this principle, we may
anticipate a continuing stream of symmetry paradoxes.
Furthermore, in the present instance, it is by no means
clear that Poiseuille flow is unstable in the usual sense, for
"infinitesimal" disturbances. The available mathematical
evidence for the plane analog of Poiseuille flow seems to
indicate that it is stable for such disturbances.33 Moreover

32 A modern formulation of this has been given by George D. Birkhoff

[57]. It is curious that formal systems of mathematical logic seem to
have dispensed with this principle, whose applicability to physics was
observed in 1894 by Pierre Curie ([4], p. 119).
33
See J. L. Synge,
" Hydrodynamical stability," Am. Math. Soc. Semi
centennial Publications, Vol. 2, 227-269, New York, 1938; H. B. Squire,
Proc. Roy. Soc., A14B (1933), 621-627; C. L. Pekeris, Phys. Rev., 74
(1948), 191-199. For recent physical data, see MM. Fortier and
Comolet, La Houille Blanche, numdro special B (1949), p. 673.
30
 §16] THE EIFFEL PARADOX

this evidence is in some sense confirmed by the fact that the

onset of turbulence can be greatly delayed by using sufficiently
smooth walls and rounded entrances ([12], Vol. 2, p. 33;
[14], p. 172), up to R = 75,000. Thus Poiseuille motion may
well be stable locally, but unstable in the large.
From all these facts, one begins to get an idea of the diffi
culties and confusion which surround the foundations of hydro
dynamics, and of the uncertainty of symmetry arguments !

16. The Eiffel Paradox

Another paradox, related physically to the Turbulence
Paradox, was found in 1912 by Constanzi and Eiffel.34 It
may be stated as follows.
Eiffel Paradox. Near a "critical Reynolds number" Re =
150,000, the fluid resistance experienced by a sphere actually
decreases as the velocity increases.
This contradicts our general physical experience; thus it is
a purely experimental paradox. It contradicts even more
flagrantly the result "proved" by dimensional analysis in
many college textbooks, that fluid resistance is proportional
to the velocity at low speeds, and to the velocity squared at
higher speeds (see Chapter III, Section 3).
It is especially astonishing if we recall that R is generally
supposed to express the "ratio between inertial forces and
viscous forces" (cf. [12], Vol. 2, Chap. I; and Chap. Illinfra).
How can the change in this ratio, from 10-6 to 0.5 X 10-6,
cause such a striking effect? Thus, unlike the paradoxes
discussed in Section 9, the Eiffel paradox seems to be a genuine
Paradox of Approximation: a minute decrease in viscosity
causes a large change in the actual flow, separation is delayed,
and the wake becomes narrower.
Of course, the cause is now well-known: the change induces
transition of the boundary layer from laminar to turbulent
flow! This transition is clearly visible in good spark photo
graphs. Thus the Eiffel Paradox is associated with the sta
bility of laminar flow, and hence to the discussion of Section 14.
It seems to me that this example throws great light on the
shortcomings of the conventional theory of partial differential
" The references are given in [12], Vol. 2, p. 98.
31
 HYDRODYNAMICAL PARADOXES [CH. I

equations. Certainly, the classical asymptotic theory of

partial differential equations ("perturbation theory") fails
utterly to predict a sudden change in topological nature of
the whole solution when one coefficient is changed from
10-s to 0.5 X 10-s. Hence it would seem that no treatment
of "slightly viscous" fluids which is based on this theory,36
can come close to describing the real behavior of such fluids.
It will never predict a turbulent wake, a vortex street, or
"transition" at Rc.
Instead, what is needed is an attempt to deduce, from o
priori mathematical considerations, the possibility of sudden
changes in regime induced by small changes in the coefficients
That

is,
of partial differential equations. mathematics must
take advantage of physical inspiration, as have already

I
stressed (Section 1).
The phenomenon of turbulence raises interesting questions
regarding the differentiability of the functions Ui(xi,xt,x3;t).
It seems well established that the w< are to some extent random
functions. And by Norbert Wiener's classic theorem, "almost
all" purely random functions are non-differentiable.36 This
fact has led some scientists to believe that one must go back
to (discontinuous) molecular considerations in fluid mechanics,
in dealing with turbulence. However, the fact that Rc the

is
same in liquids as in gases, makes this doubtful. One recent
guess, based on the statistical theory of turbulence, that the
is

ut have fifth but not sixth derivatives.

17. Stokes' Paradox

One of the great triumphs of theoretical hydrodynamics,

is

Stokes derivation of the formula

(21) D = &*ncv

Such as, for example, the work of the school of Oseen [11], which has
3•

been so useful in the opposite asymptotic case, — 0. The phenomenon

R
>

has no analog in compressible flow, though the drag of a slender-nosed

projectile has a near-discontinuity at M — (see Karman and Moore,
1

op. cit., supra).

" See L. Doob, Annals Math., 43 (1942), 351-369, and its bib
J.

of

liography. The principle that solutions of the Navier-Stokes equations

will tend not to be analytic, was deduced by P. Duhem on other grounds
in p. 121 of Vol. of his TraiU energttique.
d''
2

32
§17] stokes' paradox
for the resistance D encountered by a rigid sphere of radius c
moving with constant speed v in a fluid of viscosity. This
formula, which is verified experimentally37 for Reynolds
numbers R < 0.2, may be deduced from conditions (16)-(18),
by neglecting the "inertial" term in du/dt, and assuming
that the flow is steady and has axial symmetry — i.e. that no
Symmetry Paradox is involved. This simplification is said
to correspond to "creeping flows."
It seems reasonable to make the same simplifications in two
dimensions. However, if this is done, we are confronted with:
Stokes' Paradox. No steady "creeping flow" of a viscous
fluid past an infinite circular cylinder is possible.
Rough Proof (I know of no rigorous proof). By (3*),
neglecting du/dt and g, we get

(22) V^^A^=MV.(^)-0;
hence p is harmonic. Again, "by symmetry" (!), wi(— xi,x»)
= «i(xi,xs), while U2(— xi,xt) = — U2(xi,x2); hence the total
flux of momentum out of a large circular cylinder concentric
with the small cylinder, is zero. Hence, by the momentum
equation (cf. Section 6), the pressure-thrust on the two
cylinders must be the same. Since p is harmonic by (22), we
infer p — pK = A cos 0/nr + 0(l/r2), and so by (3*),

_,, dp A sin 20 .

•0>
But if can be expanded in a power series at infinity, with
U2
m2— > 0, we must have V2w2 = 0(l/r3); hence A = 0 and we
have a contradiction.
Oseen and Lamb38 have rationalized Stokes' Paradox as a
singular point paradox, pointing out that at very large dis
tances the acceleration terms dominate the viscosity terms
and hence should not be neglected. By definition, it can also
be regarded as a paradox of overdeterminism.
3*
[12], Vol. 2, p. 100. If the sphere is in a gas where the mean free
path is comparable with c, a correction must be made for "slipping";
cf. [6], Vol. 2, Appendix.
" [11], p. 162; [9], p. 614; see also [13], Vol. VI, pp. 29-40.
33
 HYDRODYNAMICAL PARADOXES [CH. I

Stokes' Paradox is analogous logically to the paradox of a

two-dimensional underwater explosion. Though there is a
simple and extremely useful theory of the underwater oscilla
tions of spherical bubbles,39 it is easily shown that any two-
dimensional notion of a bubble involves infinite kinetic energy,
with U = log r + 0(l/r). Hence no two-dimensional bubble
oscillations can be induced by finite forces. The paradox of
a two-dimensional underwater explosion is clearly also a
singular point paradox; at very large distances the effect of
compressibility in absorbing the expansion is dominant.

18. The Rising Bubble Paradox

Even in the case of spheres, the theory of "creeping motions "

leads to a curious paradox. Let a small air bubble rise under
buoyancy in a liquid, the bubble being so small that surface
tension keeps it nearly spherical. Assume as above the
approximation of "creeping motion," that du/dt = 0 in (3*).
Since the bubble is fluid, it is logical to replace (18) by

(18') u is continuous across the surface of the drop,

as pointed out by Rayleigh.

The mathematical problem defined in this way was solved
by Riabouchinsky and Hadamard,40 who showed that the
theoretical drag was given by the formula
,„ ,.
(21 ]
n
D _ Qnccvy.(2y. + 3m') •
(37+37)
In (21'), n is the viscosity of the medium, and that of
is
\i!

the bubble moving through it.

So much for theory; experimentally, the drag seems to be
given by (21) instead of (21') (cf. [12], vol. p. 115, or [14],
2,

That
is,

p. 215). the bubble seems to behave as were

it
if

"[9]' p• 122. An analogous paradox in the theory of heat conduction

obtained one looks for steady temperature distribution in the
if
is

plane, maintained at 100° on the circumference of circle, and tending

a

to 0° at infinity. See also [13], Vol. pp. 29-40.

6,

40
D. Riabouchinsky, Bull. Acad. Sci. Cracovie (1911), p. 40; J. Hada
mard, Comptes Rendus, 152 (1911), 1735. The experimental data are
horoughly discussed on p. 190ff. of G. Barr, Manual Viscometry,
A

of

cford, 1931. See also T. Bryn, Forschung. Ingenw. (1933), 27-30.

4

34
§19] MAGNUS EFFECT PARADOX

rigid. This contradiction between theory and experiment

may be called the Rising Bubble Paradox.
What is the explanation of this paradox? It has been sug
gested by Bond41 that the cause is surface energy, and many
of Bond's measurements lead to (21') instead of (20). How
ever, the explanation is far from clear. It may be that a
rigid film forms at the interface, especially if impurities are
in solution, but all that one can say with certainty

is,
that the
mechanics of "real" surface behavior are far less simple than
the classical treatments would suggest.

19. "Drift"; Magnus Effect Paradox

Artillerymen have known for over century that spinning

a
bullets tend to "drift" out of the vertical plane in which they
are fired; and that the drift in the direction of rotation of
is

the top of the bullet. However, this phenomenon was not

correctly understood for many years.42
A considerable vogue was enjoyed by the following quite
fallacious qualitative "explanation" of drift, due to the great
mathematician Poisson. Due to inertia, the axis of the bullet
will lag behind the tangent to the trajectory, as sketched in
Fig. 5a. Hence there will be more pressure on the underside;
hence more friction there. Hence (cf. Fig. 5b) there will be
drift in the observed direction.
a

The fallaciousness of this qualitative explanation can be

seen easily by applying the same argument to spinning
a

tennis ball: yields the erroneous conclusion that tennis

it

ball with topspin should rise! Moreover quantitative study

a

of the gyroscopic stability of bullets reveals that the stable

position of the bullet axis (with right-handed rifling) to
is

the right of the trajectory tangent. The drift of the bullet

thus due to the aerodynamic cross-force, and only indirectly
is

to the spin.
Why Poisson's explanation inapplicable to tennis balls?
is

Since every college physics student given pat qualitative

is

"
W. N. Bond, Phil. Mag., (1927), 889-898. For modern dis
a
4

cussion of the real facts of surface energy, see N. K. Adam, The Physics
and Chemistry Surfaces, Oxford, 1931, esp. Chapter II.
of

41 An historical account may be found in

[3], Chapter X.
35
 HYDRODYNAMICAL PARADOXES [CH. I
explanation of the Magnus effect, I do not think I need to
say. Instead, I should like to point out that the usual explana
tion is also dubious, in view of the following paradox.
Magnus Effect Paradox. At low rates of spin, the actual
deflection is in the opposite direction to that predicted by the con
ventional explanations.t3
Actually, possible qualitative explanation of the Magnus
a
Effect Paradox can be based on Prandtl's boundary layer
theory. Due to rotation, the low pressure "wake" is shifted

/
Less Pressure

Direction
•♦ of
Deviation

More Pressure ~.
ID
Fig. 5. Fallacious explanation of "drift."

up by the topspin (cf. [12], Vol. 2, Plates 7-9), and the stag
nation point down; this would cause a lift.
I think that Poisson's fallacious explanation and the Magnus
Effect Paradox both illustrate the unreliability of qualitative
explanations. One can argue cynically that qualitative deduc
tions have a fifty per cent chance of appearing verified, no
matter whether they are correct or not! I think that students
should be taught not to set much store by them; they are very
corrupting to one's critical sense.

20. A Role for Mathematicians

What is the moral of the preceding paradoxes? Certainly,

they show the necessity of comparing theory with experiment,
° See [6], p. 504; the data are due to Maccoll.
36
§20] THE MATHEMATICIAN

and consequently show the utter falsity of the following widely

quoted statement of Eddington: "An intelligence, unac
quainted with our universe, but acquainted with the system
of thought by which the human mind interprets to itself the
content of its sensory experience, should be able to attain
all the knowledge of physics which we have attained by
experiment.44
To my mind, they also demonstrate the need which the
theory of partial differential equations has for fresh physical
inspiration. Why should a small change in viscosity affect
fluid flow drastically, when a small change in compressibility
does not?
On the other hand, they seem to me not to support the
impression, widely current among engineers and physicists,
that mathematical deduction should be supplanted by more
"physical" reasoning. Indeed, the free use of approxima
tions, topological oversimplifications, and non-rigorous sym
metry considerations, which were seen to underly these
paradoxes, are more typical of "physical" than "mathe
matical" reasoning. I think the very phrase "physical
reasoning" implies a confusion of deductive and inductive
considerations, which corrupts the fundamental meaning of
the word "theory." This corruption is readily recognized
by the confusion of students when the "theory" is taught,
and by the unreliability of predictions when the "theory" is
applied. Thus it is all too common to read of "satisfactory"
agreement between theory and experiment, when the agree
ment should give little satisfaction to any true scientist.
I agree that oversimplifications based on the "right"
approximations, made possible by familiarity with experi
mental results, greatly accelerate technological and even
theoretical progress. Such oversimplifications led to the con
cept of an "ideal fluid," which is still very useful. But I
think that mathematicians can perform a useful service if they
will analyze critically these oversimplifications, by the deduc
tive method, and so establish their limitations more clearly.
44
Sir Arthur Eddington, Relativity Theory of Protons and Electron!,
Cambridge, 1936, p. 327. They might also give pause to social theorists
who rely exclusively on deductive logic in the infinitely more complicated
field of social relations.
37
 HYDRODYNAMICAL PARADOXES [CH. I
I realize that this service will often be deprecated by those
"scientists" who use science instead of building it. But it is
encouraging to know that its value was recognized by no less
a person than Lord Rayleigh. Apparently Maxime B6cher
had written Rayleigh, questioning the statements about
Sturm-Liouville expansions in Section 187 of the latter's
celebrated Theory of Sound. Rayleigh replied,46 with charac
teristic candor: "I doubt if I had any special method in my
mind when I wrote Section 187. At any rate, if I had then,
I have not now. I think I took this case as covered by the
general theory as in Chapter IV. I have often wished that
mathematicians would devote themselves to improving the
rigor of this and similar demonstrations, rather than to invent
ing special methods for the particular cases."

21. Mathematics as Reliable as "Common Sense''; Conclusion

Moreover mathematics is not only useful in "tidying up"

reasoning, whose broad lines were suggested by physical
intuition. In many cases, mathematical deductions will
rapidly give results which are at variance with physical
intuition, and which it would be extremely laborious to derive
experimentally. I shall close by giving two examples of such
results.
Experimental data must frequently be corrected for "wall
effects" which are much larger than common sense would
suggest. In principle, wall effects can be determined empiri
cally by removing the wall to larger and larger distances, but
this may be prohibitively expensive in practice. Thus, Laden-
burg has shown that a sphere falling slowly in a cylindrical
tube of viscous liquid, having 100 times the cross-section area
of the sphere, encounters 20 per cent more resistance than if
there were no walls.46
Finally consider the celebrated Taylor-Maccoll supersonic
flow past a conical projectile (Chapter IV, Section 4). This

" The letter is in my possession.

46 B. Ladenburg, Annalen der Physik, 23 (1907), 447; A. W. Francis,
Physics, 4 (1933), 403-406. Another surprisingly large wall effect,
arising with cavity flows, is discussed in Chapter II, Section 11.
38
§21] MATHEMATICS AND COMMON SENSE

flow has the peculiar property of having streamlines at con

stant pressure along the cone wall. It is well known that,
according to the mathematical theory of non-viscous fluids,
the region behind any streamline at constant pressure can be
replaced theoretically by motionless fluid, as in the Kirchhoff
wake theory (Chapter II). Hence, mathematically, super
sonic flow past a flat disc is possible, in which an invisible
conical air barrier separates the onrushing air from the disc.
This flow pattern is instantly rejected by intuition and
common sense, as absurdly unstable. The rejection is on the
same basis as the rejection of the reverse of the Kirchhoff
flow with wake, i.e. as the rejection of a (mathematically
possible) upstream wake. It seems highly plausible that the
effects of an obstacle should lie on the downstream side of the
cause.47
Yet in this case common sense seems to be more fallible
than mathematical deduction! Spark shadowgraphs, taken
at the Ballistic Research Laboratories in Aberdeen (see
frontispiece), indicate that a very thin needle in front of the
disc is actually capable of inducing something quite like this
"absurd" flow.
In view of these facts, I think the mathematician need have
no inferiority complex about the soundness of purely deductive
reasoning (as contrasted with "physical reasoning") in fluid
mechanics, provided he clearly understands the preceding hydro-
dynamical paradoxes.

47 The same intuition makes retarded potentials give the "physically

preferred" solutions of the wave equation, and is the basis of selecting
the corresponding solutions of the linearized equations of supersonic flow.
In the case of shock waves, this selection is bolstered by an appeal to
the Second Law of Thermodynamics; in some sense, this law also seems
relevant to the development of turbulent disorder in the wake behind an
obstacle.

39
 CHAPTER II
Recent Progress in Free Boundary

Theory
1. Wakes

The classical theory of Dirichlet and Joukowsky flows con

cerns fluid motions past obstacles whose location in space is
supposed given. A unique and fascinating chapter in fluid
mechanics concerns flows whose boundaries are not given, but
must be determined from a knowledge of the pressure along
them. Today I shall discuss some recent developments in
the theory of flows having such "free" boundaries, with
especial reference to their physical validity. Because of
this emphasis, what I shall say will overlap very little with
surveys of the field by von Karman [27] and Weinstein [34],
[34'].
I shall begin by a historical resume• of the subject.1 Free
boundary theory was initiated in 1868 by Helmholtz [26] and
Kirchhoff [27']. At that time, it was well-known that the
"Dirichlet flow" (Chapter I, Section 5) around a flat plate
held broadside in a moving fluid, has the symmetrically dis
posed streamlines sketched in Figure 6a. This flow gives rise
to the two paradoxes of zero resistance (Chapter I, Section 6),
and of infinite velocity (hence infinite negative pressure) at
the sharp edge.
Kirchhoff had the inspiration to suggest the alternative
flow sketched in Figure 6b. He suggested that only part of
the flow would be a Dirichlet flow. This part would separate
from the plate at its edges S, S', leaving a static wake of "dead
water" behind the plate, extending to infinity. In this model,
the velocity evidently changes abruptly across the boundary

1
The present material was also presented in large part at the Sym
posium on Applied Mechanics at Brown University, August 1947, and
at the Seventh International Congress on Applied Mechanics in Sep
tember 1948.
40
§1] WAKES

of the wake, giving rise to a surface of discontinuity. Helm-

holtz further assumed that this surface consisted of stream
lines, which may thus be called streamlines of discontinuity.
In a non-viscous fluid, such a streamline of discontinuity
would give rise to no (viscous) shear forces; for dynamical
equilibrium one thus requires that the pressure be continuous

oo•

Dividing
co
Streamline

Fig. 6. Flows past vertical plate.

across the streamline. But in the static wake (neglecting

gravity, by Chapter III, Theorem 4) the pressure is constant.
Hence it is necessary and sufficient for the streamline of dis
continuity to be at constant pressure. Such a streamline is
called a free streamline. By Bernoulli's Theorem (Chapter I,
Section 6), since vdv = —dp/p = 0, velocity is constant on any
such free streamline.
Hence the velocity is finite at S, S', and the paradox of
infinite velocity is avoided. We shall see shortly (Section 2)
41
 FREE BOUNDARY THEORY [CH. II

that the d'Alembert Paradox of zero resistance is also avoided.

Moreover separation does occur physically at the plate edges
S, S'. Finally, the model of a "static wake" accords with
our qualitative experience that one can find shelter from the
wind on the leeward side of a barrier. Thus the Helmholtz
model of Figure 6b represents, in the case of a flat plate at
least, enormous progress over the Euler-Dirichlet model of
Figure 6a.2

2. Mathematical Model

Following Kirchhoff (cf. [27'] or obtain

[10], p. 299) one can
a complete mathematical description of the flow of Figure 6b,
as follows.
The velocity potential U(x,y) of a locally irrotational plane3
flow is harmonic by Chapter I, Section 5; hence U(x,y) is
the real part of a complex function W = U + iV of the com
plex position coordinate x + iy. The function W so defined
is called the complex potential, and V the stream function (it is
easy to show that V = const, on streamlines). Moreover
dW/dz is the complex conjugate f = ui — iu2 of the complex
velocity vector Wi + iu*.
The conditions that the flow be irrotational and volume-

f-
preserving are thus precisely that

r.

in the sense of complex differentiation. Hence the functional

relationships between the complex variables W, f , and z are
analytic in the usual sense, and are therefore expressible by
conformal transformations. We consider the transformation
of the f -domain (or "hodograph") onto the W-domain: this
may be determined in the present case as follows.

*
In line with Chapter I, it is however interesting to recall the Hada-
mard Paradox, whereby an (analytic) surface of discontinuity cannot
be generated (J. Hadamard, Lemons sur la propagation des ondes et les
Aquations de I'hydrodynamique, Paris 1903, Note II). See also F. Klein,
Zeits. Math. Phys., 58 (1910), p. 259.
3 That flow in which the velocity vector always parallel to fixed
is,

is

plane, so that dU/dz = and U = U(x,y).

0

42
§2] MATHEMATICAL MODEL

Let the origin be placed at the stagnation point 0 in the

z-plane, and the real axis along the plate SOS'; thus we imagine
Fig. 6b rotated through 90° clockwise. Clearly f is real
along the plate, while

(2) |f | = |f(°°)| = const, on the free streamline.

Hence, under the simplest topological assumptions, the hodo-

graph is a semicircle, as in Figure 7. By proper choice of
units, |f(°°)| = 1, and so K00) = e<"• Thus the hodograph
is a unit semicircle, centered in the origin. In the present
case of normal incidence, a = r/2 and so f(°°) = i, as in
Figure 7.
oo

J Fio. 7. Hodograph of Kirchhoff flow.

The JF-domain is clearly a plane, cut along the positive

real axis, setting W(0) = 0 by choice of the constant of inte
gration in (2).
By elementary conformal transformations, we can map the
hodograph onto the W-domain. Clearly — y/W covers the
lower half-plane; a = 3^(f + 1/f) maps the hodograph onto
the lower half-plane. Hence -
\/W = (Aa + B)/(C<r + D),
with A, B, C, D real and AD ^ BC, also is a one-one con-
formal ("schlicht") transformation mapping the hodograph
onto the W-domain. And it is a classical theorem of con-
formal mapping that there are no others.
Further, in the present case, W = y/W = 0 implies f = 0
and a = °° ; hence A = 0, and we can write — s/W =
l/Ca + D). Finally, W = oo implies J• = i and so <r = 0;
hence D = 0 in the case of normal incidence. Substituting,
we get
43
 FREE BOUNDARY THEORY [CH. II

^ -A- c
cW
[fti]'' if c =
4

All choices of the constant c in (3) give similar flows; a typical

solution is given by setting c = 1. In this case, the compu
tation of z(f) reduces by (1) to the integration of a rational
function. Performing this integration, we get

-
(4) z =
wh? + \ {fti i ln ml + const-

for all values of f = { + it).

Along the plate, f is real and (4) reduces to

(4') « =
^rqrip +
5 j^qn + arctan fj .

where clearly z(0) = 0, and so the constant of integration

vanishes. At the right-hand separation point S, f = 1 and so

The pressure by setting f = £ =

is most easily computed
tan 9 along the plate, so that f/(f2 + 1) = sin 0 cos 0 =
by Bernoulli's
is,

3^ sin 20. Hence the thrust on half the plate

Theorem and (1),

(5)
r/4
cos2 26d6 =
2

t
4

=
0
9

The ratio of half the thrust to the plate half-length clearly

is

the drag coefficient, which thus

is

(6) CD = —2ir
-r- jt = 0•88 approximately.
t

[ir •+• 4)

Similar but more elaborate complications give, for the case of

oblique incidence a 5^ x/2,

.... „ 2ir sin2 a n sin 2a

tc

+ sin a' + ir sin a

4

7r

44
§3] BORDA MOUTHPIECE

Details may be found elsewhere ([9], pp. 304-

p. 103; [10],
305; [13], Vol. I, p. 287). Figure 2c shows the flow for the
case a = 15°.

3. Borda Mouthpiece

A similar procedure gives a plane jet from an infinite straight

orifice (Borda tube), as sketched in Figure 8a; the dotted
area is supposed to consist of static air. The problem is
mathematically simplified if we consider only the half of the
flow above the axis of symmetry; in this form, the problem
was solved by Helmholtz in 1869.
If we take the axis of symmetry as the real axis, and choose
units so that the jet velocity is one, we again have a semi
circular hodograph as in Figure 7. Hence a = 3^(f + 1/f)
again fills the lower half-plane. The difference is that the
present flow is bounded by two streamlines which (by choice
of units) we can take to bound the strip 0 ^ V ^ x in the
TT-plane. As before, we thus get ew = (Aa + B)/(C<r + D).
We again calculate A, B, C, D by the method of undeter
mined coefficients. Thus when W = — °° , and ew — 0, f = 0
and so <r = °° ; hence A = 0 and we can set B = 1 as before.
Whereas at W = + °° , ew = °° and f = a = 1 ; the asymp
totic jet velocity; hence D = —C. Substituting, we get

- ,-rh -
m C"
Hfm - (F^IP [Crea"

We get in this way a family of flows, all similar to the flow

(7') W = In f - 2 In (f - 1) + const.
The determination of z then reduces to

(8)
1)

= -i+ 2 In f - 2 In (f - 1) + const.

as can easily be shown using partial fractions. We can make

z real when f j 0 through positive real values, by making the
integration constant 2 In ( — 1), so that (8) becomes
45
 FREE BOUNDARY THEORY [CH. II

(8') z = --+ 2 In
y^j = - ±
+ 2 In f - 2 In (1 - f)
Again, as through negative real values at °° *, y =
ff 0
— > ir.
Hence the walls of the Borda mouthpiece are 2ir
7m(z)
units apart. Finally, as f = eie — > 1, along the free stream
line Im( — 1/f + 2 In f) —> 0, while the imaginary part of
In (1 — f) approaches x/2. Hence <fte jet width v is half the

<z
Axis of 5ymmetr>

c
Piston Container

Fig. 8. Borda flows through slot and circular orifice.

46
§4] FURTHER GENERALIZATIONS

orifice width. In other words, the "coefficient of contraction"

of the orifice is one-half.
The coefficient of contraction (but not the velocity field)
can also be obtained by simple energy and momentum con
siderations, as follows (cf. [12], Vol. I, p. 242.) Let a con
tainer with vertical walls be filled with a liquid of density p,
into which is inserted a horizontal "Borda tube" as in Figure
8b, of arbitrary cross-section shape and area A ; let the pressure
head at the level of the tube be p. We assume that the flow
separates4 from the tube at its inner end, and that the jet
from the tube tends asymptotically to a constant speed v;
this must be the constant "free streamline" velocity on the
jet boundary. Let A* be the asymptotic jet cross-section;
then A* / A is by definition the coefficient of contraction. We
compute it as follows.
Per unit time, the volume efflux is vA*; the momentum
efflux is v2A*; the kinetic energy efflux is }ipv3A*. The
momentum supplied is however pA (pressure excess); the
(potential) energy p(vA*). Hence pA = pv2A* and pvA* =
}^pv3A*. Dividing v times the first equation by the second,
we get the result
A * 1

This conclusion is not restricted to plane flows, but is clearly

equally valid for a circular tube.

4. Further Generalizations

Many other models of wakes and jets with free streamlines

have been constructed the time of Helmholtz and
since
Kirchhoff. They are fully described in the classical litera
ture,6 which has become known as the theory of wakes and
jets.
A large proportion of these flows involve cases where the
diagram in the f -plane, or hodograph, is a circular sector. It
* This is the case of "free flow"; see A. H. Gibson, Hydraulics, fourth

ed., p. 122; it is not always fulfilled.

* See
[9], pp. 94-109; [10], Chs. XI-XIII; [25], [31], and [32]. A com
plete survey of the subject is being prepared by the author and E.
Zarantonello under the title "Jets, wakes, and cavities."
47
 FREE BOUNDARY THEORY [CH. II

has recently been shown• that z(f) can be expressed explicitly

in terms of elementary functions, whenever this circular sector
subtends a rational angle (i.e. one commensurable with 360°).
Indeed, in this case a = f m/n can be made to cover the upper
half of the unit circle for some pair of integers m, n

|f|
^
1,
(in the case of free streamlines, we easily reduce to sector of

a
the unit circle). Hence = }4(<r 1/ff), which rational

is
a
r
function of Vf' covers the lower half plane.
On the other hand, stationary flow must be bounded by

a
streamlines or dividing streamlines. Hence the diagram in
the W-plane (if simply connected) consists of plane, half-

a
plane, or strip, possibly with number of semi-infinite cuts

a
parallel to the real axis. Therefore there Schwarz-

is
a
Christoffel transformation of the ^-domain onto the W-domain
of the form dW = F(t)(1t, where F(t) rational function;

is
a
this follows immediately from the theory of Schwarz-Chris-
toffel transformations.
Expressing F(t) = F(^(fm'n r~m/B)) = •B(v/f) as
+

a
rational function of y/f, we obtain by substitution in (1)

(10)

The integrand evidently rational in \/\, whence as well

is

z,
as and W, can be expressed as sum of rational and loga
a
f

rithmic complex functions of . \/f

omit the details, which
I

show that the breaking up of the integrand in (10) into partial

fractions, can indeed be computed explicitly in all cases.
Although the procedure just outlined attractive in prin
is

ciple, involves serious practical difficulties. Unless m/n

it

the final formulas are very long. Hence, unless one

3^

or
is is

1,

satisfied with the relation between few physical dimen

a

sions, better to evaluate by numerical integration than

it
is

by partial fractions.7
Using complex numerical integration, has proved possible
it

Birkhoff, Plesset, and Simmons [20]. The case where the diagram
•

in the W-plane an infinite strip had already been covered by Mises

is

([9], Appendix of second German edition).

detailed account of some difficulties of numerical calculation has
A
7

been given by D. M. Young, "The use of conformal mapping to deter

48
§5] SEPARATION POINT

to code the calculation of flows with free streamlines whose

hodograph is a circular sector for the Mark II Calculator at
Dahlgren, so that the velocity field over most of the flow can
be computed within hours.8
Similar arguments enable one to express flows past general
polygonal barriers in terms of hyperelliptic integrals, using
co = In f as independent variable. (Clearly the hodograph
is a rectangular polygon in the co-plane.)
An even more important result is due to Levi-Civita [29],
who used the Schwarz Principle of Reflection to express the
most general flow having an infinite wake, of the topological
type discussed in Section 2, in terms of analytic functions.
By this method, one can easily construct flows, with "wakes"
bounded by free streamlines, past curved obstacles. However,
the shape of the obstacle cannot be prescribed in advance,
except qualitatively; thus the effective determination of the
plane flow or flows with "wake" past a given curved obstacle
remains an unsolved problem.8*

5. Indeterminacy of Separation Point

In the cases treated in Sections 2 and 3, the topology of the

cavity and the exact location of the points of flow separation
were assumed known. These assumptions are not possible
in general, and lead to serious problems of mathematical
indeterminacy.
Thus H. Villat proved, thirty-five years ago, that even in
the simple case of flows with infinite wake past a reentrant
wedge, separating at the edges, more than one topological con
figuration was mathematically possible;9 specifically, a dead
mine flows with free streamlines," Proc. Symposium on Construction and
Applications of Conformal Maps, Inst. Num. Analysis, Nat. Bureau
Standards, Univ. Calif, at Los Angeles, June, 1949.
8 The coding and method of calculation were planned by Prof. D. R.
Hartree, Mr. J. C. Aller, and the author; useful suggestions were also
received from Prof. John von Neumann.
** For exceptions, see S. Brodetsky, Proc. Zurich Congress Appl. Mech.

(1926), 527-531.
9
hoc. cit. on p. 21 ; see also Ann. Fac. Sci. Toulouse, 5 (1913), 375-404;
Ann. Ec. Norm. Sup., 3I (1914), 455^91; Bull. Sci. Math., 42 (1918),
72-92; Apercus thioriques sur la resistance des fluides, Paris, 1920; R.
Thiry, Ann. £c. Norm. Sup. 38 (1921), 229-339.
49
 FREE BOUNDARY THEORY [CH. II

area at arbitrary pressure can form in the reentrant vertex.

This result has recently been extended by Dr. E. Zarantonello,
who has shown that, in the case of a plate with angle of attack
a ^ 0°, exactly two topological types of flow with infinite
wake and everywhere finite velocity are possible.
A more fundamental indeterminacy arises in the case of
curved obstacles, even for flows having the topological type
discussed in Section 2. It was first noted in 1911 by
Marcel Brillouin ([22], p. 174) that "la forme de M. Levi-
Civita est encore beaucoup trop g6ne>ale." Brillouin also

Fig. 9. Two cavity flows past accolade,

suggested most of the criteria mentioned below, for avoiding
this indeterminacy. As the existence and uniqueness theory
has been summarized recently elsewhere ([34] and [34']), I
shall only state the most essential facts. These are clearly
illustrated by the following fundamental theorem of Leray
[28], culminating earlier work of Weinstein and Friedrichs (see
[34],[34'])."
Let B be any circle, or more generally any symmetric convex
barrier whose curvature is non-decreasing as one moves back
wards from the stagnation point, as in Figure 9. Such barriers
are called "accolades."11 Then there is a fixed point S0 on B,

Leray's results were extended to jets and to wakes in channels by

10

J. Kravtchenko, Jour, de Math., 29 (1941), 30-303; also Ann. Inst.

Henri Poincari, 62 (1945), 233-268. See also H. Villat, Jour. math,
pures appl., 10 (1914), 231-290; S. Brodetsky, Proc. Edin. Math. Soc.,
41 (1922-1923), 58-62.
11 By
the four-vertex theorem, the circle is the only "accolade" which
bounds a convex plane region. However, other generalizations are
possible.
50
§5] SEPARATION POINT

such that for every point S behind <S0 on B, exactly one sym
metric flow with infinite cavity (wake) is possible which sepa
rates from B at S.
By this theorem, there is a one-parameter family of cavity
flows past B with infinite wakes. The free streamlines corre
sponding to two such flows are sketched in Figure 9. The
shape of each wake is asymptotically parabolic, in the sense
that

(11) y ~ y/Cx for some C; i.e., lim*-, » —

^=
= 1.
y/Cx
The "drag" D is determined by C through the formula

(11') D=^f,
of Levi-Civita [29]. (Similar formulas for the lift and moment
of asymmetrical cavity flows were obtained by M. Brill ouin
[22].)
One of these flows is preferable esthetically, on the following
mathematically equivalent grounds: separation occurs at
(i)

<S0, (ii) the minimum pressure occurs in the cavity (i.e. at

the point of separation), (iii) the cavity convex, (iv) the
is

curvature of the cavity at the separation points finite.

is

Although mathematicians have tended to stress condition

(iv), seems to me that condition (ii) far more significant
is
it

physically. Thus physically, presumably fulfilled in the

is
it

case of vapor-filled cavity,11* or In fact,

if

<g.}4pv2.
p
a

know of only one physically plausible case12 in which not

is
it

satisfied. Also, though (ii) always implies (iii), as noted by

Brillouin, condition (iv) not usually satisfied when separa
a is

tion occurs at the end of polygonal or other barrier, whereas

(ii) is.
is,

Condition (ii) of course, not fulfilled for physical wakes

IU In ordinary liquids, the pressure cannot long remain sensibly below

vapor pressure. The contrary assertion of Rayleigh ([13], Vol. p. 401)
6,

applies only to liquids whose "gas nuclei" have been removed by boiling
or other special treatment; cf. [22], p. 151. For direct confirmation, see
[15], Fig. 34; also Section 9.
"The case described by M. Lighthill, Aer. Res. Comm. Reps, and
J.

Mem., 2105, of jet sucked around bend.

a

51
 FREE BOUNDARY THEORY [CH. II

behind curved obstacles ([6], Vol. 2, pp. 422, 497), but neither
is any other part of classical "wake theory."

6. Physical Unreality

Indeed, physical unreality is well-known to be one of the

three capital defects of the classical theory of free streamlines.
(The other two are, that it is mathematically indeterminate in
general, and can be applied only to plane flows — and prac
tically only to a few plane flows past polygonal fixed walls.)
The physical unreality of the Kirchhoff wake model has
been realized for over fifty years.13 It may be shown (using
plausible, if not strictly rigorous arguments!) that a streamline
of discontinuity in a non-viscous fluid is in unstable equilibrium.
It may also be observed experimentally that it soon becomes
wavy, and then breaks up into a turbulent "mixing zone,"
which has been discussed by various writers.14 This turbu
lence is drawn into the wake and recirculated there, so that
real wakes are by no means motionless bands of "dead water"
extending to infinity. At lower Reynolds numbers (50 < R <
106 for cylinders), they often consist of alternate vortices or
"vortex streets."16
Some idea of the magnitude of the discrepancy between the
Kirchhoff model and real wakes may be had from the fact
that the real CD of a flat plate is about 1.95 ([14], p. 227), or
twice the CD = 0.88 predicted by theory. Whereas the
theoretical wake pressure is equal to the free stream pressure,
there is a large actual wake underpressure. This wake under
pressure causes the actual "streamlines of discontinuity" to

13 Kelvin (Nature, 50, 1894, pp. 524, 549, 573, 597); also Rayleigh
See
[13], Vol. 6, p. 39. Prophetically, the instability of a surface of discon
tinuity was discussed (in another connection) by Kelvin in 1871 (Phil.
Mag., Vol. 42, p. 368); see also Rayleigh [13], Vol. 1, pp. 365, 477; F. W.
Lanchester Aerodynamics (1908), pp. 128-136.
14
See [6], Sections 252-254; also L. M. Swain, Proc. Roy. Soc., A125
(1929), p. 647; S. Tomotika, ibid., A165 (1938), pp. 53-72; L. Schlichting,
Ing. Archiv. I (1930), p. 533; W. Tollmien, ibid. 4, (1933), pp. 1-15.
16 Th. von Karman and H. Rubach, Phys. Zeits., I3 (1911), p. 49;
See
M. Teissie^Solier and P. Dupin, Les tourbillons alternis, Paris, 1928;
L. Rosenhead and M. Schwabe, Proc. Roy. Soc., A129 (1930), pp. 115-135.
52
§7] DIMENSIONLESS PARAMETERS

lie inside those predicted by Kirchhoff.1' It also causes

the observed lift coefficient CL of a tilted plate to greatly
exceed that calculated on the basis of "wake theory." This
underestimate led to undue pessimism in the early days,
concerning the possibility of flight by heavier-than-air craft.
Since the magnitude of the wake underpressure appears to

is,
persist no matter how small the viscosity seems to me

it
that the flaw in theory not due so much to physical as to

is
mathematical oversimplification (lack of mathematical rigor
in treating the case of negligible viscosity). shall therefore

I
refer to the Paradox of Wake Underpressure.
similar discrepancy between theory and observation exists
A

for jets flowing out of an orifice into fluid of equal density.

a
So far as know, good agreement between classical free
I

streamline theory and experiment has been found only in the

case of liquid jets issuing into air, or some other low density
medium.17

Dimensionless Parameters p'/p and

7.

Q
A familiarity with the preceding facts leads one to focus
attention on two physical parameters which are completely
ignored in the classical theory of jets and wakes. Thus
neither even mentioned in the expositions of [9], [10], [12],
is

[34], or [34']; much of the recent progress which shall report

I

depends on recognizing their role explicitly.

The first of these parameters the dimensionless ratio
is

p'/p of the densities of the fluids on the two sides of the stream
line of discontinuity. It 1.0 in classical wake theory, and
is

about .0013 for water jet in air.

a

" O. Flachsbart, Zeits. ang. Math. Mech., 15 (1935), pp. 32-37; [6],
p. 554; Camichel, Comptes Rendus, 174 (1922), 666-669; L. Escande and M.
Teissie-Solier, ibid., 191 (1930), p. 519. For aeronautical implications,
see Lanchester, op. cit. supra, pp. 264-265; Kelvin, Collected Papers,
Vol. 205-210.
4,

17
For cases of such agreement, see [13], Vol. p. 304; R. von Mises.
1,

Zeits. Ver. Deutsche Ing., 61 (1917), 447-451, 469-473, and 493-497;

[16]. For jets flowing into media of the same density, see L. Marty,
Ann. Fac. Sci. Toulouse, %2 (1930), 41-146; E. Foerthmann, Ing. Archiv,
5

(1934), 42-54; [6], Section 255; D. Citrini, Energia Elett., 23 (1946),

133-144 and 302-315, and U (1947), 177-192; H. Rouse, Proc. [Link],
Civ. Eng., 74 (1948), 1571-1596.
53
 FREE BOUNDARY THEORY [CH. II

Thus the following plausible conjecture was made in 1931

by Betz and Petersohn [16]: if p'/p is small, then classical free
streamline theory can beapplied. (It is to be noted that the
smallness of the kinematic viscosity is not mentioned). Now
p'/p is small for jets of liquid issuing into a gas, and again in
the so-called cavity motion which occurs behind high-speed
missiles in water. In the latter case (cf. [19], [23], [35]) a
"cavity" filled with low-density air or water vapor (p'/p =
.00003) fills the region behind the missile. Hence it is plausible
that classical "wake theory" may apply to cavity motion —
it certainly does not apply to wakes!
We shall see later (Section 10) that even the conjecture of
Betz and Petersohn oversimplifies the facts. Nevertheless,
it represents an important step forward, and I shall speak
from now on of jets and cavities, instead of using the traditional
"jet and wake" terminology. Any "jets" to which I refer
should also be thought of as liquid jets moving through air or
some other low-density medium.
The other is the dimensionless ratio expressing the relative
wake underpressure: the so-called cavitation number

<12>
«-*£$*•
Here p denotes the "free stream pressure" of the undisturbed
flow; v is its relative velocity; p is the fluid density, and pc is
the "wake pressure" on the streamline of discontinuity. In
the classical theory, Q = 0 is assumed, corresponding to
infinitely long cavities.18
Now physical cavities usually have finite length and positive
cavitation number Q > 0; moreover the minimum pressure
usually occurs inside the cavity. It was observed by M.
Brillouin [22] that these conditions were incompatible with
irrotational flow in an ideal fluid, since they would lead to a
convex cavity. Such a cavity would have a stagnation point

18
As pointed out to me orally by G. Kreisel, an infinitely long cavity
in an infinite stream with Q j£ 0 is incompatible with regular flow at
infinity, by the Phragmen-Lindelof Principle. I note also that, by the
underwater explosion paradox of Chapter I, Section 17, a cavity cannot
form in plane flow.
54
§8] FLOWS WITH Q, r6 0

at the rear end of the free boundary, at minimum pressure p..

By Bernoulli's Theorem, t>2 = p. — p ^ 0 therefore, which
implies no motion at all; this is the Paradox of Brillouin.
One can escape from this paradox, and get at least a partial
mathematical approximation to physical reality, by supposing
that artificial alterations of the rear of a cavity, far from an
obstacle, do not greatly affect cavity flow about the obstacle.
I shall now discuss some such mathematical approximations.

8. Flows with Q^O

In 1921, Riabouchinsky constructed hydrodynamic
[32]
flows with free streamlines about two symmetrically placed

Dividing I III Streamline

Dividing
Streamline
»»-2
I
1 Wakec *—

Fig. 10. Riabouchinsky flow past plate.

plates (see Fig. 10), with Q > 0. A sketch of Riabouchinsky's

mathematical19 discussion follows.
The "hodograph" (i.e. diagram in the f-plane) of one
quarter of the flow is clearly a quarter-circle, and the W-domain
is a quadrant, by vertical and horizontal symmetry. Hence
the conformal map of the hodograph on the TT-domain
satisfies, much as in Section 2,

TF*_^2 + r2) +
£_cf' + 2\f* + i
as)
1 ;
co• + r2) +d
2
V + 2/if 2 + 1

By choice of units, we can reduce to the case c = 1, and f = 1

at W = oo, which implies p ** — 1. This gives

Physically, Riabouchinsky's original discussion referred to wakes,
19

where we should refer to cavities; see Proc. III

Int. Congr. Appl. Mech.
Stockholm (1930), Vol. 1, 137-158. The relevance to cavitation with
Q > 0 was noted by F. Weinig, Hydromechanische Probleme des Schiff-
santriebs, Hamburg, 1932, 294-300.
55
 FREE BOUNDARY THEORY [cH. II

(14) W = V'f4
+ 2Xf * +
(f2
- 1)
1

from which z = fdW/t can be obtained as an elliptic integral.

If v is the free streamline velocity, then f = v when W = 0,
whence f4 + 2Xf2 + 1 (f » -
f2)(f2 -
v-2), and -
X =

-H("2 + tf-2)• By Bernoulli's equation, Q = v2 — 1, whence

X = -H[(l + Q) + (1 + Q)-1]. With these formulas, it is
easy20 to obtain Cd as a function of Q.
Another important model for cavity flow with Q > 0 has
been constructed recently by Efros, Gilbarg and Rock.21

Free Streamline
Stagnation Point
*— Axis of Symmetry

Fig. 11. Cavity flow with reentrant jet.

This involves a reentrant jet Fig. 11), instead of a sym

(see
metric cavity. Such jets, somewhat analogous to those
discussed later (Sections 12-13), are now known to occur
experimentally when a symmetrical object is dropped into
water. It is therefore likely that this model closely resembles
actual cavity motion, though the two-sheeted Riemann
surface required for the z-plane cannot of course be realized
physically.
The Riabouchinsky and Gilbarg models agree with each
other in two respects. In both models, CD increases with Q
in the approximate ratio
=* + Q)CD(0) = + Q)
(15) CD(Q) (1 0.88(1
10
In 1946, T. E. Caywood extended Riabouchinsky's analysis to the
case of a wedge in a finite channel, but the numerical computations
involve untabulated definite integrals.
" D. Gilbarg and D. H. Rock, Naval Ordnance Laboratory Memo., 8718
(1945); D. A. Efros, Doklady Akad. Nauk. USSR, 51 (1946), 267-270
and 60 (1948), 29-31.
56
§8] FLOWS WITH Q, 7^ 0

for flat plates. In both models, the cavity width and length
increase without limit as Q J, 0. These . predictions also
agree generally with experiment (see Section 9), for Q small.
Of course, if Q > 1, there is usually no cavitation at all,
and a well-developed cavity usually only exists if Q < .5;
hence "cavity flows" with Q > .5 have no obvious physical
interpretation.
It is curious that the value Q = .5 should affect the drag
coefficient in much the same way as the Mach number M = 1
does. Other analogies between compressibility and cavita
tion (which is a form of compressibility) have been discussed
by Ackeret [15] and Riabouchinsky (Proc. Stockholm Congress
Applied Mechanics (1930), p. 156). One may note the further
coincidence that, in a perfect gas, M2 = v2p/yp = 2/yQ;
hence M = 1 does correspond roughly to Q = .5 — 1.0
(if 1 ^7^2). Finally, one may note Simmons' concept of
a "blocking constant," discussed in Section 11.
A simple hydraulic (and therefore non-rigorous!) derivation
of a formula closely related to (15), has been proposed by
Reichardt.22 Reichardt supposes that the energy DAs given
off in a distance As by a missile, is used up in expanding the
cavity radially to a volume A*As. (Here D is the drag, and
A* the maximum cavity cross-section area.) Since this
expansion does work against an adverse pressure difference
of p — pc (cf. (12)), we infer
D A*
ma\
(16) Ac = t
a

(P
- PcY r, or —r=-7l>
A Q
CD

where A is the missile cross-section.

It is encouraging that (16) agrees well with the numerical
results obtained with Riabouchinsky's model, if Q is small.
One can also construct flows with Q < 0 having finite
cavities, and otherwise defined mathematically over the entire
plane. Such flows avoid the Brillouin Paradox by sepa
rating from the obstacle behind the point of minimum pressure,
and by having cusps (cf. Fig. 12, innermost flow) at the rear
end.
"Reichardt, H., "The laws of cavitation bubbles," Reps, and Trans
lations No. 766, Ministry of Aircraft Production (Great Britain).
57
 FREE BOUNDARY THEORY [CH. II
Flows with cusped cavities can even be expressed analyti
cally in closed form, as shown recently by Lighthill and Allen.23
Southwell and Vaisey [33] have recently constructed an
interesting series of plane cavity flows past a circular cylinder
in a channel, with —0.4 g Q ^ 0.0. The cavity profiles are
sketched in Figure 12. The proper interpretation of these
flows is however ambiguous, for the following reasons.
First, the novel "cusped cavities" of finite length obtained
theoretically, do not at all resemble physical cavities (cf.
Section 9). Second, the flow obtained for Q = 0 is only one

Fig. 12. Cavity flows past cylinder in channel computed by relaxation

methods.

of a mathematically possible ene-parameter family, and is in

fact not the one usually preferred (see Section 5). Finally,
in the flows with Q > 0, the cavity width as determined in
[33] increases with Q, whereas all other evidence indicates
that, in an infinite stream with Q > 0, the cavity width
decreases as Q increases. Hence the contrary trend of South
well and Vaisey is presumably a wall effect.24
These remarks are perhaps timely in view of the current
interest in numerical methods. I think they illustrate the
need for accompanying numerical experimentation by a pro
found analytical knowledge, especially as regards general
existence, uniqueness, and convergence theorems.
23 M. J. Lighthill, "A note
on cusped cavities," Aer. Res. Council Reps,
and Mem. 2328 (1945), published 1949; D. N. de G. Allen, "The forma
tion of closed wakes in fluid motions," Quar. J. Mech. Appl. Math., $
(1949), 64-71.
" It may also be noted that the flows of Figure 12, which are intended
to be uniform to the right of a fixed line, could equally well be extended
by the reflection principle to give flows like that of Figure 10.
58
Plate I. Velocity field and cavity behind high-speed sphere in water.
§9] CAVITY FLOWS
In this connection, the following recent fragmentary result
of M. Schiffer,25 proved by variational methods, may be
mentioned. Let any body B of area /S and any constant
y > 0, be given. There exists at least one plane flow around
B with free streamlines, bounding a cavity of area y. Hence
there is indeterminacy as regards y; by the Brillouin Paradox,
presumably Q < 0. It is a plausible conjecture that, for a
class of bodies as general as accolades, there is just one every
where concave "cusped cavity," of area y(Q), for each Q < 0.
It would be interesting to determine y(Q) in specific cases,
and the limiting flow as Q f 0; this will presumably have an
infinitely long cavity of finite width.
Although it would seem analytically difficult to construct
simple flows with points of inflection on the cavity wall if
Q < 0, no such limitation is apparent if Q > 0. Therefore
it seems plausible that the criteria of Section 5 must be
invoked, if one wishes to avoid the embarrassment of a two-
parameter family of flows. It is an attractive conjecture
that, using these criteria, there is just one symmetric flow of
Riabouchinsky type past any "accolade," for sufficiently
small Q > 0.

9. Case p'/p Near Zero: Cavity Flows

I have already mentioned the plausible hypothesis that

classical "free streamline" theory is applicable when p'/p
is small. Mises26 has given elsewhere supporting data for
jets into air. I shall now review recent evidence for the
case of cavities behind missiles shot into water, and also for
cavities formed behind obstacles held in a water tunnel with
0.2 < Q < 1.0.
Plate I is a sample piece of evidence;27 it shows the cavity
formed by a sphere shot into water at about 150 ft/sec. Two

*• Oral communication to the author. Using reversibility, it would

seem that at least three topologically different types are possible, wake
behind, "wake" in front, and symmetrically disposed double wake.
» R. von Mises, Zeits. Ver. Deutsche Ing., 61 (1917), 447-451, 469-
473, and 493-497.
*7
The technique by which it was taken is described in [19], where
other similar photographs may be found.
59
 FREE BOUNDARY THEORY [CH. II

exposures, separated by about .005 sees., are made on a single

plate. The white dots speckled over the picture are small
bubbles, each bubble being photographed twice. It is not
hard to pair the bubbles, and the vector from the first to the
second bubble of each pair is roughly proportional to the
vector velocity of the water in the vicinity. In this way the
cavity profile and the velocity field of the flow can be clearly
visualized.
Now photographs such as these are not exactly comparable
with theory, for various reasons. Thus they represent a
decelerating body, and not stationary flow; the water surface
presents a second free boundary, which complicates the
mathematical description; the influence of air is not negligible
(cf. Section 10). Nevertheless, several pertinent general
correlations with theory can be made, based largely on first
hand observation.
First, except in the case of a water-tunnel with Q > 0.3,
real cavity outlines are relatively smooth28 and stationary for
ten or more body diameters back. This agrees with the
Kirchhoff model infinitely better than pictures of ordinary
wakes. (See the references of Section 5.)
Second, the cavity profile is nearly always convex, and
separation is in front of the cross-section of maximum diam
eter. This clearly supports the use of condition (iii) (hence
of conditions (i)-(iv)) of Section 5, in avoiding mathematical
indeterminacy.
The idea that the minimum pressure occurs in the cavity
(Section] 5, condition (ii)) is also supported by the observation
that vapor cavities always begin at solid surfaces in pictures of
cavity motion. Indeed, it follows immediately from Ber
noulli's equation (cf. Chapter I, (6), also [9], p. 19),

+ = G + const•'
(17)
^T l2u<+E
for non-stationary irrotational motion of an incompressible,
non-viscous fluid. In (17), V*G = 0, since G is the Newtonian
" Small "flutings" parallel to the flow may occur unless the solid is
xtremely smooth; see Plate I or [35], pp. 116, 129. In turbulent flow, a
ough cavity wall may occur at lower Q; see P. Eisenberg and H. Pond,
David Taylor Model Basin Report No. 668 (1948).
60
 §9] CAVITY FLOWS
gravitational potential; V\dU/dt) = d(y*U)/dt = 0, (cf.
Chapter I, (7)); Vl(wJ) = u&hu + 2Vw,Vu, ^ 0. Hence V*p
It however well-known

is,
^ 0, that superharmonic.

is

is
p
that superharmonic function must assume its minimum
a
value on the boundary. M
Fourth, would appear that arbitrarily long cavities (100
it
diameters at least), of arbitrarily great maximum cross-section,
occur behind missiles which are moving sufficiently fast.
This fact has considerable practical importance, in that
a

much of the wounding effect of high-speed bullets and bomb

fragments due to the fact that they create temporary holes
is

much bigger than themselves.30 For us, has the additional

it
significance of providing possible physical approximations to
the theoretical cavities of infinite length and asymptotically
infinite cross-section calculated by Kirchhoff .
Fifth, the rough experimental formula51

(18) CD = .55 + AQ = .55(1 + .73Q)

for the cavity CD of a cylinder broadside, agrees quite well

with theory. Thus Brodetsky had previously derived Co =
Cd(0) = .55 theoretically for the "preferred" flow (cf. Sec
tion at = and the formula for the other values of
0,
6)

Q
follow roughly formula (15), CD(Q) = + Q)CD(0).
(1

On the other hand, have seen nothing32 which at all

I

resembles the concave "cusped cavities" with < dis

0
Q

cussed above (Section would seem difficult to produce

6)
it

a
;

stable wake or cavity overpressure experimentally.

» See G. Bouligand, Jour,

de Math., (1927), p. 427.
6

10See [3], Section 74, and references given there; also E. N. Harvey,
A. H. Whiteley, H. Grundfest, and J. H. McMillen, TheMilitary Sur
geon, 98 (1946), 509-528.
Kempf and Foerster, Hydromechanische Probleme des Schiffenantriebs,
31

Hamburg, 1932; the experimental data are due to Martyrer. For

Brodetsky's result, see Proc. Roy. Soc., 102 (1923), 542-553; the sepa
ration zone appears to occur experimentally somewhat nearer the equator
than the 35° predicted by Brodetsky.
"The "deep seal" configuration of Fig. 13a entirely transient, and
is

immediately followed by cavity with reentrant down-jet (see

it
is

[24], and [19], Figs. 21, 25).

61
 FREE BOUNDARY THEORY [CH. II

10. Transient Scale Effects

The preceding results seem to confirm the plausible con

jecture already stated (Section 5, end) that classical free
streamline theory is applicable if p'/p is small, even though
it is not sufficient for v to be small. Unfortunately, recent
experimental research shows that this is true, if at all, only
in a very limited sense.

A B
Fig. 13. Deep seal versus surface seal.

If we wish to square this conjecture with the experimental

facts, we are forced to say that .0013 is not "small." Specifi
cally, there are two hydrodynamical phenomena which occur
with entry into water under atmospheric conditions, and not
if the air is evacuated. Hence no mathematical theory which
neglects p'/p ^ .0013 can correctly predict them.
The more important is the phenomenon of surface seal. If
a small sphere is dropped into still water at 10-20 ft/sec, the
cavity first closes as in Figure 13a in a "deep seal," whereas
at entry speeds of 40 ft/sec or more, the cavity first closes at
the surface, as sketched in Figure 13b. A photograph of the
velocity-field associated with surface seal is displayed as
62
§10] TRANSIENT SCALE EFFECTS

Plate [Link] existence of surface seal was discovered about

1900 by Worthington." In 1944, R. M. Davies [23], following
a suggestion of Sir Geoffrey Taylor, showed that surface seal
no longer occurs if the air pressure p is sufficiently reduced.
It was at first supposed by some that whether one got surface
seal or deep seal depended on the cavitation number Q =
p/3^pv2, which is proportional to p and intervenes directly in
other cavity phenomena. But in 1945, following my sug
gestion [18], Gilbarg and Anderson [24] used "heavy gases"
of the Freon series, so as to vary density and pressure inde
pendently. They thereby showed that, for a given sphere
diameter d and entry speed v (Froude number F = v2/gd),
the manner of seal was largely determined by p'/p. For
example, with 1-inch diameter spheres entering at 50-150
ft/sec (i.e. F = 103 — 104), surface seal occurs if p'/p > .001;
deep seal occurs p'/p < .0001.
This sudden change of regime near a very small value of
the parameter p'/p recalls the sudden changes in ordinary
wakes which occur near 1/R = v/vd = .02 (alternating
vortices) and 1/R = .000005, and in pipes when 1/R = .0005
(see Chapter I). Thus it gives us another "paradox of
approximation," and illustrates again the principle that the
nature of the solutions of partial differential equations may
change abruptly for extremely small values of the parameters.
It will be shown in Chapter III, Section 16, though by a
somewhat arbitrary physical analysis, that the change of
regime is actually associated closely with the dimensionless
parameter N = y/F p'/p, and occurs when N is about ^gn•
A second phenomenon is incompatible with the idea that
problems of water motion in air can be treated as free boundary
problems. Discovered about 1944 by various persons, it was
studied experimentally by Prof. L. B. Slichter. Prof. Slichter
showed that a smooth 2-inch diameter dural sphere, entering
the water at 50 ft/sec, and at an angle of 20° with the hori
zontal, may be refracted downwards 5° or more when it
enters. Though the exact facts are still not clear, Prof.
"See [35]; also W. Mallock, Proc. Roy. Soe., A9S (1918), 138-143,
and E. N. Harvey and J. H. MoMillen, Jour. Appl. Phys., 17 (1946),
541-555.
63
-
 FREE BOUNDARY THEORY [CH. II
Slichter made a careful (unfortunately unpublished) experi
mental analysis34 which indicates strongly that this downward
refraction is due to underpressure in the region where water
is separating from the sphere — and that in turn this under
pressure is caused by air viscosity, and its effect on the estab
lishment of the boundary layer. However, systematic
quantitative data are unfortunately still lacking.
Finally, various water tunnel studies36 show that regardless
of p'/p (which is only .00003 with water vapor cavities), the
flow varies greatly as the cavitation number Q varies from
zero to one. Hence, although the smallness of p'/p may be a
sufficient condition for the validity of classical jet theory,
it is only a useful necessary condition for the validity of cavity
theory. Much work remains to be done in this connection.

11. Wall Corrections; Stability of Pressure Coefficient

By clarifying the physical foundations of boundary

free
theory, one may obtain useful and unsuspected applications
to actual problems. I shall now (Sections 11-13) present
some examples of this.
I shall begin with the important question of wall effects.
Valcovici36 showed in 1913, using the physically inept language
of "wake theory," that whereas the theoretical cavity CD
("Cavity drag coefficient") of a flat plate in an infinite stream
was 0.88, as in Section 2, in a tunnel with fixed walls (as in
Fig. 14a), 100 plate diameters apart it was 1.15, or 30%
greater. The theoretical wall correction is reduced from 30 %
to less than 1% if the asymptotic "downstream" velocity
vc of the cavity wall is used instead of the usual "upstream"

velocity v, in computing CD = 2D/%pv2A. That

is,

the wall
correction can be made approximately simply by computing,
instead of CD,

34 Upward refraction occurs ordinarily; see K. Ramsauer, Uber den

Ricochetsschuss, Thesis, Kiel, 1903; also [3], p. 453.
" See [15]; Kempf and Foerster, op. cit. supra; Eisenberg and Pond,
loc. cit. supra; and unpublished data of Prof. R. T. Knapp.
1• V. Valcovici, Uber discontinuierliche Flussigkeitsbewegungen mit zwei
freien Strahlen, Thesis, Gottingen, 1913.
64
§11] WALL CORRECTIONS

(19) c*
^fl
= _A-
HpfM
=
(p. -5 pM'
,

by Bernoulli's Theorem. (Here p, is the stagnation pressure,

pc the cavity pressure, and A the model cross-section.)

Fio. 14. Flows past plate in closed channel, free jet, and bounded jet.

Valcovici also showed that the wall correction became

negligible if a jet with free boundaries (Figure 14b) was used
instead of a stream between fixed walls. Recently, Milton
Plesset and I [20] have noted that Valcovici's formulas could
be derived easily by the method of Section 4, since the hodo-
graph of half the flow is a quarter-circle; we also obtained
numerical results for bounded jets (Figure 14c). Similar
65
 FREE BOUNDARY THEORY [CH. II

results were obtained by N. Simmons for a class of flows with

Q > 0 including those treated by Riabouchinsky.
Now it is almost certain that the original physical interpre
tation of Valcovici's result as a wall correction for wake drag
is entirely erroneous. And it is only honest to say that my
engineering friends have been very skeptical of the interpre
tation as a wall correction for cavity drag. They point out
that the theory neglects the pressure gradient of the "undis
turbed stream." This is another wall effect caused by eddy
viscosity, which is well-known in hydraulics — if not in classical
hydrodynamics.
Nevertheless, I maintain that formula (19) is both mathe
matically and physically sound. In fact, I should like to
assert the validity of the following stronger principle of
stability of the pressure coefficient: for an obstacle of given
shape in a water tunnel (or jet!), the pressure profile function
- _ - P.)
MOM
U }
(P
(P.
- Vc)
Pc)
(p
YiPv*
is insensitive to the presence of walls and changes in Q.
Thus it was proved rigorously in that (19') is true for
[20]
wedges of any angle symmetrically placed in a water tunnel.
Again, formula (15) for the variation of CD with Q, which
was obtained by Gilbarg by study of the numerical evidence —
and which has received rough experimental confirmation
(see (18))
— is equivalent to (19')• To see this, it suffices to
note that by Bernoulli's theorem Q = xA,p(v\ — f*2)/3/6pi>*2,
where v* is the (up-stream) free stream velocity. Hence, by
simple algebra,

q = (fc/f*)2 - i, i + Q = »;/v*
and
2D =
(19") C> = ~"a =
(l + Q)C*.
pv*2A (vt)c*
Finally, I note that Ackeret [15] obtained consistent results
by assuming (19'), in extrapolating from his observations for
Q > 0.25, to the case Q = 0.
In fact, it seems to me that the "principle of stability of
the pressure coefficient" has all the earmarks of a successful
66
§12] LINED CAVITY CHARGES
hydraulic principle — including an inability to stand too close
scrutiny.
Even if it is inconvenient to measure p, — pc (which must
however be known if Q is to be measured), one can make an
estimate of the correction in (19). If T is the tunnel cross-
section, and A* the maximum cavity area, then the mean
velocity at that point is vT/(T — A*). Since the cavity
velocity vc is (the cavity being convex) at least the mean
velocity, we infer

(20) C* £ _
(1 A*/Ty•
Equality holds in Valcovici's case, and more generally if there
is a downstream asymptotic cavity velocity and maximum
cross-section.
Formula (20) reveals arearatios for space flows as the
logical analogs of diameter ratios for plane flows. This sug
gests that a 30% wall correction will apply as Q 0 j
for cavity
flows past a flat disc in a tunnel ten times as wide.
Comparison of (20) with (19") also reveals the analogy
between wall corrections and corrections for Q. Both (cf.
(16)) effectively limit the indefinite growth of cavities. In
fact (cf. Section 6, before footnote one), it is quite possible
to identify the two, setting Q = (vc•/v*)2 — 1. Similarly,
every water tunnel with fixed walls has a blocking constant;
one can prove for plane or axially symmetric flow that (cf . [20])

(20') Q>2^C„jiZ^ if CD* .0625.

Hence one cannot get very low Q in such tunnels: thus to

achieve Q = 0.05, we must have T/A ^ 400.

12. Lined Cavity Charges

So far I have mentioned almost no recent progress involving

jets. However, a very important application of the classical
theory of impinging jets was discovered in the late war, to lined
cavity charges, such as were used in the American "bazooka,"
the British P.I.A.T., and many other anti-tank and demolition
67

/"~
 FREE BOUNDARY THEORY [CH. II

weapons. I
shall briefly summarize this application of free
streamline theory; a complete account may be found in [21].
The construction and performance of such weapons may be
described in principle as follows. An explosive charge, sur
rounding a hollow "cavity" lined with metal, is detonated
from the rear. I shall consider only the cases of conical and
wedge-shaped liners; a sectional view of these is given in
Figure 15a. As the charge explodes, it drives the liner inward
and forward, and it is found that the liner so impelled has an
enormous penetrating power. How can we explain this
power?
The best known explanation proceeds from the following
plausible approximate assumptions. Assumption 1: After

A B
Fig. 15. Collapsing cone and impinging jets.

receiving an initial impulse from the explosive, the liner walls

move inward under their own momentum with constant
velocity, until they meet on the "axis" A A' of Figure 15a.
Assumption 2: Under the terrific stresses generated, the liner
behaves like a perfect fluid. Assumption 3 : Relative to axes
moving with the junction J
of opposite walls, the flow is
stationary. Assumption 4 : The surfaces of the liner walls are
free boundaries.
These assumptions reduce the problem to the classical case
of "impinging jets" (see Figure 15b). It is a matter of high-
school algebra and trigonometry, combined with the laws of
conservation of mass and momentum, to predict the velocities
J
of the jet A' and residual slug AJ
formed by the metal liner,
and to predict the ratio of their masses. Especially, the
extremely high predicted velocity of the jet is important, as
explaining the effectiveness of lined cavity charges as anti-
68
§13] PENETRATION BY LIQUID JETS
tank weapons. The mathematics of penetration of armor
by high-speed jets will be described in Section 13.
But first I
should like to call attention to the peculiar fact
that, though the advanced mathematical theory of impinging
plane jets may be found in various places,87 I know of no
previous application of this theory, or discussion of the axially
symmetric analog. Also, it seems significant to me that the
practical usefulness of the mathematical model does not
come from the relatively advanced use of complex variable
theory and conformal mapping (the "hodograph" is a circle,
and 'the method of Section 4 is applicable in principle).
Instead, it is humble mathematical methods, combined with
simple physical conservation principles, which are really
useful. This is especially true because, as in the determina
tion of the coefficient of contraction of the Borda tube (Sec
tion 3), it is only the simple laws of conservation which are
applicable to the axially symmetric case of greatest practical
importance.

18. Penetration by Liquid Jets

It has been known since Worthington's early photographs

[35], that cavity "deep seal" (Figure 13a) is followed by the
formation of upward jets. Gilbarg and Anderson [24] have
recently shown that down-jets also occur, and the cavity flows
with reentrant jets described in Section 6 may be regarded as
crude approximations to the flow induced by such down jets.38
These jets are not to be confused with those arising from lined
hollow charges (Section 12).
A mathematical model for stationary plane up-jets has
recently been suggested by T. E. Caywood and myself [19];
it is surprising that it should have been discussed so slightly
in the classical literature.39 The so-called "physical" z-plane
"See [10], p. 280; [25], pp. 11-12; W. Schach, Ing. Archiv, S (1934),
245-265. In the plane, half the flow corresponds to a jet impinging on a
plate, which is easily realized physically. One gets the impression from
the figure on p. 381 of the 4th edition of Gibson's Hydraulics, that even
the usual topological picture is not realized physically in the case of
impinging jets.
38 Photographs displaying the velocity-fields of up-jets and down-jets,
like those of Plates I—II, may be found in [19].
I
39 was only able to find, in [25], one passing reference.
69
 FREE BOUNDARY THEORY [CH. II

is shown in Figure 16, with the up-jet along the x-axis SS'.
The "hodograph" of half the flow is a semi-circle, while the
W-domain is a half -plane with semi-infinite slit SOS'. Hence
the method sketched in Section 4 is applicable; the details
follow. By the elementary theory of Schwarz-Christoffel
transformations, the TP-domain can be mapped onto the
upper half <-plane by a transformation

dW = ( 1
-
~
dt, or W = t - In t.
(l
(21) ~A
i)
The most general transformation of the hodograph is given
as before by

(21') t-A' + B_ P + 2\t +

\ where

«■ = s (f + r1).

But by Figure 16, f = -

when t = + <» at S' and T; f = 0
1

when t — 1 at 0; and f = 1 when I = 0 at S. Hence X = — 1,

ju = 1, and by suitable choice of units we can make

-
<21") <
fc$
The preceding model may easily be altered to describe
penetration of an incompressible fluid by a two-dimensional
jet of different density, as of tank armor by the jet from a
wedge-shaped liner. The "dividing streamline" TOT' of
Figure 16 may be taken as a streamline of discontinuity (not
a free streamline, however!). Above TOT' is the penetrating
jet of density p moving downwards with speed v, obtained
from the model of the preceding paragraph by multiplying all
vector velocities above TOT' by the scalar —v. Below TOT'
is the material being penetrated, of density p'; we assume
that it behaves like a liquid; and obtain its velocity field by
multiplying all velocities below TOT' in the preceding para
graph by the scalar — y/p/p' v. By Bernoulli's equation, the
pressure is continuous across the streamline of discontinuity.
70
§14] STATIONARY CAVITIES

Thus we have a stationary streamline of discontinuity not a

free streamline; this type of flow seems to be new.40 By super
posing a constant vertical velocity on the whole, we get a
model for the penetration of a stationary target by a liquid jet.
What about axially symmetric jets? Again, the important
physical quantity, the rate of penetration, can be obtained by
use of algebra and energy conservation (disguised as the

Fia. 16. Collapsing cavity — penetration by jet.

Bernoulli equation) alone, without the benefit of higher

mathematics! The details are mathematically so trivial that
I shall not bore you with them (they may be found in [21]).
14• Stationary Cavities in Accelerated Motion

A novel type of non-stationary plane motion with stationary

free boundaries has been found very recently by von Karman.41 .

Consider the Bernoulli equation (17) of Section 9, with

40
See [17], where flows of this type are discussed in detail. Con
tinuous analogs of such changes of density have been introduced by
M. Munk and R. Prim, Proc. Nat. Acad. Sci. U.S.A., 33 (1947), 137-141.
41
Th. von Karman, "Les sillages non stationnaires " ; Prof, von Karman
kindly lent me a copy of this unpublished note. Karman's use of simi
larity is analogous to that of Wagner, described in Chapter IV, Section 8.
71
 FREE BOUNDARY THEORY [CH. II

gravity neglected,

for non-viscous fluid. We look for flows involving accelerated

motion past a fixed obstacle, in which dependence on time
and space can be separated, so that

(23) U = f(t)<t>(xi,x2,x3), and p = g(f)f(xi,xt,x3).

If we substitute from (23) in (22), we get

(22') f'4> + i/*(V*V*) + & =

C(t).
£ p

<t>,
For this to be an identity, since v</>V</>, ^/p, are functions of
alone, clearly and must bear constant ratio to f2, and
/'
x

a
g

and constant ratio to V0V0.

4>

a
^

Setting/' = — kp, we get — kdt = df/P, or (by appropriate

choice of time-origin) =
1/kt. This of course the case of
is
/

constant acceleration coefficient, and not of constant acceleration.

Now a, free streamline one with = const.; hence we have
is

\p

as necessary and sufficient condition

a

(24) -k<t> + %(V<t>V4,) = C,

or

= + C*/t.
V

(25)

In the case =
of no acceleration, (24) reduces of course
0
k

to the classic condition (Section 1), that velocity constant

is

on with stationary motion. Differentiating

free streamline
a

(25) along the streamline, since = dU/ds, we get (indicating

v

differentiation with respect to by an accent), — kU'

+
s

U'U" = Except at stagnation points where U' = this

0,
0.

amounts to U" = or = V = ks C: velocity linear

is
+
k,

a
v

distance along the streamline. Choosing constants

of

function
of integration suitably, we easily reduce to

U' = ks, = Viks2 + C*/t.

U

Using (25), one can apply hodograph method, and von

a

Karman has by this means obtained explicitly an accelerated

72
§15] AXIALLY SYMMETRIC FLOWS
plane cavity flow past a flat plate held broadside. Curiously,
a solution is possible in the large only for one value of k\
The preceding refers to a fixed plate in a fluid accelerated at
infinity. We can transform to the case of an accelerated plate
in a fixed fluid, by superposing a "virtual" gravity field,
constant in space at any instant. Now if the cavity is
filled with fluid of density p, such a field — or any other con
servative field — has no effect on the flow (Chapter III, Section
12). Hence von Karman's model describes a wake behind an
accelerated or decelerated object, for the case p'/p = 1 and
constant acceleration or deceleration coefficient such as occurs
with constant CD.
It is perhaps of interest to obtain an analog of (25) for
streamlines of discontinuity not free streamlines. Along any
fixed streamline, let u = /(f)v(x), where f(t) = l/kt. The
component of the Equations of Motion parallel to the stream
line reduces to

du/dt + udu/ds + dp/pds = 0,

where u is the scalar speed of flow, and s measures distance
along the streamline. Substituting for u = v/kt, we get
—
du/dt = u/t, whence we have the generalized Bernoulli
equation

\ + =
j I ^
t;2

w* vds, or ±
+ = vds
(26)
k

-p
[

in the case of fluid of constant density p. Since con

is
a

tinuous across any streamline of discontinuity,

-
Pi[|
«*]

p[1^-] =
u\

(26')
f
f

Ul<fc],
\

where subscripts refer to opposite sides of the streamline.

15. Axially Symmetric Flows

Until the past few the only significant non-planar

years,
flows with free streamlines which had been computed, were
the (axially symmetric) jet through circular orifice in
a

flat plate, and the impact of circular jet on flat plate. The
a

73
 FREE BOUNDARY THEORY [CH. II

former was calculated by Trefftz,42 using successive approxi

mations based on integral equations; he computed the coeffi
cient of contraction to be 0.68. The latter was calculated by
Schach, using the same method.
Recently, Southwell and Vaisey [33'] have shown the power
of "relaxation methods," by calculating the jet through a
circular orifice, from a container of different dimensions, and
computing besides the axially symmetric jet through a Borda
tube in the shape of a circular cylinder. In both cases
the coefficient of contraction was assumed — as "generally
accepted" to be 0.68 in Trefftz' case, and as 0.50 in the Borda
case (we showed this in Section 3). In addition they com
puted a flow with a "cusped cavity" past a sphere.
These achievements are obviously most promising, and it
is to be hoped that "relaxation methods" may also obtain
flows past a sphere with Q 5; 0, as these seem physically more
interesting.43 Also, it is to be hoped that the ambiguities
specified above (Section 8) can be eliminated. Indeed, a
rigorous justification of either the Trefftz or Southwell- Vaisey
method would be most valuable, as would an accurate com
parison of their numerical results.
In the same connection, one should mention the very recent
use of the electrolytic tank analogy to obtain the axially
symmetric Riabouchinsky flow past a disc in a channel.44
When (15) was used to extrapolate the results to Q = 0, the
estimate CD(0) = 0.823 was obtained, which agrees very well
with experiment.
Another impressive recent achievement is the derivation
by Levinson [30] of analogs of formulas (11)— (11'), for the

" [Link], Uber die Kontraktion kreisformiger Flussigkeits strahlen,

Dissertation, Strassburg, 1914; also Zeits. Math. u. Phys., 64 (1916),
p. 34. Schach's results are in the Ingenieur Archiv, 6 (1935), 51-58.
Note also Weingartner, Gott. Nachr. (1890), and H. Reissner, Zeits. ang.
Math. Meeh., 12 (1935), 25-35.
43
It would seem difficult to estimate "wall effects" by relaxation
methods; so many mesh points would be needed to get any accuracy;
cf. Section 11.
44 " Determination deslignes de jet dans les mouvements plans et
de revolution," Problems D359A, Travaux du Laboratoire de Calcul
Experimental Analogique, by P. Marchet, under the direction of L.
Malavard.
74
§16] SUMMARY

axially symmetric case. Assuming only that for some s with

0 < s < 1, the cavity profile can be expressed by

(27) y = x-g(x), where lim^. ?£& = 0,

Levinson proved that the cavity profile must satisfy

(28) y = y/Cxg(x), where

(In x)-H- < g(x) < (In x)-*+'
for all e > 0 and sufficiently large x. He proved further that
if (18) is strengthened to xg'(x)/g(x) = 0(l/ln x), then

(i.e.,
~ VCx .

l\
^^|
(28') y (In *)-* =
Jlirn^
and the drag given by
is

(28") a-nr1
However, Levinson failed to prove the existence of solution.

a
In view of the complexity of (28'), and of the Paradoxes of
Overdeterminacy which are known (cf. Chapter I), one may
well wonder there exist axially symmetric cavity (alias
if

"wake") flows, in the classical mathematical sense.

similar but much easier result has also been recently
A

obtained for plane and axially symmetric cavity flows with

reentrant jets ([6], Figure 10). have shown [17] that the jet
I

cross-section A* and the cavitation number are related by

Q

- + + ®A*
Q

2<*
(29) CD
+/1 whence

= Q-0.
^

if
C-i

The asymptotic considerations needed to prove (20) are quite

like those used in deriving the d'Alembert paradox and Levi-
Civita's formula (11').

16 Summary Recent Tendencies

of
i

In
closing this chapter, should like to stress the trend of
I

recent progress in free boundary theory. think clear

is
it
I

75
 FREE BOUNDARY THEORY [CH. II

that there have been important developments dealing with

an increasing variety of mathematical problems, inspired by an
awareness of the great variety of real physical problems.
Thus the trend of recent progress has been in the same
general direction as modern theoretical physics. Twentieth-
century physics has also been characterized by a broadening
of foundations, both in quantum mechanics and relativity, and
by an acceptance of the philosophy that we are not yet able to
deduce all the laws of Nature from a single set of convenient
and harmonious mathematical hypotheses.
I hope that a realization of this community of spirit between
modern physics and fluid mechanics will lead to a greater
interest of physicists generally in the theoretical and experi
mental problems of fluid mechanics. It seems surprising
that this subject, to which so much was contributed by Stokes,
Helmholtz, Rayleigh, and other physicists of the nineteenth
century, should owe all its more recent progress to mathe
maticians and engineers. I feel confident that a renewed
interest by physicists in the accurate correlation of theory with
experiment, would help greatly in providing it with a more
solid scientific basis.

76
 CHAPTER III
Modeling and Dimensional
Analysis
1. Introduction

The use of models to study fluid mechanics has an appeal for

everyone endowed with natural curiosity. What active boy
has not played with ship and airplane models, or crude models
of dam and drainage systems? Even in the most advanced
technical engineering, such models play a fundamental and
indispensable role.
And yet in few departments of the physical sciences is there
a wider gap between theory and practice, between scientific
knowledge and the state of the art, than in the use of models
to study hydrodynamic phenomena. This is partly because
engineers with adequate laboratory facilities, who are most
familiar with the realities of the situation, are usually pre
occupied by special practical problems concerning one type
of model. Whereas the academic scientist, preoccupied by
the difficulties of logical exposition, too often ignores experi
mental reality.
It is my aim to narrow this gap; this can be done only by a
critical examination of the whole field from both sides.
Being a mathematician, I shall begin with the theory of
modeling. Here there are a number of simple, yet funda
mental, mathematical remarks which have hitherto escaped
general attention — no doubt because mathematicians have
not seriously studied the subject. I hope I can make my
remarks in sufficiently plain language, so that physicists and
engineers will see their relevance.

2. Dimensional Analysis

A prioritheoretical arguments for the use of models usually

follow one of two main lines: so-called dimensional analysis,
77
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill
and what may be called inspectional analysis.1 As the former
is simpler and more widely known, I shall begin (Sections 2-7)
with a brief historical resume of its origins.
Dimensional analysis arises from an attempt to extend to
physics the Greek concepts of geometrical similarity, ratio,
and proportion.2 It was first applied by Galileo to predict the
strength of beams of given material, as a function of their
linear dimensions. He assumed it as intuitively obvious that
for any given material, failure of a beam occurred when the
force per unit area exceeded a certain maximum. He con
cluded that the safe load per unit volume was inversely pro
portional to the length, and various other classical results.
Other applications were given by Mariotte3 and Newton,
but it was Fourier ([47), loc. cit.) who first stated that there are
certain "fundamental units," in terms of which every physical
quantity has certain "dimensions," to be written as exponents.
To be sure, this idea is almost obvious, if one thinks syste
matically about the "conversion factors" needed to convert
physical quantities from one system of units into another.
By simple changes of units, Fourier found it easy to treat
problems involving heat conduction in small spheres, and
problems involving the earth, by the same analytical formulas.
It is relevant, since we are interested in the relation of theory
to fact, that Fourier's deductions were sometimes badly off,
because they ignored the effect of radioactive materials in the
earth's core! Nevertheless, his procedure for eliminating
parameters by changes of units is now classic, and has been
used with notable not uniform!) success in many parts of
(if

physics. In the hands of Stokes, Savart, Froude, Rey

nolds, Vaschy, and many others, suggested laws of basic
it

importance.

For dimensional analysis, see Bridgman [38]. The useful phrase

1

"inspectional analysis" due to Ruark [52]; the idea much older.

is

is

Thus Fourier [47], p. 128, mentions the Greek awareness of the

8

dimensions of area and volume. Rayleigh always referred to "simili

tude"; cf. the current phrase "dynamically similar." For Galileo, see
his Two New Sciences (1638), Second Day.
Mariotte, Traiti de la percussion des corps, (1679) and Traiti du
8

mouvement des eaux (1686); Newton, Principia mathematica (1686),

Vol. Sect. See also Bertrand [36].
J.
2,

7.

78
§3] ASSUMPTIONS

Above all, Rayleigh focused attention on the fruitfulness

of the procedure as a general method. This led to an intensive
discussion by leading physicists during the decades 1900-1920
of the assumptions tacitly involved, and a realization of the
limitations of dimensional analysis.
Though still frequently ignored in the engineering and
popular literature, these assumptions and limitations are
clearly stated in the classic monograph of Bridgman [38].
However, I do not wish to presuppose any familiarity with
the subject, and shall introduce dimensional analysis with a
simple example.

3. Statement of Assumptions

The following illustration of the procedure of dimensional

analysis may be found in innumerable physics textbooks, but
it is usually considered too confusing to the student to label
the assumptions as such.
Example 1. Let us assume (or remember!) that the speed
v of a wave in deep water is determined by its length X and
the acceleration g of gravity, so that v = f(\,g). Let us also
assume that the relation is the same in all "fundamental
units" of length and time. Having made these assumptions,
we can argue mathematically as follows.
Let a wave of length X in a gravitational field of intensity g
move with speed v, in one set of units. Choose a new unit of
length equal to X old units, and a new unit of time equal to
t old units. An easy computation of "conversion factors"
shows that, in the new units, the same wave will have length
X' = 1, and speed v' = w/X, while the acceleration of gravity
will be g' = gr2/\. Hence, setting t = •\/\/g, we get

v/Vfr =V = f(\',g') = /(1,1) = C,

where C is the velocity of a wave one unit long in a unit gravi

tational field, in one (hence all) systems of units. Multiplying
through by V^X, we get v = C y/g\.
An analog of the preceding argument can be applied under
the following more general mathematical conditions.
79
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill
I. There are certain independent "fundamental units" g,•

such that for any positive real numbers an [i «■ 1, • • •

,n]
we can
" change units " according to the formula

(1) T(qi) = cnq{ [ai > 0].

= 3
(In mechanics, n and the g, are length, time, and mass.)
II. There are "derived quantities" Q, (such as density,
velocity, viscosity, etc.) which are homogeneous in the sense
that under (1) each Q, is multiplied by a "conversion factor"
given by

(2) T(Qj) = Q,«?" • • • a;*.

The exponents a,* are called the "dimensions" of Q,•; if they

are all zero, then Qj is called "dimensionless." Clearly, any
product of powers of homogeneous quantities is homogeneous.
III. The quantity Q0 is determined by Qi, . . . , Qn,
through a relation Qo = f(Qi, • • • ,Q»).
IV. The equation Q0 = f(Qi, ■ • • ,Qn) is unit-free, in the
sense of being preserved by any transformation (1).
V. The quantities Qo, . . . , Q» involve all n fundamental
units. (In Example 1, n = 2, as mass was not involved.)
Then we can argue just as in Example 1, to get

Qo =
■ ■ ■
(3) CQV Q*n;

where C = /(l, • • • ,1). Moreover the x< are by IV such

is,

as to make (3) dimensionally homogeneous, that of the

same dimensions on both sides in every fundamental unit.
Hence

a0i = xidu • •
for = ■ • ■
n.
+
+

xnani
1,

(4)
i

This gives an easy way to calculate the x,•.

The nature of the results which can be obtained in this way
illustrated by the following examples.
is

Example Assume that the fluid resistance D for a rigid

2.

object of given shape and aspect inertial, in the sense

is
a

that determined by the fluid density p, speed and

it
is

v,

object diameter Then (IV) equivalent to

d.

is

MLT-* = (ML-3)*(LT-l)"L*,
80
§3] ASSUMPTIONS

so that equations (4) reduce to

1 = x, l=-3x + y + z, -2 = -y,
whence x = 1, y = z = 2. Hence if the relation is unit-free,
then D = KDpv^d?, where KD is a constant. (Actually,
KD = vCd/S, which is called the ballistic drag coefficient,
varies slowly.)
Example 3. If D is determined by v, d, and the fluid
viscosity n, if inertia is negligible (Stokes' "creeping motions"),
and the determination is unit-free, then D = K*pvd for
another constant K*, by a similar dimensional calculus.
When the number m of independent fundamental units is
less than the number n of variables in the right-hand side of
a unit-free relation Qo = f(Qi, • ' • ,Qn), other well-known
and useful formulas can still be obtained.
Example 4. Assume that D = f(p,v,d,n) is a unit-free
function of p, v, d, and n, under all changes (1) of units of
length, time, and mass. The dimensionless quantities
KD = D/pv2d2 and R = pvd/n (Reynolds number) are invari
ant under such changes. By one such change, we can how
ever4 reduce p, v, d simultaneously to 1, this will carry m into
n/pvd = l/R. Therefore

(5) D = KD(R)pv2d', where KD(R) = / (l,l,4)

Example Assume that D is similarly determined by
5.
p, v, d, and the compressibility —d(l/p)/dp = dp/pHp of
the undisturbed fluid. These yield the dimensionless power
products D/pv2d* = KD and v2dp/dp (of dimensions (LT1-1)2
(il/L-'XMLr-2!/-2)-1 = 1). A more physical interpretation
of v2dp/dp is obtained if we recall that dp/dp = c2, where c is
the speed of sound in the fluid. Repeating the argument of
Example 4, we then get similarly

(6) KD = /(Af2), where M = - is the Mach number.*

This is essentially Vaschy's argument; see also D. Riabouchinsky,

4

L'Aerophile, Sept., 1911.

'Called the Sarrau number in France. The term "Mach number"
was proposed by J. Ackeret, Schweiz. BauzeUung, 94 (1929), p. 179.
81
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill
The physical applicability of formulas (5), (6) will be discussed
in Sections 13, 18.
Example Let there be a stationary distribution of turbu
6.
lent energy E per unit mass among different eddy sizes X, so
that dE = E'(\)d\. Assume that this distribution is deter
mined by an inertial mechanism of turbulent energy transfer
to smaller eddy sizes X. Then the rate of energy transfer,
per unit mass and time, to smaller eddies, clearly has dimen
sions V2/T = L2/T3; hence it is multiplied by a2/y3 under any
change of scale L—>aL, T— > yT. Moreover to maintain
E'(\) unchanged, this rate must be independent of X. Hence
the mean time T(\) required to break up eddies of size X into
smaller eddies, must be proportional to Xw: T has the dimen
sions Ltt under change of scale. Now consider the frequency
spectrum of energy : dE = F(k)dk, where k = 2ir/X is the wave
number. Since dE has dimensions V2 = L2/T2, while A; and
dk = 2irdX/X2 have dimensions 1/L, F (k) has dimensions
L3/T2, or Lw, or k-^. This gives Kolmogoroff's formula for
turbulent energy distribution: F(k) « k-H.
This formula involves the well-known paradox of infinite
turbulent total energy density per unit volume for large eddies;
I shall not discuss the rationalization of this paradox.

4• Pi Theorem

Instead of giving more examples of special cases,6 I shall

proceed directly to the general Pi Theorem of Vaschy and
Buckingham, which may be formulated as follows:
Theorem 1. Let positive variables Qi, . . . , Qr transform
by (2) under all changes \1) in fundamental units Ji, . . . , qn•
Let m ^ n be the "rank" of the matrix ||a,*|| defined by (2).
Then the assertion that

(7) f(Qi, ■ ■ ■
,Qr) = o

is a unit-free relation, is equivalent to a condition of the form

(8) *(nlf • • •
,nf_) = o,
•
Numerous other examples are given in Bridgman [38], Chs. I, VI;
in Porter in B. A. Robertson, Gen. Elec. Review, 33 (1930), 207.
[50]; and
See also Rayleigh, Phil. Mag., 34 (1892), 59, and 8 (1905), 66; also
Nature, 95 (1915), 66.
82
§4] PI THEOREM

for suitable dimensionless power-products Hi, ... , Hr-m of

the Qi.
Explanation. The first sentence restates Assumptions I,
II of Section 3.
Assumptions III
and IV are generalized in
(7). Nothing is required like Assumption V, which is the
special case m = r — 1.
There is a possible ambiguity in condition (7). It means
not merely that the equation /
= 0 is true, no matter what
fundamental units are employed; it means that the locus
defined by /
= 0 is invariant under all transformations (2).

To see the distinction, consider the case of a particle under

constant rectilinear acceleration a; denote distance by s, time
by t, and velocity by v. Then in any system of fundamental
units,

(9) \/(v - at)(a + v) + \/t(v - at) = 0,

and
v + v2 = 2as + at.
(10)

Equation (9) is unit-free in our sense, since it is equivalent

to v = at in the octant v > 0, a > 0, t > 0. Equation (10) is
not, though as it represents a linear combination of v = at
and v2 = 2as, it is true in any system of fundamental units.
Thus the meaning of (10), though not its truth, depends on
the choice of fundamental units; this equation will be dis
cussed again in Section 5.
Proof. By matrix theory, ||a,*|| has a non-singular m X m
minor.7 By permuting the Qi and qit we can suppose that
this minor involves Qi, • • •
, Qm and qlf • • ■ , qm. (Thus,
in the usual language of physics, the other fundamental
units are not involved independently.) Then any vector
a, = (a,,, • • •
,&,n) with > m j
is a linear combination
a, = cn&i +
• • •
+cymam of the vectors ai, • • •
, am.
Hence n, = QjQj*11 ■ • ■ Q^''"1 will be dimensionless, with
Qi = n,QQf» • • • QmQ%". We define g by

(ID g(Qi, ■ ■ ■
,Qm•,ni, ■ ■ ■
,nr_) =/(Q1, • • •
,Qr).
7 For the properties of matrices assumed here, see for instance [37]
Ch. X, esp. p. 291. A non-singular m X m minor is simply a square
submatrix whose determinant is not zero.
83
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill

In the "octant"8 Qi > 0, • ■ • , Qr > 0, the transformation

of independent variables is one-one in the large. Hence
/ = 0 is equivalent to (i.e. defines the same locus as) g = 0,
and g = 0 is consequently also unit-free. But since the minor
\\aij\\ with i, = I, • ■ ■j, m is non-singular, the simultaneous

linear equations an log on + • • • + aim log am = log Q,

can be solved for arbitrary positive Qi, • ■ ■ , Qm, by suitable
choice of positive cti, • • • , am. Since g = 0 is unit-free, the
locus is therefore the same for all Qi, ■ ■ ■ , Qm hence (7) is
equivalent for example to

(8) </>(n,,
• • •
.e^h.)
= 0(1, • •
,i;ni, • • •
,nr-m) = o.

This completes the proof.

If, for example, this procedure is applied to (9), we are led
to define II = at/v, and to get after substitution,

(8.1) \/(i - n)(ii2 + i) + -v^T^n = o.

However it should be stressed that this method assumes that

(9) is unit-free; the impression that (8.1) gives an a posteriori
demonstration that (9) is unit-free, is false; indeed, it would
be hard to prove that (9) was unit-free, by direct algebraic
substitutions such as those considered above.
Historical Note. There is some controversy over the
origins of the Pi Theorem. Vaschy9 stated the result in 1892,
but without formulating his assumptions. He indicated the
method used here; but his logic was so cryptic that nobody
reproduced his demonstration. Buckingham ([40], [41]) gave
in 1914 the first proof of the Pi Theorem, but for the special
case that /
could be expanded in a Maclaurin series; until
recently, this was the only accepted proof.10 Recently,
8
In case the c,™ are integers, this can be generalized to other octants.
The behavior on the hyperplanes Qi = 0 is more difficult, and has been
discussed by D. Riabouchinsky, Comptes rendus, 217 (1943), 220-223.
9
A. Vaschy, Annales UUgraphiques, 19 (1892), 25-28. The ideas of
Riabouchinsky were evolved in a series of papers (L'Aerophile, Sept.
1911; Comptes Rendus, 217 (1943), 205-208, and 225 (1947), 837-839).
My discussion was presented in [56], before I knew of this French work.
10
In fact, Bridgman [38], p. 16, raises the question of whether more
general functions can be considered. In this connection, see Section 18.
A good dimension' ess function for which expansion in Maclaurin series
is impossible, is the a(0,M,v) in the Taylor-Maccoll determination of
Chapter IV, Section 4.
84
§5] DIMENSIONAL HOMOGENIETY

Riabouchinsky and A. Martinot-Lagarde [49] have clarified

the ideas of Vaschy, so as to obtain a much more general
proof. The proof given here represents, I think, a further
clarification, as regards the restriction to positive a,• and Q,
and the theorems on matrices assumed.11

5. Dimensional Homogeneity

A relation f(Qi, ■ •
is called " dimensionally
■
,Qn) = 0
homogeneous" if and only if it satisfies Euler's relations

under transformations (2). It is easy to show that any

dimensionally homogeneous identity is unit-free. For by
(12), d(logf) = Xidai + • • •
+ Xndan. Integrating, we see
that /
= 0 is transformed into the equivalent equation
anx.y = o under (2).
. . .
aiXi
Conversely, most unit-free relations arising in practice are
dimensionally homogeneous. I shall now try to give this
well-known fact a mathematical motivation. Indeed, from
Theorem 1, we obtain immediately
Corollary 1. Any unit-free relation is equivalent to a
dimensionless (hence dimensionally homogeneous) relation.
More generally, consider relations between variables Qi
which, like (10), are not strictly unit-free, yet satisfy

(13) f(Qi, ■ ■ ■
,«r) = 0

for all fundamental units, for some function /. We shall call

such relations true in all units. Under any particular system
of fundamental units, if we define g by (11), / = 0 in all units
is equivalent to g = 0 in all units. As in the proof of the
Pi Theorem, by suitable choice of units, we can make Qi,
. . . , Qm assume arbitrary values. Hence the m-parameter
set of dimensionless equations in II i, , 11,—™ defined by ...
(14) 0 = f/oOIi, • • ■
,nr-m)
= g(Qi, • • ■
,Qm•, ni, • • •
,nP-„),
11
See also H. L. Langhaar, Jour. Franklin Institute. 2A.2 (19471, 459-
463.
85
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill

holds for all Q = (Qi, • ■ ■

,Qm). The converse is obvious,
proving
Corollary 2. Any relation which is true under all units
is equivalent to a set of dimensionless relations.
Though the theory of functional dependence is notoriously
difficult,11* it is plausible that at most n functionally independ
ent relations can exist, in general. Hence it would seem as if
any ordinary law, true in all units, could be reduced to a
finite set of dimensionless identities.
If /(Qij • • • ,Q») is a polynomial, or can be expanded in a
Maclaurin series,12 a neater treatment is possible. We can
show that / = 0 is true in all units, if and only if every dimen-
sionally homogeneous constituent of /
vanishes identically.
Indeed, consider the identity (14) as a function of the a,,
when the fundamental units are varied. Since expansion in
Maclaurin series is unique, it is equivalent to the vanishing
of every dimensionally homogeneous constituent of g = f.
For example, (10) is equivalent13 to the pair of identities
v = at and v2 = 2as.
It follows as an immediate corollary that every binomial
or trinomial relation, true in all units, is dimensionally
homogeneous.
The preceding results substantiate in part the claim that
only "dimensionally homogeneous" equations can be regarded
as having fundamental physical significance — to the extent
that all relations involving positive quantities, and which are
true in all units, can be reduced to dimensionally homogeneous
(hence to dimensionless) form. However (see Section 7) not
all physically significant equations are true in all units; hence
the claim has only a limited validity.
n* Assuming sufficient differentiability, many cases can be treated by
implicit function theorems (cf. G. A. Bliss, Fundamental Existence
Theorems, New York, 1913, Chapter I). If only continuity is assumed,
all the difficulties of topological dimension theory arise.
11
Since only the uniqueness of the expansion is really involved, it
would seem that expansibility in Laurent series was sufficient.
13 This
result, and the example of (10), are due to Bridgman [38],
p. 42; it led Bridgman to propose a "vector calculus" of identities.
From a mathematical point of view, in the case of polynomial relations,
this amounts to characterizing a locus by the ideal of polynomial equa
tions which it satisfies; thus we are led to classical ideas of algebraic
geometry. For the question in general, see Bridgman [38], pp. 41-46.
86
§6] VALIDITY OF ASSUMPTION II

6. Validity of Assumption II
In many instances (as when velocity is measured indirectly
by dividing the distance traversed by the time elapsed), the
validity of Assumption II
of Section 3 follows directly from the
physical definition of the Qj. However, it must be admitted
that not all experimental measurements are related to funda
mental units in such a simple way. What it means to say
that "they could all be so reduced," beyond an expression of
faith, it is difficult to say.
One can also characterize14 homogeneous quantities by
Bridgman's property of "absolute invariance of relative mag
nitude." A function of the variables qi, . . . , qn is said by
Bridgman to have this property when it satisfies the functional
equation
/(gl, • • •
,gn)
=
/(«lgl, • • •
,"ngn)
QgN
/(gl, • • •
,gn) /(«lgl' • • •
,«ngn)

for all positive on, ?, and = The result

■

■ ■
I,
q'i

»].
[i

follows.
Theorem Let = f(qi, ,qn) be a positive quantity,
2.

■
■

•
Q

whose numerical magnitude varies continuously with the funda

mental units q^ and which satisfies (15), so that ratios numerical
of

magniiudes are invariant under changes in the fundamental

units. Then must satisfy (2).
Q

Proof. Let = a, and all other variables

q\

1.

on =>am,
After transposition, (15) becomes
- /(o-«
■

• • ■ •
(15') X/(<rm,l, ,1) ,1),
•
1,

where = /(<r,l, • •
,1)//(1,1, •
,1). By induction on
•

•
•
X

m, we get for all positive and negative integers m, .

• • - X»/(l,l,
■

•
/(•• ,1) ,1).
•

(15")
•
1,

14 See Bridgman [38], p. 21. Equation (15) evidently states that

ratios of homogeneous quantities are invariant under changes of the
fundamental units. The assumption that the q, be positive, though not
stated, also required in Bridgman's proof, since he forms fdqi/qi. And
is

indeed, negative on usually have no physical meaning. The extension

from differentiable to continuous functions due to A. Martinot-
is

Lagarde, "Sur l'application des variables sans dimension aux ph6-

nomenes discontinus," Comptes Rendus, M3 (1946), 136-137.
87
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill

Hence, setting qi = 2n"n, a = \^2, and /(l, • • • ,1) = C,

we get /(ffi,l, •••,!)= CX« and /(2,1, ■ • • ,1) CX». -
Denning a = n log* [/(2,1, • • •
,1)//(1,1, ■ ■ •
,1)] n -
log2 X, so that Xm = 2mo/B, we get by substitution,

(16) /(2n/n,l, •••,!) = C2n"»/n =

C(2m/n)°.

But the powers of with rational exponents are an every

2
where dense set of positive real numbers. Hence, if is /
continuous — in fact, unless /is non-measurable and every
where discontinuous,15 f(qi,l, • • • ,1) ™ Cq\ holds for all
positive qi.
Repeating the argument for the other indices, we get the
conclusion of the theorem.

7. Validity of Assumptions I and IV

The validity of Assumptions I and IV has been hotly
debated. The self-evident principle that "all units of meas
urement16 are equally valid" is often loosely interpreted to
imply that all such units must give rise to the same universal
physical laws.
That this implication is not logically necessary is easily
seen if one recalls that, in a sense, all coordinate systems for
space-time are equally possible. But a geocentric system, such
as used in Ptolemaic astronomy, does not give rise to the
same physical laws as a heliocentric system.
Moreover the interpretation is not true, even for the units
of length, mass, and time in mechanics.17 Indeed, the main
15
For everywhere discontinuous counterexamples, see G. Hamel'
Math. Annalen, 60 (1905), 459-462. By a logarithmic transformation'
(15) reduces to a well-known additive functional equation. It is easy
to show that, by (16), if /is discontinuous at one point, it must be every
where discontinuous. For the nonmeasurability of /, see S. Banach,
Thiorie des opirations liniaires, Warsaw, 1933, p. 23.
14
For a discussion of the meaning of a scale of measurement, see
Norman Campbell, Measurement and Calculation, 1928. Strictly, one
should distinguish between enumerative scales, ordinal scales, scales in
which equality of differences can be observed, and true linear scales with
zero.
17 One
can, of course, rescue Assumption IV by the sophistry of
defining a "fundamental unit" in a theory to be one for which Assump
tion IV holds; this may even be useful.
88
§7] VALIDITY OF ASSUMPTIONS I AND IV

point of special relativity is that the laws of mechanics, which

include the basic force law

Fp _ d(m0u/\/l - u*/c2)
'
(17)
It
are notunit-free for length and time independently, but
involve the speed of light.
There are many universal dimensional constants other
than c. Thus such constants are involved in Newton's
universal law of gravitational attraction,18 F = ymm'/r2; in
Planck's quantum law E = hv, in the diameter of every
ultimate particle, etc. We are thus impelled irresistibly to
the conclusion that there are no known "fundamental units,"
with respect to which all known physical laws are unit-free.1*
Indeed, the decision to call certain units "fundamental"
(or primary) and others "derived" (or secondary) is one of
convention, and not of physical necessity. Thus it is some
times convenient to treat force as independent of mass, length,
and time.20
However, the preceding statements are subject to one
reservation. The universality of "universal constants" may
not be absolute. Thus before Newton's discovery of gravi
tation, g must have seemed a universal constant (this is still
inherent in the use of engineering units). In electromagnetic
theory (though not in relativistic mechanics), c is not usually
" In fact, according to Eddington (Land. Phys. Soc., 1918, p. 91),
Nature practically dictates choosing units of length, time and mass in a
unique fashion, so that y = c = h = 1. The limited laws of "gravi
tational similitude" have been discussed by Labocetta, Elettrotecnica,
1932, p. 1629.
" For further examples, see Bridgman [38], p. 103. In view of these
examples, Tolman's widely quoted statement that "the fundamental
entities of which the physical universe is constructed are such that from
them a miniature universe could be constructed exactly similar . . .
to the present universe," is clearly false. We would have to alter the
values of y, c, or h. Of course, the idea is very old; note Swift's Lilliput
and Brobdingnag!
10
See Bridgman [38], p. 65. In thermodynamics, we have the remark
able paradox of Riabouchinsky (ibid., p. 10); also the Stefan-Boltzmann
radiation law E/Area. = KT*. The artificiality of the concept of a
"fundamental unit" appears even more clearly with electromagnetic
units; see J. Jeans, Electricity and Magnetism, Cambridge Univ. Press,
1941, pp. 14-15; also Brylinski, Comptes Rendus, 215 (1942), 104.
89
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill

treated as a universal constant, but understood to depend on

the dielectric constant and magnetic permeability of the sub
stance under discussion. Hence it would seem premature to
dismiss as without foundation efforts such as those of Tolman
[53] and Eddington21 to deduce relations between universal
constants from epistemological principles.
What is more relevant to hydrodynamics, I shall now show
that in the Newtonian mechanics of fluids, length, time, and
mass can be taken as independent fundamental units. And it
seems almost certain that the laws of Newtonian mechanics
very nearly determine the behavior of fluids under ordinary
conditions.

8. Inspectional Analysis

In order to become convinced of the truth of the principle

italicized above — which always underlies the validity of
Assumption IV in fluid mechanics — one must resort to
"inspectional analysis." By this I mean, in general, testing
for invariance under the transformation (1), every funda
mental mathematical equation on which a given theory is
based. This test was in fact used by Fourier, Stokes, etc.;
it was clearly in the back of Rayleigh's mind when he referred
to "similitude"22; and it has been used very often since. It
occurs most commonly in connection with "dynamic simili
tude," as now defined for fluids.
Definition. Two fluid motions $ and <t>' are called
dynamically similar if they can be described by coordinate
systems23 which are related by transformations of space-
" Relativity Theory of Protons and Electrons, Cambridge Univ. Press,
1935. See also Bridgman [38], Chapter VIII; and E. T. Whittaker,
Space and Spirit, Edinburgh, 1946.
"See Rayleigh, Phil. Mag., 34 (1892), 52, and 8 (1905), 66; also
Nature, 95 (1915), 66. Similarly Bridgman ([38], p. 17) writes "The
advantage of (dimensional analysis) is that it is rapid, . . . but . . .
it does not give as complete information as might be obtained by . . .
a detailed analysis"; also "dimensional analysis physically is not so
instructive as the similarity condition." See also L. Schiffer, Zeits. ang.
Math. Mech., 24 (1944), 289-293.
"Actually, we mean "Newtonian" coordinate systems, in which
Newton's laws of motion are valid. Since the assumption that these
exist begs the question of the possibility of dynamic similitude, there
is some circularity in this definition.
90
§9] RELATION TO GROUPS

time-mass, of the form

= axi, = fit, m' = ym.

x't
(18)

t'
We can easily test the equations of hydrodynamical theory

a
for invariance under (18). For example, consider the Navier-
Stokes equations of motion for an incompressible fluid,
expressed in formula (3*) of p. These may be written in

7.
modified notation as

„„ ^.,^ ,,_i.
+
Here D/Dt = d/dt + "Lukd/dxk, so that Du/Dt ordinary

is
acceleration; the gravitational field; and = n/p the

is
is
g

v
"kinematic viscosity." Clearly (19) "dimensionally homo

is
geneous," of dimensions LT-2; hence invariant under

is
it
(18), provided the constants g, and p, and the variable p,
v,

are transformed appropriately.

Relation to Groups
9.

It would be a mistake to suppose that the applications of

" inspectional analysis" were limited to transformations of
the form (18). Actually, inspectional analysis is applicable
to any group By group of transforma
of

transformations.2*
a

tions, mean of course a set of transformations which con

I

tains the identity, the inverse of any one, and the product of
any two of its members.
This assertion based on the logical axiom, discussed in
is

Chapter that the hypotheses theory are invariant under

of
I,

if

then so are its conclusions. n

G,

a group Conversely, the set

all one-one transformations leaving any set equations
of

of

invariant forms a group.

After the group of transformations of the form (18), the
most important group for mechanics the ten-parameter
is

** This point was made by Mrs. Ehrenfest [46], p. 261 but as she gave
;

no applications, has been generally ignored. For elementary facts

it

about groups, see [37], Chap. VI; for orthogonal matrices, see Chap. VIII.
" This principle, though widely assumed by mathematicians and
physicists, seldom explicitly formulated. See however G. Bouligand,
is

Thiorie ginirale des groupes, Paris, 1935, p. 3.

91
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill

Galilei-Newton group. This group is generated by the three-

parameter subgroup S of space-translations

(20) x\ = xi + a [i = 1, 2, 3],
the one-parameter subgroup T of time-translations

(21) a = t + c,

the three-parameter subgroup R of rigid rotations

= a***,

x'i
(22)

^
where the most general 3X3 orthogonal matrix,
is

||a,•*||

and the three-parameter subgroup M of changes to moving

axes being translated with constant velocity
= xi -
4
(23) bit.

Now easy to verify that Newton's three laws of motion

it
is

areinvariant under (20)-(23), and that these transforma

tions leave unchanged the definitions of such material con
stants as density, viscosity, etc. (mass being assumed the
Hence, theoretically, Newtonian mechanics invariant

is
same).
under the Galilei-Newton group, as well as under the group of
transformations (18) of dynamic similitude. Experimentally,
this principle has been verified in many different ways, with
very great precision, except at speeds comparable with that
of light.

10. Theory Modeling,

of

shall now discuss critically justifications for model experi

I

ments by priori considerations. shall begin by recapitu

a

lating methods already in the literature.

The Pi Theorem mathematical theorem; hence say
is

(I
a

this for the benefit of physicists and engineers) has absolute

it

validity. This suggests that dimensional analysis the

is

proper method for justifying the use of models. The method,

though the one most commonly appealed to, however
is

inadequate for two reasons.26 In the first place, although

" The inadequacies of dimensional analysis have also been exposed
by Norman Campbell [43], A. E. Ruark [52], Mrs. Ehrenfest [45], [46],
and many others.
92
§10] THEORY OF MODELING

Assumptions I, II, and IV seem to be sufficiently valid in

fluid mechanics (see Section 9), dimensional analysis provides
no basis for deciding a priori when variables Qi, . . . , Qn do
determine Qo with sufficient accuracy. Thus Bridgman27
" cannot
says flatly that this crucial question (Assumption III)
be decided by the philosopher in his armchair," but only on
the basis of considerable physical experience. I shall give
many experimental instances of this difficulty (in Sections
12-18); others are given by the "paradoxes of approximation"
discussed in Chapters I and II. They are related to the
mathematical questions raised in Section 11.
The second limitation of dimensional analysis is that
"transformations of similitude" corresponding to changes of
units do not have enough mathematical generality to cover
useful methods of modeling. One wants (for example) more
general affine and conformal transformations as well.
Inspectional analysis would seem to resolve both of these
difficulties. One writes down the differential equations on
which theoretical fluid mechanics is based, and simply tests
them for invariance under all transformations in which one
is interested. I shall now give an example which illustrates
both inspectional analysis, and the care with which it must
be applied.
Consider the Navier-Stokes equations (19) for an incom
pressible non-viscous fluid. Following Ruark [52], they
can be reduced to dimensionless form as follows.
Let V, L, and P be respectively a characteristic velocity,
length, and pressure in the model, presumably controlled on
the boundary of the fluid flow. If we multiply (19) by L/F2
to make it dimensionless, and introduce the dimensionless
variables u\ = Ui/V, t' = Vt/L, x\ = xi/L, p' = p/P, and
the dimensionless constants R = VL/v, F = V2/LG, and
Q* = P/pV, we get

Q*dp'
(iy) Ut^F\Gj '
Dt' R dx't

where </,/(? are the direction cosines of gravity.

" [38], pp. 13-14; see also p. 50.

93
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill
There is a remarkable correlation between the dimensionless
differential equation (19') and engineering experience; we can
derive from it three of the most important rules of thumb
involved in model work.28 Thus it suggests that if gravity,
compressibility, and cavitation are unimportant, one should
use models at the same Reynolds number R. If compressi
bility, cavitation, and viscosity are unimportant, one should
use models at the same Froude number F. If compressi
bility and viscosity are unimportant, but gravity and cavita
tion effects are involved, one should preserve both F and the
"cavitation number" Q* (see Section 16).
It also gives the plausible suggestion that quantities are
unimportant precisely when corresponding coefficients in (19')
are small, and so appears to give a complete theoretical basis
for model work based on F, R, and Q*.

11. Theory of Modeling, II

Although the argument just given is suggestive and worth
remembering, it suffers from the defect of considering only
one of the three fundamental equations of fluid mechanics,
the equation of motion. Thus it ignores the "equation of
continuity"

™ £+2^-°. - ^"+^ =
0.

and the "equation of state," which can be written

(25) P = /(p).
For this reason, Ishall call it "partial inspectional analysis,"
and shall designate the corresponding procedure, when, a set
of conditions completely determining a flow has been con
sidered, "complete inspectional analysis."
One might suppose that the partial dimensional analysis of
Section 10 was legitimate in incompressible fluid dynamics,
for the following reason.

18 Vol. 2, Chap. I, for a very similar derivation, but without

See [12],
use of differential equations so explicitly. The prescription is that
'ratios of forces" must be preserved; Q* is not discussed.
94
§11] THEORY OF MODELING

Theorem 3. The equation of continuity is invariant under

all transformations of similitude; so is the equation of state
Dp/Dt = 0 of an incompressible, inexpansible fluid.
Proof. One can check the invariance of (24) and Dp/Dt —
0 under (18) and (20)-(23), by straightforward inspectional
analysis. Or one can observe the invariance under these
transformations of the laws of mass and volume conservation
which (24) and Dp/Dt = 0 express.
However, Theorem 3 no longer applies if expansion occurs
through cavitation. In this case, we have usually
(25.1) p = po if V > Pv, and
p = pv, p = pv elsewhere.

Since pv > 0, clearly (25.1) is not invariant under29 (18).

Hence the inferences of Section 10 regarding Q* are theo
retically unsound; I shall give a correct discussion later
(Section 16).
The suggestion that 1/F and 1/R can be neglected in (19')
precisely when they are small, will be discussed in Section 12.
But first I should like to discuss generally the completeness of
an inspectional analysis based on (19') and (24)-(25) alone:
i.e. the question of whether these equations (with boundary
conditions) completely determine flows.
It is natural to suppose •that they do, because they express
the time-derivatives of the variables p, p, «i, uz, M3 in terms of
their space-derivatives. (Determinism Principle.)
This is most evident in the case of compressible non-viscous
flow. One writes the equation of state (25) in the usual form
p = F(p), and thus eliminates p in the equations of motion30
(19). Then (19) and (24) evidently express the time-deriva
tives of p, Mi, U2, Ma in terms of their space-derivatives.
The case of incompressible flow is more complicated — really
by the theoretical possibility of "action at a distance." The
*9 From now on I shall not mention (20)-(23), as invariance under
them is true of Newtonian mechanics generally. But it should be
noted that the use of fixed models in a moving stream to simulate moving
models in a stationary fluid is based directly on (23), which is not of the
form discussed in ordinary dimensional analysis.
10 With non-viscous
flow, r «• 0. In the case of compressible viscous
flow, the validity of (19) is questionable.
95
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill

time-derivatives at a point are not determined by local space-

derivatives, but are influenced by the movements of rigid
bodies everywhere in space.
One rescues an appearance of determinism as follows. One
can ignore the variation of p, and (24) reduces to div u = 0,
i.e. to the condition that u is solenoidal. One introduces the
kinematic vorticity vector f = curl u. After some calculation,
one gets from (19) and (24) the vector differential equation

= *VVt +1 « if g = vG.
^
•
(26) V)u

This expresses the time-derivatives of vorticity in terms of

local space-derivatives, and shows that vorticity (unlike
velocity) is diffused, and not instantaneously communicated
at a distance. Finally, one remarks that the vorticity deter
mines the velocity up to an irrotational solenoidal velocity-
field (Dirichlet flow),31 and this is determined from the
normal derivatives at the boundaries, by potential theory.
This completes the argument.
To a physicist, the preceding argument may seem to be
mathematical perfectionism, or at least so abstruse that he
has no interest in pursuing the matter further. But this is
not so; it should be obvious to any twentieth century mathe
matician that the discussion of boundary conditions is vague.
Thus the "Determinism Principle" is really a very weak
uniqueness theorem, projecting conditions at t = 0 infini-
tesimally into the future. If all functions are assumed to be
analytic, a local projection into the near future is possible by
the Cauchy-Kowalewski Theorem; but no mathematical
theorem is known which justifies projection into the indefinite
future (see Section 18 for physical evidence corroborating
this mathematical argument).
Moreover, in most modeling problems, it is boundary con
ditions in space that are modeled, not at t = 0; for these
boundary conditions, and solutions in the large, the Cauchy-

81 In plane flows, V) =0; moreover we can obtain a solenoidal

(5
•

field with suitably bounded scalar vorticity £{x,y) explicitly, by the

stream-function V = //{ log rdxdy.
96
 §12] INERTIAL MODELING
Kowalewski Theorem is useless. No general uniqueness
theorem appropriate to the situation is known, and it looks
as if none is true. There are only special theorems, which
depend on whether the differential equations are of elliptic,
hyperbolic, or parabolic type.32
Thus for a complete inspectional analysis of hydrodynamic
modeling, fundamentally new theorems about partial differ
ential equations are needed. I have already expressed my
opinion that these theorems will be inspired by a familiarity
with the experimental facts. In this connection, note that I
the fundamental classification of partial differential equations
into elliptic, hyperbolic, and parabolic types, already men
tioned, was suggested by applications.
The rest of this chapter will therefore be devoted to summa
rizing the experimental facts concerning various types of
models, in their relation to mathematical theory.

12. Inertia! Modeling

In practice, considerations of experimental convenience and

economy often dictate not only the particular fluid (such as
air or water) used, but also the size of the models and the
speed of flow. In such cases, one has .to assume invariance
of the behavior under all transformations (18), or else to
work out a set of approximate corrections for "scale effects."
I shall call such use of (18) for all a, 0, 7, "inertial scaling."
In the early days of the airplane, all airfoil tests were made
by inertial scaling;33 moreover the quasi-empirical results so
obtained were far more reliable than the most convincing
theoretical formulas (of Rayleigh, such a formula as (6') of
Chapter II) then known. Today we know the pitfalls of
inertial scaling, and I shall now try to give an objective and

3' For such theorems, see R. Courant and D. Hilbert, Methoden der
Mathematischen Physik, Vol. 2. For results relating specifically to
hydrodynamics, see L. Lichtenstein, Math. Zeits., 26 (1927), 196-323;
K. Maruhn, ibid., 45 (1939), 155-184; J. Leray, Acta Math., 63 (1935),
193-248.
13
See G. Eiffel, La risistance de I'air et Variation, translated into English
by J. Hunsaker, Boston, 1914. Tests based on inertial scaling are still
widely used in airplane design.
97
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill

up-to-date account of some theoretical and experimental facts

about it.
Many hydrodynamical theories, contradictory in other
respects, agree in being invariant under inertial scaling. Thus
this is true of Newton's primitive kinetic theory of air resist
ance, and of Euler's quasi-empirical formula expressing drag
and lift as definite integrals.34 It is also true of the theories
of Dirichlet and Joukowsky flows;38 of two-dimensional airfoil
theory; of the Helmholtz-Kirchhoff theory of wakes and jets
(Chapter II); and of Karman's theory of vortex streets
(alternating vortices). I shall omit the " inspectional analy
sis" required to prove these statements.
It is also true of Oseen's theory,3' which purports to take
account of the asymptotic effect of viscosity, as R — > « .
Further, it is true of Stokes' theory of "creeping motion,"
which is a limiting case as R — > 0; since the terms in Dui/Dt
and g, are neglected here, the argument of Example 3 above
applies to this case.
I have already discussed (Chapter I, Sections 14 and 18)
the inadequacies of these asymptotic theories to explain
observed phenomena at l/R = .02, .001, and .00005 such as
alternating vortices, turbulence, and "transition" of the
boundary layer; I have also pointed out how clearly they
show the need for fundamental new developments in the
theory of partial differential equations. By the same token,
they show the insufficiency of inspectional analysis to justify
modeling, until such developments have been made.
Nevertheless, the plausible idea that inertial scaling may
be used with large R has some validity. There are extensive
"regimes," corresponding to large ranges of R, for which
variation of the Reynolds number gives experimental results
which essentially follow inertial scaling. Thus this is true
34
These theories are discussed at great length in P. Painlev6, Learns
sur la risistance des fluides, Paris, 1930; and in C. Cranz [3], Chapter II.
Until very recent times (1935), no better means for calculating bullet
resistance were known.
36 Dirichlet flows and two-dimensional airfoil theory have been dis
cussed in Chapter I, Sections 3, 5. As these two theories are reversible,
v need not even be positive in (18).
"For Oseen's theory, see [11]. For Stokes' theory, see [11], or [12],
Vol. 1, pp. 261-264.
98
§13] REYNOLDS MODELING

of alternating vortices,37 of jets and cavities subject to the

limitations described in Chapter II, of friction in pipes of
given roughness, of the resistance of bullets and bombs at
speeds up to M = .4, and so on. It is also valid, if 1/R is small,
in measurements of virtual mass for acceleration from rest
(see Chapter V) ; in this single case, the boundary conditions
for the Cauchy-Kowalewski Theorem apply.
It does seem possible to neglect gravity if F is very large.
One can also often neglect gravity for arbitrary F, because of
invariance of the basic equations under transformations quite
other38 than (18).
Theorem 4. In a fluid of constant density p0 (hence without
free surface), gravitational forces can be neglected, provided the
hydrostatic buoyancy of submerged models is allowed for.
Proof. Let G be the gravitational potential; then one can
reduce equations (19) and (24)-(25) to the case where G = 0
by the transformation

(27) p' = p - PoG.

The hypothesis p = p0 is of course not fulfilled in a compres

sible fluid; and (27) does not carry "free surfaces" (with
p = p0) into free surfaces, unless they happen to be horizontal
(i.e. gravitational equipotentials).
Theorem 4 shows again the inadequacy of the partial
inspectional analysis of Section 10.

13. Reynolds Modeling

Inspectional analysis based on Theorems 3 and 4 suggests

that the dynamics of an incompressible viscous fluid bounded
by fixed walls should be invariant under all changes of scale
37
See P. Dupin and M. Teissi6-Solier, Les tourbillons aUernis, Paris,
1928. For flow in pipes, see the curves of Nikuradse (V.D.I. Forschungs-
hefl, 361, 35b) reproduced in Fig. 117 of H. Rouse, [14]. For the drag of a
plate, see [14], p. 227. The case of a sphere is more doubtful. Much
experimental evidence for the principle that inertial scaling is possible at
sufficiently high R is given in C. Camichel and L. Escande, [42],
38 See Th. von Karman, Jour. Aer. Sci., 8
(1941), 337; also Avanzini,
1st. Naz. Ital., tomo 1, part 1 (1807). To be invariant under translations
—
9i > ?< + Pi means that there is no zero-point in the scale of measurement.
This is true of temperature in the theory of heat conduction.
99
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill

which preserve the Reynolds number R = VL/v, in the nota

tion of Section 10. In such "Reynolds models," pressure is
transformed by
+
(27') V + Of =
P0F/2(;o
PoG)
const.

This principle has received impressive experimental confirma

tion, with the most diverse liquids and gases.39 Thus (see
Figure 17) the breakdown of Poiseuille flow in pipes comes at
the same Reynolds number in air, water, and many other
fluids. At Mach numbers below M = 0.3, drag coefficients
of spheres and cylinders satisfy (5) with the same KD(R) in all
fluids, and at all sizes and speeds. A corresponding result
holds for the skin-friction of plates parallel to the wind.
In verifying these facts, two precautions must be taken;
otherwise Reynolds modeling will not work. First, one must
use models with similar surface roughness. The onset of
turbulent flow, and transition in the boundary layer from
laminar to turbulent flow, are greatly influenced by this
factor. Thus near RcAt, the drag of a sphere can be greatly
reduced by roughening it suitably.
Second, especially in wind-tunnels with closed circuits, the
turbulence of the free stream must be the same.40 It is found
that the R^t of spheres in wind-tunnels may vary by a factor
of two, depending on the turbulence of the wind-tunnel. To
get consistent results, it is usual to introduce a factor X for
each wind-tunnel, and to define an effective Reynolds number
Rett = WD/v, where X is chosen to make (R,s)cm assume a
certain standard value.
It is not easy to model large Reynolds numbers on a small
scale. The trouble is that, if the fluid is given, reducing
39 See T. E. Stanton and J. R. Pannell, Phil. Trans. AM4 (1914),
119-224, from which Figure 17 is adapted; also [50], pp. 16-17. The
theory is due to Stokes, Collected Papers, Vol. 3, p. 17. Since the turbu
lent behavior of liquids is dynamically similar to that of gases, it seems
hardly possible that turbulence should be related to kinetic theory
except indirectly, through viscosity.
40
Goldstein [6], p. 431. This explains the Paradox of DuBuat (Prin-
cipes d'hydraulique, 1816 ed., Vol. 2, p. 194) whereby the force on a stick
pulled with speed o through a still lake may exceed the force exerted on
the stick when held still in a stream flowing past with the same speed.
The moving stream is usually turbulent!
100
 06»
I—*
-
—
.007
Surface Friction with Air and Water
Experiments

♦
.006
♦ • Water in Pipe

♦
* ♦ Air in Pipe

•
♦ ♦
_
.005 • • m
+
<

♦ ♦•

.
♦

R
+ ■z
* ♦

4
* •♦
pv2

•
.004
orD

* ♦
♦ w
*

•
♦
♦ • ♦
2OO

.003 \- • ♦
• w
•

r
z0

—
.002

VD

1
l
I

111
1 taooo 23,100 1111
1

1
i

1
1

.001

3.0 3.1 3.4. 3.6 3.Q 4.0 4.2 4.4 4.6 4.8 5.0
V
D

i
- te
Log.
of
of

Fia. 17. Dimensionless plot of pipe friction, after Stanton and Pannell (courtesy Royal Society London.)
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill

diameters requires increasing velocities in the same ratio.

With air,41 such modeling is facilitated by "variable density
wind-tunnels," using compressed air. Unfortunately, there
seems to be no liquid whose v is much less than that of water,
though many have much greater v. Hence wind-tunnels
provide the only known economical Reynolds models of
phenomena of water flow.

14. Froude Models

A "Froude model" is defined as one at the same Froude

number F = V2/GL. Moreover on the earth, G is sensibly
constant, so that Froude modeling amounts to making
V « y/L: the velocity should be proportional to the square root
of the linear dimension. Generally speaking (cf. [12], Vol. 2,
p. 10), Froude models may be used when gravity forces and
inertial forces are predominant. In practice this occurs most
commonly in flows at fairly high speeds, which possess "free
surfaces" exposed to the air.
One can justify Froude models by inspectional analysis as
follows.
Theorem 5. Let viscosity be neglected, in a fluid of constant
density po, and let the gravity field be constant. Then equations
(19) and (24)-(25) and the property of being a free surface are
preserved under all transformations (18) which preserve F.
Proof. By Theorem only consider the trans
3, we need
formation of (19) under (18), ignoring the term in v. (The
neglect of v is valid under just the same circumstances as in
has little effect).
is,

Section 12; that where the variation of

R

We multiply (19) as in deriving (19') by L/V2, where

V
is
a

typical velocity, and get the dimensionless equation

(iy - •

F\G)
>

Dt' dx'i

41 In almost independent of the density, at fixed tempera

gases,
is
11

tures (Maxwell's Law). Hence 1/v = p/fi (and so R) proportional

is

to p. The of water increases substantially with temperature. For the

v

use of wind-tunnels for water studies, see C. Keller, Escher Wyss News
(1940).
102
§14] FROUDE MODELS

Hence, assuming that there is only one pressure distribution

compatible with the boundary conditions, we have dynamic
similarity, with equal differences in the pressure coefficient
2p/poV2 (not equal p/P) at corresponding points. Hence
surfaces at constant pressure (coefficient) are also preserved
by (18), completing the proof.
Froude modeling has proved important in turbine research,42
but is most generally known as a means of estimating the
wave resistance of ships. It is also used in harbor models, for
studies of the breaking of waves on beaches, and in studies
of water entry (see Chapter II).
There is a long-standing controversy as to whether Reech
or Froude should get the credit for the idea of "Froude
modeling" in ship resistance model tests. As the facts are
rather curious, I shall state them
Reech43 certainly proposed, in 1831, exactly what is com
monly referred to as "Froude's law," namely, testing ship
models at equal Froude number, and estimating full-scale
resistance by the similarity transformation (18). Froude's
great merit consists in having gone beyond this simple law.
In most merchant ships, ninety per cent of the resistance is
frictional, and hence susceptible to Reynolds and not to
Froude modeling. In order to estimate hull resistance from
model tests, one must separate ship resistance into its fric
tional and wave components. This was first proposed in 1874
by Froude [48]. The exact model laws are still far from clear,
and it is noteworthy that different rules for interpreting model
results are used in the United States, England, France, and
probably other countries as well.
Experience with Froude models quickly leads one to recog
nize other limitations. In harbor models, the interpretation
of model. results is complicated by capillary waves44 (as

42Its relevance was first mentioned in 1848 by Bertrand [36]; for a

modern account, see G. F. Wislicenus, Fluid Mechanics of Turbomachin-
ery, New York, 1946.
*' F. Reech, Cours de I'Ecole d' Application du Ginie Maritime, Lorient,
1831.
44
have not discussed the modeling of capillary effects. To model
1
these,one should preserve the "Weber number" W = KT'/pL3 =
K/pV*L, where K is the surface tension (Bashforth and Adams, Capillary
103
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill

regards erosion by breakers), and by the fact that the attenu

ation of waves by viscous effects does not correspond under
Froude modeling. Again, in Froude models of dams, water
falls, etc., the entrainment of air is also much less than in the
full-scale.

15. River and Harbor Models; Nearly Horizontal Flows

Although inertia and gravity effects can be scaled by using

Froude models of rivers, harbors, ships, etc., viscous effects
cannot be scaled, for the reasons described at the end of
Section 13. In models, viscous forces are relatively greater,
and hence turbulence ("eddy viscosity") less. The usual
cure is to induce turbulence in models, by artificially roughen
ing surfaces or even by obstructing fluid motion with coarse
wire meshes. This increases the eddy viscosity, so that the
viscous forces in the model are even greater, relatively, than
they would be otherwise.
However, at sufficiently high Reynolds numbers, this some
what paradoxical procedure may be partially justified by
inspectional analysis, as was pointed out to the author by
Dr. S. K. Roy. The viscous forces are then much less than
the "Reynolds stresses" puty, where u' denotes the turbulent
vector velocity, and a bar denotes averaging (see [6], p. 192).
Hence, if the relative turbulence is the same at all points, the
mean velocity distribution may be expected to be similar.
What is perhaps more important, separation of flow tends
to occur at the corresponding points, and suspensions are
prevented from settling by the same mechanism, if turbulence
is modeled. However, it should be emphasized that in prac
tice an attempt is usually made to model only the intensity of
turbulence, and certainly only the intensity and mean eddy
size.
The transport of solid
particles (dirt, sand, gravel) by
moving water is frequently studied in models, because of its
great practical importance for rivers, harbors, and estuaries.

Action, 1883). The poor modeling of air entrainment is discussed by

L. Escande, Genie Civile, 16 Dec. 1939; and by C. Camichel and L.
Escande [42].
104
§15] RIVER AND HARBOR MODELS

The use of such models is a highly individualized art, which

involves very complicated considerations.45 Considerations
of size make Reynolds modeling out of the question. Froude
modeling roughly preserves the velocities due to gravity head,
and (in habor models) wave motions as well, but only in the
turbulent regime or if viscosity is negligible.4'
The relative particle size is frequently exaggerated in the
model, partly to avoid cohesive forces, partly to preserve R,
and partly for ease of fabrication. This increased size tends
to keep the model particles from being stirred up by the water,
a tendency which is compensated47 for by reducing their
negative buoyancy pi — p.
The common use of different horizontal and vertical scales
in such models, also deserves comment. In England it is
usual to exaggerate the vertical scale (following Reynolds
and Gibson) to avoid the tendency to excessive shallowness.
The validity of this exaggeration is often disputed48 in France,
where there is tendency to use larger models instead. M.
Craya [44] has justified exaggeration of the vertical scale, for
the case of rivers with small curvature without waves; I shall
present a justification which allows waves to exist, in the case
of nearly horizontal flow. This case corresponds to a very
common hydraulic approximation.
Thus consider the theory of shallow water waves. If the
surface height is h(x,y,t) above a horizontal bottom, and
v(x,y,t) is the mean horizontal velocity, then one has the approx
imate equations

46
A lengthy discussion of some complications may be found in Proc.
Am. Soc. Civ. Eng., 71 (1944), Trans. No. 3, Part 2. In certain tidal
models, the Coriolis force must be modeled by changing river curvatures ;
see [39], p. 78.
" One can exaggerate the tilt of the bottom of the river or estuary to
obtain appropriate mean velocities of flow when R is small. Roughness
is an important factor.
47
See C. Camichel and L. Escande, Comptes Rendtts, 199 (1934), 992.
An excellent review of scaling laws for silt transport is given by H. A.
Einstein and R. Muller, "Uber die Ahnlichkeit bei flussbaulichen
Modell-versuchen," Schweiz. Archiv, 5 (1939), No. 8.
48
C. Camichel, E. Fischer, and L. Escande, Comptes Rendus, 199
(1934), 594. The experimental counterexample given by them is not
one of nearly horizontal flow.
105

jT
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill

(28)
— = div (hv), and = g grad h.
•^

If one defines p as h, and p as —

ah2
+ const., then one has a

well-known analogy with compressible flow, with y = 2 in

the equation of state. It is easily verified that these equations
are preserved, not only under Froude modeling, but also under
the change of horizontal scale with vertical scale fixed,

(29) x' = ax, y' = ay, t' = at; v, g, h constant.

A similar scaling law holds for any "nearly horizontal flow"
of an incompressible non-viscous fluid under gravity,

(!)
defined mathematically as follows.
Let u(x,y,z,t]\) be defined for all x/\, y/\,

^
0

^
1,
X
t/\ within with ho(x/\,y/\,t/\)
R,

region and

^
a

z
hi(x/\,y/\,t/\,\). Let = h0(x/\,y/\,t/\) be fixed wall, and

a
z

= free surface under gravity and inertia. Then

h,

if
a
z

u(x/\,y/\,z;t) = F(x,y,z,t) approaches limit as — we

0,
>
X
a
shall say that defines "nearly horizontal flow." For such
F

scaling by (29)
F,

valid.
is is

Scaling by (29) clearly not special case of (18). Other

a

examples where different scales are used in perpendicular

directions will be given later4' (Section 19; Chapter IV,
Section 7).
16. Modeling Cavitation
of

It important to avoid cavitation in the use of ship pro

is

pellers and of hydraulic turbines. Therefore only natural

it
is

that model studies should have been devised to reproduce this

phenomenon to scale.
Until twenty years ago, cavitation was generally stated to
depend on the homogeneous dimensionless parameter Q* =
P//4pV2 of Section 10, which directly suggested by ordinary
is

dimensional analysis.60 The incorrectness of this view (and

Algebraically, these examples can also be considered from the point
49

of view of dimensional analysis; see W. Williams, Phil. Mag., 34 (1892),

234-271.
50
See F. Lorain, L'hilice propulsive, Paris, 1923, p. 129; E. Bucking
ham, Jour. Am. Soc. Naval Eng., 48 (1936), 147-198; D. W. Taylor, The
Speed and Power Ship*, 3d ed. (1943), p. 17.
of

106
§16] MODELING OF CAVITATION

hence of the partial inspectional analysis of Section 10) is

immediately seen, if the role of the vapor pressure p, is con
sidered. This was implicitly observed in 1924 by Thoma,61
and the "cavitation number" Q = (P — Pv)/}4pV2 is now
generally used in place of Q*.
It is noteworthy that the use of Q can be justified theo
retically, by an inspectional analysis which neglects only
viscosity. For it is evident that a transformation of similitude
(18) preserves the equation of state (25.1) for water if and
only if

(30) (P - p.) = aW - H
7
That
is,

and only There no linear

Q.
preserves
it

is
if

if

homogeneous transformation of space, time, and pressure

which will do. This shows, not only that full inspectional
analysis to be preferred to dimensional analysis modeling
is

if
seriously desired, but also that more general groups than
is

the homogeneous linear "changes of scale" contemplated by

dimensional analysis may be important.
Thus, in a Froude model at a linear scale, with velocities
1
:

in the ratio y/-a, we can preserve both gravity and cavita

1
:

tion effects (but not viscosity), by transforming (P — p„) in

the ratio l:ot. This "reduced pressure Froude modeling"
now commonly used in model studies of ship propellers; in the is
model, vapor pressure may be far from negligible.
One caution should be observed in the use of cavitation
modeling. Water which has had all gas taken out of solution
capable of withstanding very large tensions.62 Hence cavi-
is

D. Thoma, "Experimental research in the field of water power,"

51

Trans. First World Power Conf., Vol. (1924), 536-551; see also H. B.
2

Taylor and L. F. Moody, Mech. Engineering, U (1922), 633-640. An

explicit statement was given by H. Lerbs, on p. 290 of Hydromechanische
Probleme des Schiffsantriebs, Hamburg; see also H. E. Rossell and L.
B. Chapman, Principles Naval Architecture, New York, 1947, Vol.
2,
of

p. 177.
" This importance of air content was observed by Spannhake, V.D.I.
Special Issue (1933). Systematic measurements have been made by
Numachi, Rep. Tohoku Imp. Univ., 13 (1939), No. Vuscovic (Escher
1.

Wyss News, I3 (1940), 83-90), showed that the increase in cavitation

with air-content did not either increase erosion or decrease turbine
efficiency.
107
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill

tation modeling may be considerably affected by reducing

the air-content of the water used; increased air content has
less importance.

17. Modeling for Entry into Water

For underwater ballistics, it may be important to model

the phenomena of "surface seal" and "down-nosing" which
accompany entry into water, as stated in Chapter II, Section
10. This leaves us with the problem of trying to reproduce
these effects to scale.
On the basis of experimental analogy, one might be tempted
simply to use reduced pressure Froude modeling, preserving
F and Q or Q*; this has indeed been proposed. However,
I am glad to state that, for once, the correct solution53 as
regards surface seal appears to have been given, not by an
engineer on the basis of physical experience, but by a "mathe
matician in his armchair," using inspectional analysis —
namely, myself [18]. The answer was suggested by the
following considerations.
A little reflection makes it plausible that it is the air density
which causes surface seal — there being a reduction of J^pV2
(p' = air density, v' = air velocity) in the pressure at the
neck of the cavity, which causes the splash and neck to con
tract. This is not scaled if p' is reduced by lowering pressure;
while if one does not lower the pressure, presumably the size
of the bubble after seal is not modeled.
Since one is not dealing with water vapor, the equation of
state is no longer (25.1), but is nearly

(31) p' = kp'y.

The same equation is satisfied (approximately) by underwater

gas bubbles; unless we are dealing with saturated gas, we do
not have p' = pv + Up'y. For this reason, we should use Q*
instead of Q.
Mathematically, the Equations of Continuity, State, and
Motion are preserved under the transformation

M For the experimental data, see Gilbarg and Anderson [241.

108
§17] ENTRY INTO WATER

Xi = ax, ti = at, Ui = VaU, Pi = <*Vi,

Vi = «P', Pi = P, Pi = P, ki = ak,

where suffixes denote transformed variables. Hence we could

obtain a model if we could get a gas whose equation of state
had the form p' = afcp'1•408, where p' = kp'lAW was that for air.
For example, air at a low temperature would provide a good
solution; but this is not practical from an engineering point
of view.
It appears to be more practical to use Freon, or some other
"heavy gas" with 7 7* 1.4, but having a density several times
that of air at reduced pressure and atmospheric temperature;
this has been done on a small scale (cf. [24]). One can then
only model p and dp/dp in the gas.
The modeling procedure outlined above presumably does
not model "down-nosing." If down-nosing is a viscous
effect, as suggested in Chapter II, it can only be modeled by
preserving R, and this is impractical. However, since the
maximum possible underpressure decreases to zero with
Q* (assuming tension p < 0 impossible for the durations
involved), down-nosing should at least be mitigated by
reduced pressure Froude modeling, with or without a heavy
gas.
The need for a morecareful analysis than that provided by
ordinary dimensional analysis is also illustrated by the dimen-
sionless parameter N = •\/F p'/p, which has recently64 been
shown to determine roughly whether entry into water is
accompanied by surface seal or deep seal. The derivation
of N as a criterion rests on the empirical fact that the time
Ti. required for deep seal is roughly proportional to y/L/g,
combined with the fact (deducible from inspectional analysis
of the inertial mechanism of surface seal) that the time T„
of surface seal is proportional to pL/p'V, where L is a typical
length and V a typical velocity. The condition TA, > T„
for surface seal therefore assumes the form N > Ncru. Using
ordinarily dimensional analysis, one would probably have
M G. Birkhoff and R. Isaacs, "The mechanism of cavity motion," to
be published as a Bureau of Ordnance Report. For deep and surface
seal, see Chapter II, Section 10, supra.
109
 MODELING AND DIMENSIONAL ANALYSIS [CH. HI

argued that the average pressure differences causing surface

seal and deep seal, respectively, should be proportional to
Kp'F2 and 2pgL, suggesting the use of the dimensionless ratio
N' = Fp'/p as the criterion for surface seal. This disagrees
strikingly with observation.

18. Mach Modeling

It has been known since the early experiments of Robin65

(1742), that the drag of bullets is not proportional to the
velocity squared; hence no form of inertial scaling is valid.
In the notation of Example 2, KD has a notable rise near the
speed of sound. As a result, K D has been commonly tabulated
as a function of v.
It was recognized from the beginning, that this was due to
the compression of the air, but the more rational tabulation
of KD as a function of the Mach number M dates since World
War I. It follows logically that there should be a temperature
effect on the range of projectiles, in addition to the effect of
density already obvious from the definition of KD. However,
this was apparently first stated explicitly after the first
World War."
Beginning about 1935, with the advent of high-speed planes,
the aerodynamicists have become interested in Mach modeling.
Whereas wind-tunnels operating at 100 ft/sec can be used to
simulate flight conditions up to 400 ft/sec, provided the
"effective" Reynolds number is properly controlled, they
give no idea at all of the compressibility effects which arise
at higher speeds. Since 1935, therefore, the aerodynamicists
and ballisticians have joined forces in the study of compres
sible flows.
The laws of Mach scaling were first derived by Langevin,"
" See Cranz [3], pp. 44-45, for an historical account.
" G. Darrieus, Mem. Art. Francaise (1922), p. 242; H. W. Hilliar, Dept.
Sci. Res. Exp. Report RE 142 /19 (1919); Darrieus states that the vari
ation in range due to this cause can be as much as 1%. The article by
Langevin cited below comes immediately after it in the same journal.
Mach modeling was justified, on the basis of dimensional analysis, by
Buckingham [41], pp. 275-278. For practical aspects, see R. H. Kent,
Mech. Eng., Sept., 1932.
" See footnote 56.
110
§18] MACH MODELING

from an "inspectional analysis" of the equations for a com

pressible non-viscous gas neglecting gravity, in the case of a
polytropic equation of state. I shall give a slight generaliza
tion of his results.
Going back to Theorem 3, we see that all transformations
of similitude preserve the equation of continuity. Obviously
also, the equation of state is preserved under any transforma
tion which preserves p and p at corresponding points. Hence,
in particular, it is preserved under the transformations

(32) xi — > aXi, t —> at; p, p, u unchanged.

By inspection, the terms Dui/Dt = Zut dui/dxt + dui/dt and

dp/pdxi in (19) are both multiplied by 1/a under (32). We
conclude
Theorem 6. The fundamental equations for any compres
sible non-viscous fluid are invariant under (32).
I recall the doubt, discussed in Section 11, regarding the
completeness of these "fundamental equations." This doubt
is confirmed by the fact that the important Rankine-Hugoniot
conditions for shock waves cannot be derived from the "funda
mental equations" of continuity, state, and motion.68 How
ever, since these conditions can be derived from the laws of
conservation of mass, momentum, and energy, which are
preserved (per unit volume) under (32), they are scaled also.
The scaling law (32) holds also in elasticity. In fact,
whenever the stress tensor is a function of strain only, and
not of the rate of strain, the scaling law (32) is valid. In
fact, kinetic energy per unit volume and (potential) strain
energy per unit volume are the same. Curiously enough, it
also holds in relativistic fluid mechanics.
The partial inspectional analysis, described in Section 10,
has been thought by some to give a basis for Mach scaling.
Let c = y/dp/dp denote the local speed of sound in (19),
and let C be the sound velocity of the free stream. Then
if the forces of viscosity and gravity are neglected, and we
write M = V/C, (19') becomes

68
The great mathematician Riemann was misled by the apparent
completeness of (19) and (24)-(25), to ignore this possibility.
lll
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill

n q'm>
(iy Ml
dt'
:
+ V uidM*
^ ^ dx'k m*
J- (A* J*di
\cj pdxt P

This suggests the plausible rule : scale at the same Mach number,
which for a given free stream is equivalent to (32).
However, this rule is false in general, if we consider behavior
in different gases (gases with different y in (31)), or even the
same gas under different temperatures and pressures. Com
pare for example dynamically similar adiabatic flows of a gas,
in which the free stream conditions correspond to two points
along the same adiabatic. By (18), v and p are proportional.
Hence by (19) (neglecting g and v), grad p is in a fixed pro
portion a(a), where a is the ratio of free stream densities.
Hence if F(p) — p(p) — p/ (p/ = free stream pressure),
F(ap)/F(p) = a(a) is independent of p. We thus have for all
a, F(ap)/F(p) = F(ap')/F(p'). But this clearly
p',

is
p,
equivalent to (15) of Section Hence by Theorem
6.

2,
F(p) = kpy, and so
Theorem For models compressible flows to be dynami
of
7.

cally similar at the same Mach number under all free stream
conditions, the equation state must have the special form
of

= fcpT + const.
p

(33)

It sufficient that be the same along all adiabatic curves.

is

Generalizations to non-adiabatic flow are obvious.

What about the effects of viscosity? We have already
seen the danger of neglecting viscosity, in incompressible
flow. In ballistics, skin friction most commonly of the
is

order of 10 per cent; this obviously affected by viscosity

is

and surface roughness. Otherwise, the Reynolds number does

not seem to be of much importance in ballistics, though the
contrary view has been widely disseminated.69
However, the Reynolds number does influence various
minor phenomena, such as the thickness of shock-waves, and

59 The theoretical conclusions in the well-known book The Mechanical

Properties Fluids, Blackie and Co., 1935, are based on single table
of

on p. 37 of Cranz [3]. Nothing remotely resembling the data of this

table has been found in the United States, with ordinary bullets (flat-
based or boat-tailed).
112
§19] LINERALIZED MACH MODELING
the X-shocks found by Ackeret.60 "Condensation shocks,"
due to sudden condensation of water vapor, are also certainly
determined by factors which certainly do not scale under (32),
let alone Mach modeling in general.

19. Linearized Mach Modeling

In the case of slightly perturbed uniform flow, more general

transformations of approximate (i.e. asymptotic) generalized
dynamic similitude are possible. This case includes "thin
wing" theory, and the theory of the resistance of "slender"
bullets.'1 I shall now derive the equations to be tested by
inspectional analysis.
Consider steady irrotational flow (with negligibly "weak"
shocks) of a compressible non-viscous fluid. Then the equa
tions of motion are the space-derivatives of the Bernoulli
equation, and so are satisfied if

— + = 0.
(34)
2 udu*
The equation of state p = /(p) is equivalent to dp = cHp,
where c = c(p) is the local speed of sound. Hence the equa
tions of motion and state are both satisfied if c(u) and p(c)
are defined from (34) by dp/p + 2ukduk/c2 = 0. This leaves
the equation of continuity, which can be written

(35) 0 = - div
' (pu)
v"~/ "" u
= div +' 7 w,•
~~y
Bp
P ^
Li pdxi

Substituting from the preceding equation, we get

(35') div u -I£ u,uh g- 0 [c

= c(u)],

as a single differential equation containing all the conditions

for compressible flow.
60
J. Ackeret, Mitt. Inst. Aerodynamik Zurich, No. 10 (1944).
n For thin wing theory, see N. Rott, Mitt. Inst. Aer. Zurich, No. 9, 247-
251; W. D. Hayes, Quar. Appl. Math., 5 (1947), 105-106; H. S. Tsien,
M.I.T. Jour. Math. Phys., 25 (1946), 247-251. For the resistance of
slender bullets, see Th. von Karman and Moore, op. cit., Chapter I,
Section 12.
113
 MODELING AND DIMENSIONAL ANALYSIS [CH. Ill

Now we are ready for asymptotic considerations. We let

(36) u = (U,0,0) + XV*.

Here U is the velocity of uniform flow (not velocity potential)

<t>
to be perturbed; X is a parameter, the normalized velocity

is
potential of perturbation. Actually, depends on but this

X,
*
does not appear in the notation. We also write u = d<t>/dx,
= d<t>/dy, w = d<t>/dz.
v

Since duk/dxi = \d2<t>/dxidxk, clearly

^.g-x^ OOT.

+
(37)

Again, since differentiable function of the velocity

c2

is is
a

squared, which U* + 2\Uu + 0(X2), we have = C2

c2

+
0(X), where the velocity of sound in the free stream.
is
G

Finally, by (36), div u = XA2*, and so (35') becomes

Qj

XV2* =
0(X)j (\U2
0
+

+
(38)
0(X2))•

Dividing by and simplifying, this reduces to

X

V2* =
•g
0

(38') + 0(X)

Thus as —
<t>

satisfies partial differential equation

0,
>
X

which approaches

v2^ =
(39)
M2^.
In the equivalent to V'2* =
< 1), this
(M
-
subsonic case
is

0,

where V'2 = d2/dx'2 d2/dz2, and x' = x/y/l M2;

+

d2/dy2
hence reduces under an affine transformation to the incom
it

pressible case. In the supersonic case (M > 1), we have

a

similar reduction to

which the wave equation. In either case, we get affinely

is

114
§19] LINERALIZ ED MACH MODELING

similar flow at corresponding Mach numbers around all affinely

equivalent models. The sonic case must be considered inde
pendently.62
Thus changing M is equivalent (at least in theory)•1 to
changing the "thickness ratio," for affinely similar flows.
Hence, besides the ordinary Mach modeling of (32), one can
scale two perpendicular directions independently, much as in
the theory of "nearly horizontal" flows.

" See Th. von Karman, "The similarity law of transonic flow,"
M.I.T. Jour. Math. Pkys., t6 (1947), 182-190. I have here omitted a
discussion of boundary conditions. Essentially, d4>/dn is given on a
surface 2, parallel to the x-axis, as U times the slope of the "thin solid,"
in a plane perpendicular to Z and parallel to the x-axis.
•3 In view of
Chapter I, Section 12, the theory may well be viewed with
some skepticism. Thus, it involves no shock waves.

115
 CHAPTER IV

Group Theory and Fluid

Mechanics
1. Introduction

In Chapter III, Itried to show that the group concept is

valuable in fluid mechanics, for three purposes. First,
through
" inspectional analysis" it provides a mathematical
basis for the use of models, which is far more adequate than
the "dimensional analysis" commonly appealed to. Second,
it provides tests for the validity for mathematical theories of
hydrodynamics, even in cases where it is impossible to inte
grate the partial differential equations involved in such
theories. And finally, like dimensional analysis (but more
generally), it often permits one to reduce the number of
parameters which need to be considered; thus it provides
important simplifications.
I shall now discuss its utility as a method for integrating
the differential equations of fluid mechanics, and indeed of
mathematical physics generally. Most of what I shall have
to say in this connection has been said in one way or another
in the mathematical literature. But
if,
somewhere as
believe, we have only begun to explore the applications of the I
group concept to differential equations, seems worth col
it

lecting these ideas in one place.

shall first describe what may be called the method of
I

search for symmetric solutions of partial differential equations.

Suppose that system of partial differential equations
is
2
a

invariant under group G; the method consists in searching

a

for solution which also invariant under G.

is
a

This method has been applied so often to physical problems

that surprising that should not have been more explicitly
it

it
is

recognized long ago.1 shall now illustrate its effectiveness

I

by various specific applications.

It seems to have been first formulated by K. Bechert, [54] and [55].
1

more explicit formulation has been given by L. Sedov [66], and

A

I.

116
§2] SYMMETRIC SOLUTIONS

2. Symmetric Solutions of the Heat Equation

As the method of "search for symmetric solutions" is not

restricted to fluid mechanics, I take first an application to the
heat equation.
Suppose we are interested in the diffusion of heat from a
point, in a medium of constant thermal diffusivity K. The
equation of heat conduction in solids is

dU r*n d2U
(!)
i-^U^KlJ-^ (n- 1,2, or 3);

here U is the temperature at the point x = (xi, • • •

,xn) at
time t.
Now since we have rotational symmetry, we look for a
solution of the form U(r,t), where r2 = 2xf. This exhausts
the purely geometrical symmetry of the problem. But we also
have physical symmetry, in the sense that, for any positive
number m, the differential equation (1) is invariant under the
group of self-similitudes

(2) r' = or, t' = aH, U' = amU,

of thermal space-time. This symmetry extends the classical
Law of Times. Hence we look for solutions U(r,t) of (1) which
are invariant under (2).
In the present case, since the group (2) consists of scalar
multiplications, we can use the language of dimensional
analysis. We have the two dimensionless variables x = r*/t
and V = U/tm/2. In any solution which is invariant under
(2), by the Pi Theorem we have V = F(x), and so

(3) U =
F'W(x)
K}
Our general principle, to be proved later, asserts that (1) will
have solutions of the self-symmetric form (3).
Knowing that such solutions exist, it is easy to find them.
We obtain the ordinary differential equation satisfied by

K. P. Staniukovitch [67]. I later, but independently, formulated the

same method (cf. [56]).
117
 GROUP THEORY AND FLUID MECHANICS [CH. IV

''(x), by the following series of direct computations, in

dimensions.

dt
l
[j^-rf"]
t»«-W'

d-^
V2U = + =

+
r^±d£ 2nF'].

Equating and dividing out suitable power of we get

t,
4KxF
" +
(2Kn + xW - =

|F
(4)

0
Since this an ordinary differential equation, can easily be
is

it
integrated numerically, for any real m, and "initial values"
of and F'. Thus &K = (4) has Laguerre polynomials
F

if

1,
as solutions.
However, not all thesj "symmetrical" solutions have equal
physical interest. Interest largely confined to solutions
is

for which — — so that limx_ «,F(x) = 0. We

U

as <x>
>
>
0

are also interested in the total heat, which proportional to is

Urn-ldr, and hence to aman = am+n, or to t(.n+n)/2.

/

One special case of interest that of constant total heat,

a is

corresponding to diffusion of fixed quantity of thermal

energy, initially at the origin. Here m = — n; we set
if

n = 2h, (4) reduces after collecting terms to

= 0.
x(4KF' + F)' h(4KF' + F)
+

(4.1)

This can be easily integrated; and F' tend to zero as

F
if

r-> oo, then \KF' +F = and F(x) = e-*'AK = e-"'**',

0

whence

= Clrn'*e-rtHKt
U

(4.1')

Another interesting case point source producing heat

is
a

(by chemical or radioactive process) at a constant rate,

beginning at = Here m n = or m = n — whence
0.

2,
2,
t

(4) becomes

4KxF" (4Kh + x)F' - h)F = 0.

+

+
(1

(4.2)
118
§3] PRANDTL-MEYER EXPANSION WEDGES

The integrals of this differential equation (obtained by other

means) are expressedin closed form in Carslaw and Jaeger,
Conduction of Heat in Solids, Section 104; the case n = 2 is
especially obvious, as then (1 — h) = 0. Even if no such
expression were known, they could be integrated numerically.

3. Prandtl-Meyer Expansion Wedges

Having shown that the method of "search for symmetric

solutions" applies to physics in general, I shall now concen
trate on some typical applications to fluid mechanics.
I shall begin with stationary flows of a compressible non-
viscous fluid. The differential equations of such a fluid are
invariant, as we have seen (Chapter III, Section 12), under the
group of Mach modeling

(5) xi — > axi, t —> at, p, p, u unchanged.

The method of search for symmetric solutions suggests looking

for flows which are invariant under (5); stationary flows will
be invariant under the group

(5') t ^> t + T,

all other variables unchanged, as well.

Plane flows invariant under (5) and (5') can most easily
be discovered using polar coordinates (r,0) and the associated
radial and angular velocity components ur, w». By definition,
invariance under (5) and (5') means that all quantities are
functions of 6 alone. Hence we write

ur = g(6), ue = h(6), p = p(6),

(6) p=/(p), *-rw-A

dp
dp

We shall call flows satisfying (6), wedge flows. They will be

generalized in Section 13 (see Figure 20).
The assumption of irrotationality amounts to <f>urdr + uerdO
= 0 for all closed curves, whence 0 = d(ru$)/dr — dur/dd =
h — g', and we get

(7) h = g'.
119
 GROUP THEORY AND FLUID MECHANICS [CH. IV

Since the flow is irrotational, the Equations of Motion are

equivalent to the Bernoulli equation, which we can write as

or in the differentiated form

(£)•
0 = udu + = g'(g + g") +

e«
(8.1)

dj•
With polytropic equation of state = kpy + po, = ykpy-1,

c2
p
a

and so fdp/p = kypy-l/(y — = c2/(y — 1); hence in this

1)
case

C?
= l(g* g'*)+
+
C

(8') w

(y
»
'
'
'

1)
2

So far everything equally valid for irrotational flows along

is

meridian planes through an axis, in which all physical quan

tities are functions of the latitude only. We shall call such
flows, conical flows.
Now with wedge flows, the Equation of Continuity, div (pu)
= can be written in the form
0,

= — (prur) — (pu,) = pur (put)' = p(ur u») + p'u,.

+
+

+
0

Substituting from (6)-(7), this equivalent to

is

(^)

g") = 0.
+

+
g'
(g

(9)

Multiplying (9) by and subtracting from (8.1), we get

g'

-
(£)

+ g')(c g') =
0.

(9')
(c

Clearly (9)-(9') are equivalent to the equations (8.1), (9)

of

motion and continuity. We have two families of solutions.

Case = 0. Substituting in (9), g" + =
0,

whence
p'
I.

= —
cos a). Using (6)-(7) to get w», we see that
A

ur
(6

this uniform flow with constant vector velocity.

is

120
§4] TAYLOR-MACCOLL CONICAL FLOWS

Case II. g'2 = c2,

or 0* = +c. We see that the radii
d = const, are characteristics, in the sense that the component

of velocity perpendicular to them is always the sound velocity

c. This gives so-called Prandtl-Meyer2 expansion waves,
which are often seen to fill wedge-shaped regions in Schlieren
pictures. At the ends, they join onto regions of uniform flow.
In the polytropic case, substituting c2 = g'2 in (8'), we get
immediately the differential equation

(10) (7 + I)?" + (7 - 1)9*

= 2(7 - 1)C.

This can be easily integrated, the nature of the solutions in

the adiabatic case y > 1 being quite different from what it is
if 7 = 1, if— 1 < 7 < 1, or if 7 = — 1 (circular flow). Evi
dently in the non-polytropic case, we can still obtain equation
(9) and c = g'. Using (8) to solve In p = ^(c) = i>(g'), we get
(g + g") + g'g"^'(g'), which can be integrated numerically.
Hence Prandtl-Meyer expansion waves are mathematically
possible with a general equation of state.
Mathematically, Prandtl-Meyer "compression waves" are
also possible, with flow in the direction of increasing pressure.
But such waves do not seem to occur physically (Reversibility
Paradox).
4• Taylor-Maccoll Conical Flows

Inn dimensions (the physically interesting case is of course

n = 3), the Equation of Continuity for stationary axially
symmetric flows assumes the form

0 = •r- (prn-i cosn-20 ur) + — (prn-2 cos—80 «,)

or off
= (n — l)p cosB-20 rn-!Mr + prn-!(cosn-20 «>)'
+ p'r—* cosn-*0 ut.

After dividing by prn-2 cosB-i and substituting from (6)-(7)

d,

(we recall that equations (6)-(8') are valid in space), we get

(11) (n
- 1)0
- -
(n 2)0' tan d + g" + g' fcl = 0

»
[63]; see also T. Meyer, V.D.I. Forschungsheft, 62 (1908), 31-67.
121
 GROUP THEORY AND FLUID MECHANICS [CH. IV

in place of (9). Here (8.1) is still valid; it is equivalent to

(8') in the polytropic case. In this case, by (8.1)-(8'), we
have

Substituting in (11), which can be written

(n- 2)(-g'

(£}
(11') (g" +g) + tan + g) + g' =

0
B

we get

(g" +
g)[l+ ^-jj -

+
(12) <H

|j

c]
(</2

- 2)(-0'tan0 +

'J
+ =0

(n

sO
This has been integrated numerically for = 1.408
(air) in

a
celebrated paper by G. Taylor and W. Maccoll [68].

J.
I.

3
An attempt was made to find a flow satisfying (11)— (12),
which would be tangent to rigid cone (projectile head) and
a

be separated by a conical shock wave of constant intensity,4

also invariant under (5), from region of uniform flow.
a

It was deduced that the resulting "conical regime" depicted

in Fig. 18 was only possible for Mach numbers considerably
greater than one. The predicated limits of the conical regime,
the pressure on the conical head, and the angle of the attached
shock wave as function of Mach number have been closely
a

confirmed experimentally.

Non-stationary One-dimensional Compressible Flows

5.

There also exist important classes of non-stationary flows

having the internal symmetry (5), and which can therefore
be found by the method of search for symmetric solutions.
shall here discuss the one-dimensional case only; flows in
I

higher dimensions, which are invariant under (5), will be

discussed in Sections and 10. shall also assume poly
8

a
I

tropic equation of state, — p0 = kpy.

p

B,

See also [8], and the bibliography of [60], p. 449, Part especially
3

J.W. Maccoll, Proc. Roy. Soc., A159 (1937), 459-472.

Such a shock yields irrotational flow and a single equation of state
1

behind the shock front.

122
§5] COMPRESSIBLE FLOWS

In the one-dimensional case, it is convenient to use Lagran-

gian coordinates. If we let a measure the mass between a
particular material point and a general material point, and
denote position by x = f(a,t) and density by p = h(a,t), then
the Equation of Continuity is equivalent to having the specific
volume a = 1/p = df/da.

5hock Wave

Conical Projectile

Fig. 18. Conical flows of Taylor and Maccoll.

Since velocity u = df/dt, invariance of the flow under (5)

is equivalent to f(aa,at) = af(a,i), and hence to f(a,at) =
af(1,t) = ag(t). Setting r = t/a, this is

(13) x = f(a,t) ag(r) =

=ag(£)=
i].

[r
Thus (13) expresses our symmetry condition.
It remains to consider the Equation of Motion. In Lagran-
gian coordinates, acceleration
is
(0)
(g'
d

^/ _ _ ,"(t)
dt2 dt a

Again, dp/p dx = dp/p dx, being held constant. But

c2

d/p dx = d/ da, and with polytropic flow, = ykpP1-l; hence

c2

123
 GROUP THEORY AND FLUID MECHANICS [CH. IV

dp
_ ykpt-1 dp
_ - 7V+1 3(l/p) = _ ykpy+1 d2f
p dx da da da2

But if f(a,t) = ag(t/a), then d2f/da2 = t2g"(t/a)/a3, by direct

calculation. Hence the Equation of Motion is equivalent to

<14> o-g + Jg-itf'M-tf^V'M}.

Equation (14) has two families of solutions.
If ff" = 0, then / = a[Ci + Ct(t/a)] = da + Ctt. This
implies uniform flow, u = a = const., and is uninteresting.
Otherwise, 1 = ykpy+1t2, which implies

But d//da = — (t/a)g'(t/a); hence the condition is

#(</a)

(15) g - rg' = [ykt2]u^+l\ unless y - -1.

This first-order linear non-homogeneous ordinary differential
equation can easily be integrated formally. The general
solution is

(15') s-(t) = Ct + At2/*+1\

where A = [(y + 1)/(-y - and C is arbitrary-

1)][7fc]1/(l-+1>,
There no solution = —1, since then
\y\

provided
is
1.

if

5^
7

t = const, by (14'). If = the general solution g(j) =

is
1,
7

Ct + \/yK t In t.
6. Spiral Viscous Flows

The method of "search for symmetric solutions" also

is

extremely useful in obtaining viscous flows. We illustrate

by the classical case of the "spiral flows" of incompressible
viscous fluids.
It well-known that the equation of continuity equiva
is
is

lent, in the case of plane flows, to the introduction of "stream

a

= —
function" M2&1) such that (—dV/dy,dV/dx)
V

f(uidx2
the velocity vector. Then V2F = du^/dxi — dui/dx2 the
is
is

vorticity. Moreover the Navier-Stokes equations of motion

124
 §6] SPIRAL VISCOUS FLOWS

are equivalent6 for such plane flows to

U0; "V
d(x,y) r d(r,6)

r\_dr d6 d6 dr j
Here d(p,q)/d(x,y) = p,qy — qxpy is the ordinary notation for
the Jacobian; v is as usual the kinematic viscosity p./ p.
Dimensional analysis tells us that under geometrically
similar conditions, the behavior of incompressible viscous
fluids will be a function of the single dimensionless parameter
R, which plays the role of KT2/D in the example of Section 2.
We now look for self-similar plane motions. To this end, we
consider the spiral group

(17) r' = e"r, 6' = 6 + ca,

where the parameter c is a fixed real number.
The transformations (17) are self-similitudes of space.
Since p is fixed, they will make the motion self-similar at the
same R, if and only if rur and rue are the same at corresponding
points. But differences in the stream function V are pro
portional to distances times velocity, since dV = (dV/dx)dx
+ (dV/dy)dy. Therefore differences in V must be invariant
under the spiral group (17). Hence

(18) V(e°,x + ca) - V(e°,ca) - V(l,x) - 7(1,0) = F(X),

is a function of x alone, for any a. By similarity also, V(e",ca)

is proportional to a for self-similar motion. Hence

(18') V(r,6) = V(e°,x + ca) = Ca + F(x),

where a = In r and x = " — c That

is,

In r.

V(r,6) = In + F(6 - In r).

C

(19)
r

This expresses the motion in terms of an arbitrary constant

C

and function of one variable.

F
a

This result may be derived from Chapter III, (26), since curl =
5

0
g

in conservative field, and since = because = = d/dx3 =• 0.

&
•
V

£i
a

The result may also be found in [10], p. 536, Ex.

7.

125
 GROUP THEORY AND FLUID MECHANICS [CH. IV

As before, we now substitute in the general differential

equation (16); the calculations follow. In general, writing
X = In r, we have

V2 = + whence VW = e-^(c2 + 1)F".

\z (^ ^}
Since d/d(x,y) = 7-2d/d(X,0), we infer by substitution in (16),

r2j_dX dd dd d\ J•

Carrying out the with V = CX +

F(6 -
cX), we get
operations indicated,

+ 4cF'" + 4F"]
(21) *[(c2 + 1)F1V
= (C -
c/')F'" + 2F'F" + eFF".
This is the ordinary differential equation obtained by Oseen;6
it would be difficult to find another equally simple motivation.
Equation (21) can be made somewhat less repulsive through
the substitution G = F'\ moreover it is satisfied whenever
F" = 0. In any case, solutions can be found by numerical
integration.

7. Laminar Boundary Layer

We next consider the plane laminar flow parallel to a fixed

semi-infinite flat plate, placed on the positive x-axis. Clearly,
if mm is the free stream velocity, we have the boundary conditions

(22) w = v - 0 on 2/=0, x ^ 0,

and

(22') lim^-, » u(x,y) = ux, Iim,^ . v(x,y) = 0, for any fixed x.

We also have the equation of continuity

«
See C. W. Oseen, [Link] Mat., I-II, 1927-1928, or [65], ChapterII.
126
§7] LAMINAR BOUNDARY LAYER

and the Navier-Stokes equations (Chapter I, (3*))

(24) !g»,
dt
v«u-±fe,
p dx
etc.

Here as d/dt is the substantial derivative d/dt +

before
ud/dx + vd/dy; gravity is neglected.
We shall now simplify the preceding equations by the thin
boundary layer approximation of Prandtl, namely, that the
flow is nearly parallel to the plate, and changes rapidly when
one moves perpendicular to the plate. This implies7 that
u » v and d/dy ^> d/dx. It follows, since the pressure
gradient is zero in the free stream, that we can set dp/dx = 0
in (24); moreover in the steady state case, d/dt = 0. Sub
stituting, (24) becomes

du du d2u
t- ,
—
+ vv»£ = —,,
(24*) u^
dx dy v^.
v
dy1

Now consider the symmetry of the problem. Since we are

studying asymptotic conditions, we may hope to treat x and y
as independent dimensions, and to find symmetry under a
non-trivial subgroup of the four-parameter group of affine
transformations

x —> ax, yu, v — > Sv.

(25) y—>0y, u—>

Because of (22'), 7 = 1; and in fact substitution in (22)-(22')-

(23)-(24*) shows that

(25*) x^>px, y^>Py, u->u, t>->-

defines the subgroup specified above.

Having found this symmetry (or guessed it by "physical
intuition "), we can apply the method of "search for symmetric
solutions," and look for solutions which are invariant under
' For more thorough discussions of the approximation, see R. von
Mises, Zeits. ang. Math. Mech., 7 (1927), 425-431; V. L. Streeter, Fluid
Dynamics, McGraw-Hill, 1948, pp. 242-244. The method of search
for symmetric soutions works for other velocity distributions outside
the boundary layer. See D. R. Hartree, Proc. Comb. Phil. Soc., 33 (1937),
223-229; S. Goldstein, ibid. 3B (1939), 338-340; W. Mangier, Zeits. ang.
Math. Mech., 23 (1943), 241-251; J. Oldroyd, Phil. Mag., 36 (1945),
587-600.
127
 GROUP THEORY AND FLUID MECHANICS [CH. IV

(25*). To do this, we first find, by inspection, a space

variable »j = y/y/x which is invariant under (25*), and use
it and x as independent variables. This leads to the classical
solution of Blasius [58] ; I shall now sketch the details.
We first transform (24*) to the new independent variables
x, i). In the convenient thermodynamic notation for partial
derivatives, whereby subscripts denote variables held fixed,
we have

- ic
(26)
\i\ =
(£), (in): w.
=
v*~ w.•
This may be conveniently proved by substituting in

*- $.* +
(©.*

We then set u = F(i)). This gives, by (26), du/dx = -i)F'/2x,

du/dy = F'/y/x, d2u/dy2 = F"/x. We next eliminate (23)
as usual, observing that it amounts to asserting that udy — vdx
dV, where V is the stream function.
is a total differential
Integrating V = Judy = y/x /wdij from the boundary stream
line, 7 = 0 when y = 0, we get V = \Zxf(v), where/' = F.
We now calculate

.. "te _^L
y/x
2
+ J*L
y/x 2
by(26).

Substituting the preceding expressions in (24*), and /' for F,

we get, after multiplication by 2x and obvious simplifications,

(27) 2pf" + ff" = 0.

This can be integrated numerically, relative to the boundary

condition /'( oo) = u„,/(0) = /'(0) = 0.

8. Other Symmetric Flows

The preceding examples should give a clear picture of how

the method of "search for symmetric solutions" can be
128
18] OTHER SYMMETRIC FLOWS

applied mathematically. I should now like to illustrate

further its importance for fluid mechanics by mentioning
briefly some other problems to which it has been applied, but
which do not always lead to boundary value problems involv
ing ordinary differential equations.

Solid Cone

Fig. 19. Vertical impact of cone on water.

Consider the entry at constant speed of a wedge into water

(Fig. 19), at sufficiently high speeds so that gravity can be
neglected. As in Sections 3-5, the transformation
x —> ax, y —> ay, — > at; p, p, u unchanged,
(5) <

leaves inertial fluid mechanics unchanged — indeed, we could

suppose the fluid compressible! Hence the method of "search
for symmetric solutions" leads one to look for solutions of the
form

(28) U(x,y;t)
-<4i>
This method of reduction from three to two independent
variables is the basis of a well-known paper of H. Wagner on
the impact of seaplane floats.8 We can reduce to functions
8
H. Wagner, Zeits. ang. Math. Mech., 5 (1925), 26.

129
 GROUP THEORY AND FLUID MECHANICS [CH. IV

of a single complex variable, by the considerations of Chapter

II, Section 2, but we have complicated boundary conditions.
The same method is evidently applicable to a cone entering
the water at a constant velocity, suggesting

V
U(x,y,z;t) = t*

(*,

Vj,
(28')

i.e. a reduction from four to

independent variables.
three In
the case of right circular cone entering the water vertically,
a

solutions having the circular symmetry of the hypotheses can

be expressed, through

(28") U(x,y,0;t)
-<4f>
in terms of single function of two independent variables.
a

Similarly, the invariance of the theory of compressible

non-viscous fluids under changes of space-scale suggests the
search for the so-called "conical flows" invariant under
the same group. This suggestion due primarily to A.
is

Busemann,9 and suggests the consideration of flows of the type

(29) u = u(4>,d)

in spherical coordinates. Such flows occur with delta wings,

because delta wings also have conical symmetry.
In the preceding examples, the reduction in the number
real independent variables involved in the basic partial of
differential equations was to n = whereas in the examples
2;

of Sections 2-7, we reduced to the case n = of ordinary

1

differential equations.

General Local Existence Theorem

9.

It natural to ask, how generally can one expect the method

is

of search for symmetric solutions to give mathematical solu

tions to physical problems? stress the word mathematical,
I

because we have already seen (Chapter Sections 14-15)

I,

that the symmetric solutions may well be unstable, so that

N.A.C.A. Tech. Memo., 1100 (1947), and refs. given there; also,
9

See
A. E. Puckett, Jour. Aer. .Set., 13 (1946), 475-484.

130
§9] EXISTENCE THEOREMS

apparent symmetry in the mathematical statement of a

problem need not be reflected in symmetry of the real physical
solution.
We shall begin with the case which is simplest from the
mathematical point of view — namely, where the symmetry
is so great that we are left with only one independent variable,
so that the partial differential equations describing the internal
conditions reduce to an ordinary differential equation, as in
Sections 2-7 above.
In this case, we have well-known theorems for differential
equations of the form

(30, g-4.»., ••,£*}

which prove the following fact, under extremely general
conditions. For any "initial values" of y, y', y", . . .
dn-1y/dxn-1, at x = a, there is one and only one "solution"
of (30), in some neighborhood of x = a. On the other hand,
it is obvious that
yn + y2 + 1 = 0
(31)

has no solution, even locally, in the real domain.

It may be hoped that in "well-formulated" problems arising
from physics, absurdities of the type given by (31) will not
occur. It may also be hoped that we will not be left with
two ordinary differential equations — which will not of course
usually have any common solution. However, we cannot
hope to avoid difficulties inherent in the mathematical prin
ciple that solutions compatible with given boundary con
ditions need only exist locally. An outstanding example is
furnished by the "conical flows" of Taylor and Maccoll
(Section 4) ; these authors showed that the mathematical limits
for the regime of the conical flows coincided very nearly, with
the physical limits for "attached conical shock waves" having
the conical symmetry postulated. Another example is fur
nished by Prandtl-Meyer expansion waves.
Thus even in the simplest case, of reduction to ordinary
differential equations, mathematical solutions need exist only
locally, and then only for the "initial value problem." It is
131

<
 GROUP THEORY AND FLUID MECHANICS [CH. IV

comforting to find that at least this mild local existence and

uniqueness theorem is valid quite generally.
To see this, recall10 the Cauchy-Kowalewski existence
theorem for analytic partial differential equations, of the form

(32)
-at FiV"Wt'to£te,' )
In words, suppose that the time derivatives of all functions
<t>i(x;f) [x = (xi, . . . ,£n)] involved, can be expressed in terms
of space derivatives. The Cauchy-Kowalewski theorem asserts
that (32) has one and only one local analytic solution for each
set of given analytic initial conditions at t = 0.
But now suppose that (32) is invariant under a group G.
Let <(>i(x;0) = C?,(x) be a set of analytic initial conditions invari
ant under G. Then the unique solution defined locally by the
preceding theorem, will also be invariant under G. Hence we
get a local existence (and uniqueness) theorem for the reduced
differential equation obtained by the method of search for
symmetric solutions, whenever we have one for the original
differential equations.

10. Group Theory and Separation of Variables

Solutions of physical problems possessing internal sym

metry, with respect to a suitable group, can usually be given
a simplified mathematical form by the introduction of suitable
variables associated with this group. I shall now show how
this leads to various "separations of variables" of great use
in fluid mechanics.
Thus consider again the invariance of the theory of a non-
viscous (but not necessarily incompressible) fluid under the
group

(5) t —> at, xi —> axi, p, p, u fixed.

By definition, individual "self-symmetric" flows, invariant

under the group (5), can be expressed in the form

(33) m = /.(jc), p = p(x), P = p(p) = p(p(x)),

10 J. Hadamard, "Le probleme de Cauchy," Paris, 1932, Chapter I.
132
§10] SEPARATION OF VARIABLES

where

(34) X
= (xi,X2,X3) =
[j, j, j) =
j•

Clearly, (34) is a special case of

(35) = h(t)x.
K

We shall now find all flows of a non-viscous fluid which admit

the formal "separation of variables" expressed by (33) and
(35); we shall show that, in this case, (35) necessarily reduces
to (34). Gravity is supposed negligible throughout.
Formally, it is evident that (35) implies
,.. dF =
h' dF , dF hdF
(36) ^7
at TX*^—
h dxk
and X7=/tT-'
dxi dxi

for any function F(x). Hence the equation of continuity,

dp/dt + div (pu) = 0, is equivalent to

if (33) holds. Likewise, the equations of motion for a non-

viscous fluid are equivalent to

(370
■»[(*)*£.
+ »* + £]_*
If
h'/h2 is a constant in time, say —C, then 1/h = C(t — to).
Hence, after an obvious change of time origin and unit, we
can reduce to the case of flows satisfying (34), and hence
possessing internal symmetry under the group (33).
Otherwise 2x* dp/dxk = 2x* dUi/dxk = 0, as can be shown
by differentiating (37)-(37') partially with respect to time, for
fixed £• In this case, in spherical coordinates, the radial
derivatives of x and u vanish, so that x and u, and hence
p(p), are functions of direction x/lxl (i.e. of latitude and
longitude) only. Substituting back in (33) and (35), we see
that they are stationary flows, and satisfy (33) and (35) for
any h(t), including the h(t) = 1/t of (34).
In summary, the method of "separation of variables,"
expressed by (33) and (35), gives only the flows possessing
133
 GROUP THEORY AND FLUID MECHANICS [CH. IV

internal symmetry under the group (5). Hence, in this case,

the methods of separation of variables and of search for sym
metric solutions are effectively equivalent to each other. We
shall see shortly (Section 12) that this equivalence does not
always hold. But first we shall derive an interesting property
of flows satisfying (33) and (35), from a slightly different
formal treatment.
In the case of irrotational flows, (33)-(34) are equivalent
to the slightly less trivial "separation of variables"

(34') U =
tF(x)
-"®
In this case, clearly w,• = dU/dxi = dF/dxi) and (37') is equiva
lent to

(37») _x^ + ^ + 4- = o,

since h'/h2 = - and dui/dxk = d2F/dxk dxi =

- x* dF/dxk, dW/dxi dxk
1
= duk/dxi• Moreover dU/dt = F and so

d(dU/dt) xkd*F = _ xkdUk

dxi dxi dxk dxi
Hence (37") is equivalent by a space integration to

(38)
^f + lVlNU f /?"
--
' ^C•

Here C = C(t); but by (33) C = C(jc) also; hence C is a

That all irrotational flows of
is,

constant. compressible
a

non-viscous fluid, which are invariant under (5), satisfy the

generalized Bernoulli equation (38). In the incompressible
case, one deletes the assumption = p(p) in (33). This
p

deletion leads to the usual Bernoulli equation for accelerated

motion in an incompressible fluid, with = C(t).
C

11. Application the Special Group Mach Modeling

of

of

My intention
is,

however, less to enumerate all known

results which can be regarded as applications of group theory"
11
Others, closely related to the considerations of Sections
8,
3,
4,

10,
are found in G. Taylor, Reps, and Mem. Aer. Res. Comm., 1381 and
I.

134
§11] SPECIAL GROUP OF MACH MODELING

than to show the variety of such results. Accordingly, I shall

now discuss a quite different "separation of variables" sug
gested by the special group of Mach modeling, applicable in
the case of polytropic equations of state, and described in
Chapter III, Theorem 7. This is the group

x -> ax, t^>pt] u ->

(|J u,
(39)
p -> hp, (p - p0) -> k&y(p - p0),

which automatically leaves invariant the polytropic equation

of state (p — p0) = kpi, and the equation of continuity. It
leaves invariant the equations of (non-viscous) motion, if and
only if a*-1 = P/a*.
Clearly, all flows which are invariant under a subgroup of
(39) which involves time, admit of a separation of variables
of the form12
Ut = </.(*), p = <v(jc),
(40)
p = kpy + p0, where x = tTx.
We shall now discuss flows satisfying (40).
Such flows always satisfy the equation of state. The equa
tion of continuity, dp/dt + div (pu) = 0, reduces for such
flows to

<«>
H»+»£+<~'['.£+e&])-°•
This is equivalent to a choice between the equations:
c + t + 1 = 0 and
(42) , de de dfk
dxk dxk dxt
and the alternative equations

(43) „ + n£.A£ +
.£.o.
1382 (compressible circular and spiral flows), and in [66] and [67]. Still
another was treated above in Chapter II, Section 14.
"This case includes the "separation of variables" introduced with
out motivation by Love and Pidduck, Phil. Trans., A222 (1922), 167-
226; by R. H. Kent, Physics, 7 (1936), 319-324, and Am. Math. Monthly,
43 (1941), 13; and by Courant-Friedrichs in [60], p. 419.
135
 GROUP THEORY AND FLUID MECHANICS [CH. IV

The alternative (42) asserts that the "dimensions" of u equal

those —t of x, minus those of t, and that a certain partial
differential equation of "stationary flow in x-space" holds.
The alternative (43) asserts that div (ef) = 0, in a stationary
flow in x-space, and that the radial derivative of e in x-space
is J/e/lxl* whence e(jj) = Ijcl^fa/lxD, where E is a function
of direction (latitude and longitude) only.
The equations of motion involve the three terms

P dxi dxi dxi

oxk dxk

For the sum of these terms to vanish, we have a choice between

the following alternatives :

e + r + 1 = 0, (y - l)r) = 2t, and

(«) *~* +« + » + .»>£-«,
or

(45) « + t+1 = 0, e = const.,

and
, , (ft + txk) dfi
tfi +
_ u'
a£
or

(46) e
- 1 = r + (7
- 1)i», /* ~ = 0,

and

key-*p- + tf< + txt&=0,

dxi dxk
or

(47)
2, = (7 - 1),, kei-* j| ■+ h |£ - 0,

and

+ =
rxk^
rft

0.

dxk
136
§12] VISCOUS FLOWS

The condition that (y — 1)ij = 2e is simply the condition

that the Mach number w?/c2 = u?/(dp/dp) be a function of

X which is independent of time.
Thus we see that the formal "separation of variables,"
suggested by the special Mach group (39), yields six alterna
tive cases, of which that expressed by the simultaneous con
ditions (42) and (45) alone corresponds to invariance under
(39)! This case involves fewer simultaneous differential
equations than the others, and hence is the most interesting;
it includes the flow described above in Section 5. In fact, I
have been unable to discover any non-trivial flows in the other
five "degenerate" cases.
However, in the viscous case, we do have significant flows
in the "degenerate case" arising from a corresponding sepa
ration of variables, as I shall now show.

12. Separation of the Time Variables: Viscous Flows

Consider the flows of an incompressible, viscous fluid which

are self-similar at all times, in the sense that

(48) Ui = g(t)fi(x), where x

= MO*.

Clearly the equations of state and continuity are together

(Chapter I, Section 13) equivalent to the single condition
div u = 0; since g and h are non- vanishing, this is equivalent
to

<«>
2i-°.
from which t has been eliminated.
It remains to consider the Navier-Stokes equations of
motion; by the remark of Chapter III, Section 12, we can
neglect gravity. The other terms are, by direct computation
from (48) in (24),

137
 GROUP THEORY AND FLUID MECHANICS [CH. IV

dp fa(t)h(t)\ dp ., , .

Hence the Navier-Stokes differential equations can be

expressed in the "separated form"

(50) J/.^OflWX)-O,
where Fi = Ft = gh'/h, F3 = p*A, Ft = F6 = gh\
g',

afc,
and

e.-A y^li
(

and G6 =
dxi *-< to
Clearly condition (50) equivalent to the requirement that
is

the vectors = (Fj,/WV*)

and = (ffhOt,O3,0t,Ot) be G
F

confined to orthogonal subspaces. Depending on the number

of linearly independent relations satisfied by the G„ the
"F-subspace" spanned by the may conceivably have

3,
1,
2,
F

or dimensions. shall first investigate the non-degenerate

4

Fj I

case, when all the are proportional, so that the F-subspace

has one dimension.
We can make Fi proportional to F3, by making a = ag2.
Again, = (we can ignore constant factors) equivalent to
is
h
g

the proportionality of F3 and F6, and implies the propor

it

tionality of Fi and f2. The remaining condition that Fi

is

and F3 be proportional, or that = (—p/2)g2h for some

g'

constant Since = this amounts to —2g'/g3 = or

ft,
0.

h,
g

1/g* = — By properly choosing the origin and scale

l$(t to).
of time, this can be reduced to = l/\/t, and hence to
g

-^1

Pk)
[•jc

(51) U<(x;0 = and = =

p

-^/<(X)
—
T- =
+

+
/,•

3"
/*

(51 X*
A
2 jj

"

a a
*•
\I

/
)

3x*/ dxi
I

dx* Po ax*

Hence we have again reduced the number of independent

138
§13] "inverse methods"
variables by one — by a search for a more general type of
symmetric solutions! I omit the discussion of (51') itself.13
This corresponds to the invariance of the Navier-Stokes
equations under t — » at and x —> /3x, provided the Reynolds
number R = vd/v « (0/a)fi [v = const.] is unchanged — so that
/3
<* y/a.
The solutions (51) are precisely the flows invariant
under this group.
There are also ' ' degenerate ' ' solutions of (50) . For example,
consider the "parallel flows," parallel to the avaxis, with

(52) mi = g(t)fi(y,z), u2 = w3
= p = 0.

In this case, (49) is always satisfied. Since h = 1, Ft = 0;

since p = 0, Ft = 0. What matters more, G3 = /i dui/dx = 0
for all i; hence F3 is unrestricted. It remains to satisfy
q'U = vg(d2fi/dy2 + d2fi/dz2), which, g being a function of t
and/< of x = (x,y,z), reduces to g'/g = —k and

(52') m = e-»My,z), where £& + **- -„*/,.

This is the well-known14 exponential decay of parallel viscous
flows, for example of flow in the two-dimensional channel
—a ^ y g a, with Mi = e-ki cos iey/2a, with k = jr2/4a2i'.

iS. "Inverse Methods" andGroUp Theory

The flows described in Sections 10-12 are characterized

formally by the "separability of the time variable." They
are dynamically similar to themselves at all times. This
property gives a simple group-theoretic characterization of
the method of separation of the time variable, which is not
limited to the groups discussed in Chapter III, which were
the only groups used in the applications of the method of
"search for symmetric solutions," discussed in Sections 3-8.
" In the plane, (49) is equivalent to the existence of a stream function,
and p can be eliminated from (51') by using curl (grad p) = 0. This
leaves the fourth-order partial differential equation (16).
"
See G. I. Taylor, Phil. Mag., Jfi (1923), 671-674. A similar expo
nential decay of circular flows is possible, with p = p(r) sufficient to
account for centripetal acceleration. Cf. [65], and Ratip Berker, Sur
quelques cas d'inUgration des Equation du mouvement d'un fluide visqueux
incompressible, Lille, 1936.
139
 GROUP THEORY AND FLUID MECHANICS [CH. IV

It may be that the method of "separation of variables" has

also a deeper group-theoretic foundation than that expressed
in the self-similitude of products such as F(x) J(y,z), as x
changes. An interesting case in point is furnished by the
cases in which separation of variables is possible in the Laplace
equation V2C7 =0 in space. Redheffer has recently shown
[64] that in all cases one family of coordinate surfaces must
consist of parallel planes, planes through a line, spheres
tangent at a common point, concentric spheres, the plane and
set of spheres obtained when one set of bipolar coordinates is
revolved about the line joining the poles, or a set of spheres
all passing through a single fixed circle. These are evidently
all equivalent under the group generated by inversions in
spheres to parallel planes, planes through a line and con
centric spheres — i.e. to one set of coordinate surfaces for
Cartesian, cylindrical, or spherical coordinates. This suggests
that the problem could be attacked directly by conformal
methods, invariant under the conformal group.
However, though this idea is very attractive, it is still
only a vague conjecture. (In this connection, it may be
remarked that the Schroedinger-Infeld factorization method
(see [62]) in ordinary differential equations definitely does not
depend on group-theoretic ideas.)
A safer, if less satisfying, explanation of the applications
of the method of "separation of variables" to fluid mechanics
is in terms of the concept of an inverse method, recently dis
cussed systematically by P. Nemenyi.16 Whenever group
theory indicates the existence of flows with separated variables,
or any other property P, an a priori postulation of property P
will yield these solutions at least, and possibly others.
For example, group theory assures us of the existence
(locally) of Prandtl- Meyer expansion waves, in which

,pi. the vector velocity is constant along each straight

line of a one-parameter family.
The "inverse method" consists in finding all stationary
irrotational flows of a compressible non-viscous fluid, which
have property PI. This can be done as follows.
15
Nemenyi, "On inverse and semi-inverse methods in the mechanics
of continua," Proe. Int. Mech. Congress, London, 1948.
140
§13] INVERSE METHODS

We know (Chapter I, Section 5) that with stationary irro-

tational flow, the equations of motion are equivalent to the
Bernoulli equation }4u2 + fdp/p = C. Hence, by numerical
integration, there is for each "stagnation pressure" (i.e.
constant of integration) one and only one pair of functions
p{u) and p(p(u)) such that the equations of state and motion
are both satisfied. Moreover the flows with property PI are
the flows with such p(w) and p(p(u)), which are irrotational
and satisfy the equation of continuity (mass-conservation).

Fig. 20. Coordinates for Prandtl-Meyer expansion fan.

Inthe present problem, considerable geometrical insight

can be obtained by using special coordinates appropriate to a
one-parameter family of straight lines. These special coordi
nates are: the angle 0 made with the x-axis by the lines,
and the directed distance h from a fixed curve cutting the
lines orthogonally, as in Fig. 20. If we recall that the lines
represent "in general" the tangents of a suitable plane curve
T, we will see immediately: the loci = const, are the
(i)

given straight lines, (ii) the lines = const, are the orthogonal
h

family of "evolutes" of (iii) ds2 = dh2 rW, where

r,

= the radius of curvature of the evolute, and

+

is

s(6)
h
r

denotes arc-length along r.

141

'.
 GROUP THEORY AND FLUID MECHANICS [CH. IV

In terms of these intrinsic geometrical coordinates, we can

easily express the two conditions of irrotationality and mass
conservation.
By definition, irrotationality is the condition that the circu
lation around any closed curve y vanishes. If a denotes the
angle from the straight line 6 = 0O to the vector velocity
u(0o), of magnitude q, the circulation around y is

J"Sukdxk = ^g(cos adh + sin ardd).

By Green's Theorem, this vanishes identically if and only if

— (q cos a) = (rq sin a).

^
Since r = h + s(0), dr/dh — 1; while by PI, q and a are func
tions of d alone. Hence irrotationality is equivalent to

(53) -j- (q cos a) = q sin a.

ad

Letting g(6) = q cos a, and h(6) = q sin a, this is g' = A,

generalizing (7) of Section 3.
To express the condition of mass conservation, observe that
the rate of mass flux outward through a curve y is

h pq(— sin adh + cos ardd).

By Green's Theorem, this vanishes identically if and only if

(54)
— (
— pq sin a) = -r (rpq cos a) = pq cos a.

In the notation just introduced, this reduces to the equation

(—ph)' = pg, where the accent denotes differentiation with
respect to 6. Substituting h = g' and simplifying, this is
p'g' + P9" + P9 = 0,
(g" + g) + (p'/p)g' = 0, which is
or
(9) of Section 3. Hence all flows satisfying Condition P1
can be obtained from Prandtl-Meyer flows, by replacing the
radii from a fixed corner by the tangents to a fixed curve T,
and using the same vector velocity at corresponding points.16

" This
result is due to F. H. Lees, Proc. Camb. Phil. Soc., 22 (1924),
350-362; see also W. Tollmien [69]; and [60], pp. 273-278.
142
§14] HODOGRAPH METHOD

The preceding illustration is a special case of the more gen

eral "inverse problem" of finding all flows with one-dimen

is,
sional hodographs— that all flows whose velocity vectors
are confined to lie on single curve.11

a
These investigations are all very interesting, and has

it
been freely conjectured that they are closely connected with
group-theoretic ideas. However, there as yet no con

is
clusive evidence to support this conjecture.
What does seem to be certain the following. Most of

is
the special coordinate systems which lead to useful sepa
rations of variables or other formal simplifications in fluid
mechanics have been suggested by group-theoretic ideas
relating to internal symmetries of the partial differential equa
tions involved. Formal postulation of these coordinate
systems then yields indirectly, through the "inverse method,"
flows which are usually closely related to the "self-symmetric"
flows directly suggested by group-theoretic ideas. The postu
lation of formal properties which are not related to groups
will usually yield only uninteresting flows, or no flows at all.

14• Hodograph Method and Groups

Another method of simplifying the equations of flow the

is
so-called "hodograph method." shall now exhibit two
I

connections between this method and group theory.

We have already seen (Chapter III, Eq. (35') that the
irrotational plane flows of compressible non-viscous fluid
a

correspond one-one to the velocity potentials which

U

satisfy the non-linear partial differential equation

V*U = 2UxUvUxy + LyUvUvv}

+

(55) -2{ UxUtUxx

.

Here subscripts represent differentiations with respect to the

the local velocity of sound — equal
-
variable involved, and
c2
is

to }4(y — the constant "escape veloc

<72)' where a
is

1)(a2
"
ity, in the case of a polytropic equation of state = kp'.
p

17
For the incompressible, viscous case, see W. M tiller, Einfuhrung in
die Theorie der zdhen FlOssigheiten, Leipzig, 1932; also Zeits. ang. Math.
Mech., I3 (1938), 395-408. For the compressible, non-viscous case,
see [69]; also unpublished papers of Giese and G. F. Temple.
J.

143
 GROUP THEORY AND FLUID MECHANICS [CH. IV

It is well-known18 that (55) is equivalent to either of the

linear partial differential equations

q*VM + q (l + £j V, + - V„ =

jj)
(l

0,
(56.1)

or

jV„ + - +, - = 0.

+
(56.2)

f|)
f|)
q

(l
(l
Here the stream function; qeie the complex velocity
F
is

is
vector, so that Ux = cos and Uv = sin single-

c2
6;

is
a
q

q
0
valued function of by the Bernoulli equation; and =

<t>
q,
U —
xUs — yUv the dependent variable of the Legendre
is

contact transformation giving (56.2).

Now not difficult matter to calculate (56.1)-(56.2)
is
it

from (55) and the preceding definitions, but quite obscure

is
it
why the hodograph variables leading to these linear equations
should have been tried. It may be that one reason was the
success of the hodograph method in free streamline problems
(as in Chapter II, Section 2). shall now give an alternative
I

motivation in terms of three group-theoretic considerations.

The first group-theoretic consideration the invariance of
is

the laws of non-viscous fluid dynamics under the group of

Langevin transformations

(5) x->Xx, u->u, U->\U, V-+W, 0->X0.

We do not have invariance under — x, u — \u, except in the
>

>
x

incompressible case (cf. Chapter III, Section 18). It follows

that partial differential equation expressing these laws must
a

be non-homogeneous (and hence non-linear) the x, are taken

if
U,
V,

as the independent variables, and or as dependent

<£

variables^ — while will be homogeneous the w, are taken

it

if
U,
V,

<t>

as independent variables1' and or as the dependent

variables.
18
S. Bergmann, Method in the Theory
The Hodograph Com
of

See
pressible Fluid, Brown University, 1941; or H. Seifert, "Die hyper-
geometrische Differentialgleichungen der Gasdynamik," Math. Annalen,
120 (1947), 75-126.
19
This possible locally except in the case, mentioned in Section 12,
is

of a one-dimensional hodograph. It not generally possible in the large,

is

and this leads to mathematical difficulties in applying the hodograph

method, which have not yet been surmounted.
144
§14] HODOGRAPH METHOD

I confess it is not yet clear to me a priori why this homoge

neous equation should be linear when V and 4> are taken as
the dependent variables, and I
have not determined what
kind of equations would arise if the equally plausible U was
used as dependent variable.
A second relevant group-theoretic consideration is the
obvious invariance of the laws of fluid motion under the group
of rotations 0 — > 6 + a, when q, U, V, This

<t>
are held fixed.
implies that in (56.1)-(56.2), will only appear in differential

6
operators, and not in coefficients. Hence we have a group-
theoretic motivation for the use of as independent vari

q,
6
ables20 in place of u = Ux and = Uv. It makes the coeffi

v
cients of our linear homogeneous partial differential equations
functions of one of the two independent variables only.
An analogous third group-theoretic observation the obvious

is
invariance of the laws of fluid motion under the group of
This
U,

translations x — <t>
V,
+ a, when u, are held fixed.
>
x

has been commented on in Chapter III, and too equiva

it

is
lent to the occurrence of x and in (42) as differential operators
y

alone, and not in coefficients.

This same group-theoretic observation seems related to the
classic procedure for solving problems in free streamline theory
(Chapter II) by the hodograph method. Indeed, let W
f,
z,

be the complex position, complex conjugate velocity, and

complex potential respectively. Any physical condition
<?(z,f, W) = involving these three quantities will be invariant
0

under the complex translation group — z0; hence

+

it
>
z

must be equivalent to relation of the form 6(f W) = or,

0,
a

solving, W = F(f to relations of this form that conf ormal

it
is
)
;

transformations lead. Again, any first-order ordinary differ

ential equation H(z,t;,d^/dz) = to be invariant under the
0,

translation group —
Zo, must be [61] integrable by
+
>

a
z

quadrature — and to such quadrature, of the special

it
is

form = /(^"(f)df/f) that the hodograph method leads.

z

Lest the connection between quadratures and invariance

under translations seem too abstruse and mathematical,
I

10
The use of io = In instead of suggested by complex variable
is
g

theory, so as to make w i8 the complex logarithm of u

+

*t>; see
+

Chapter II.
145
 GROUP THEORY AND FLUID MECHANICS [CH. IV

shall now give an example21 where this connection is utilized

repeatedly.

15. Nonlinear Theory of Inertia! Plane Motion

Let missile be moving in a plane, so rapidly that only

a
"inertial" forces (Chapter III, Section 12) need be con
sidered — gravity and friction are thus considered negligible.
We let x = qi and y = q2 denote the coordinates of the
missle, and = q3 the angle between the x-axis and an axis
\f/

fixed in the missile. We assume22 that the instantaneous

coordinates of position and rate of change of position of the
missile determine its future trajectory through force-reactions
and Newton's laws, so that (subject to the usual limitations
on differentiability etc.)

= Fi(quqt,q3Ai,4t,b) = F<(q;q).
q\

(57)

This evidently a sixth order system of ordinary differential

is

equations, which are in general nonlinear.

shall now show how can be reduced to a second order
it
I

system and four quadratures, by use of group theory. The

method indicated generalizes in principle the usual method of
"ignorable coordinates,"23 from Lagrangian to non-Lagrangian
dynamical systems. After describing the reductions, shall
sketch (in Secton 16) the further extension to general ordinary I
differential equations invariant under group.
a

First, the system (57) may be expected to be invariant

under translations x — x + xo, — + yo of space-coordi
>

>
y

nates.24 This has very simple mathematical interpretation:

a

permits us to replace (57) by the fourth order system

it

21
For rigorous modern mathematical discussion of this classic rela
a

tion, in the case of (ordinary) homogeneous linear differential equations,

see E. R. Kolchin, "Algebraic matric groups and the Picard-Vessiot
theory, etc.," Annals Math., 49 (1948), 1-42. For the non-linear case,
of

see Dickson [61].

22
This hypothesis satisfied by all the forces usually considered
is

including those of drag, virtual mass, etc.

23 As
described for example in [70], p. 54. With non-zero Cu, the
system discussed here non-Lagrangian.
is

24 In
practice, this means that such factors as the variation of density
with altitude can be neglected.
146
§15] INERTIAL SCALING

= F,(g,;4), = F,(g,;4),
(58)
^ ^
F,(9"q),
-dJ-q3' -dt-
and the two quadratures

(58') qi = J^eft, q2 = iqt/dt.

Thus two-parameter group permits us to reduce the order

a
of our system by two, replacing integrations by quadratures.
Second, the system (57) is presumably "isotropic", that

is,
invariant under rotations of coordinates. To express this fact
analytically convenient to use, as new variables, the
is
it

scalar velocity = (i2 + J/2)^, and the inclination of the

6
v

trajectory to the x-axis. This makes x = cos and =

y
6
v
sin Clearly and the "pitch angle" = —

<t>
are
B.

6
v

i>
invariant under rotations; hence (58) equivalent*6 (if
is
isotropic) to system
a

the second of which reduces to the quadrature = f&dt.

6

Finally, the hypothesis that all forces are "inertial" means

that they are proportional to velocity squared, i.e. that space
trajectories are invariant under the group of changes in the
time scale. But clearly = fvdt are invariant
<t>

and distance
s

under this group. Hence we replace by the independent

if

variable we will have

s,

=
a-^MT.
%

<M>
%-<, "<*•*'>.

(For example, mvdv/ds the tangent force-component; this

is

v2 times the force H*(v/v,<t>,<j>/v) = H*(l,<l>,<t>') which would

is

act all velocities were reduced in the ratio v.l). In sum

if

mary, (57) is equivalent to the second order system (60), plus

26
Strictly speaking, as long as does not vanish.
v

147
 GROUP THEORY AND FLUID MECHANICS [CH. IV

the five quadratures (58') and

(60') v = voe!"*"',
t=f^, 6 = f 6dt.
16. Generalization

The preceding reduction can be considerably general

ized. Let 2 be any nth order system of ordinary differential
equations:

dr•
= =
F&u
2jl'

■
■ • • •

1,
(61) ,xn) n].

[i

,
Suppose invariant under
group of transformations
is
2

r
a
— y(x) of (xi, ,xn)-space, meaning by this that
• • •

if
>
x

x(<) satisfies (61), then so do its transforms y(x(t)), for all

7«I\ We shall show that this information of real assistance

is
in the explicit integration of (61).
In the case one-parameter group, the procedure
is

is
r
a

easy. In this case, except near singular points, locally

is
T
equivalent26 by change of coordinates to the group of trans
a

lations yi— j/i + a; j/j, • •

yn constant. Under these
•
>

coordinates, (61) becomes dyt/dt = (7,(2/1, •

,yn)• Since
•

subtraction of constant from j/i does not alter any dyi/dt, we

a

see that the Gi are really independent of j/i, giving

= Gi(ylt ■ ■ ■ = •
^

1,

,yn)
•
•

(62) n].
[i

Hence we reduce the integration of (61) to the integration of an

-
1)"' order system dyj/dt
= GX2/2, ■ = •
(n

[j

2,

•
•
•
•

,2/n)
,

n], and the quadrature yt = fGi(y2(t),

■
■
■

,yn(t))dt.
More generally, let be any r-parameter solvable Lie group
T

of transformations of (xi, . . ,xn)-space, which leaves (61)

.

invariant. Then, almost by definition, possesses a series

r

of local27 Lie subgroups Si < S2 < Sr =

T,

■ •
such
■

<

*•See L. P. Eisenhart, Continuous Groups Transformations, Prince

of

ton, 1933, p. 35. In general, shall not give detailed references for the
I

results from Lie theory assumed.

17
The concept of a local Lie subgroup is given in [59]. The older books
are somewhat vague on this point.
148
§16] GENERALIZATION

that (i) jS,-i is normal in Si, (ii) Si is generated by <S,-i and

a one-parameter subgroup I\.
Locally, let us assume that the subsets of transitivity of
<S,-i are the /i-dimensional subspaces yh+i, . . . , yn constant,
i.e. the parallels to the (yi, . . . ,y*)-plane for some h. Sup
pose further that by consideration of the invariance of (61)
under <S,-i, we have been able to reduce the integration of
(61) to the integration of

= GAVh+i, ■
■
- +
3jS

■ • • •

\j
(62')

1,

n]
,Vn)

,
and quadratures. shall show that the corresponding result
I

holds for Su
There are two cases. If the subsets of transitivity of

<S<
are
A-dimensional, the result trivial. Otherwise, since £,-i a

is
is

normal (i.e. invariant) subgroup of

S,
the sets of transitivity28
of Si-i are permuted non-trivially by IV By proper choice
of coordinates, we can suppose that I\ effects the translations
— a; 2/*+2, ■ • • Vn constant. Hence, as with (62),
+

j/m•i
>

2/a+i
,

we can reduce the integration of (62') to the integration of

= G,W, =
*g

• • •
+ • • •
2,

n],
\j

(62") ,yn)
h

and the quadrature yn+i = $Gh+i(yK+2(t), ■ This

■

,yn(t))dt.
•

completes the proof by induction of the following

Theorem. Let an nth order system ordinary differential
of

equations invariant under solvable Lie group having

2

be
a

m-dimensional sets transitivity. Then the integration can

of

of
2

an —
be reduced to the integration m)th order system, and
of

(n

28
quadratures.
In Section 15, Ti was the group — x + r2 was
a, —>
y
>
x

+ T3 was 6— + a, x—>x cos a — sin a, — x sin a

b,

>
y

y
>
6

-f- cos a, and I\ was —> t/\, x — x, —> \v, etc. Hence the
>
y

v
t

By definition, "set of transitivity" of Si-i

is,

18
for some point y,
a

the set of all <r(y) Since ytTi and o«S,•-i implies y-1aytSi-l,
y

[ffe<Si-i].
the set of all •y(<r(y)) = (orHy), equals the set of all a(y(j)), and hence
also a set of transitivity of (S>-i. See A. Speiser, Gruppentheorie, 2nd ed.,
p. 117.
149
 GROUP THEORY AND FLUID MECHANICS [CH. IV

preceding theorem, which appears to be new, is a direct

generalization of Section 15.

17. Summary

We have seen that group theory often makes possible

reductions in the number of independent variables involved
in partial differential equations, directly through the method
of search for symmetric solutions and the method of "sepa
ration of the time variable," and indirectly through "inverse
methods." Moreover the method of search for symmetric
solutions can be relied upon in general to give solutions
locally (Section 9).
Even after the number of independent variables is reduced
to one, so that no further reduction by the preceding methods
is possible, the resulting system of ordinary differential equations
can often be most easily integrated by use of group-theoretic
considerations.
The preceding methods apply to non-analytic as well as to
analytic, and to non-linear as well as to linear differential equa
tions; thus they are free from the limitations of the usual
methods of expansion in series or integrals. Hence group
theory certainly plays a fundamental role in solving the differ
ential equations of fluid mechanics.
The role of group theory in fluid mechanics
is,

moreover,
not restricted to its usefulness in solving differential equa
tions. We have already seen that the theories of modeling
and of "dimensional analysis" can be regarded as special
cases of group theory. In my next lecture, shall try to
I

show that group theory also underlies the classical force

equations for rigid body in perfect (i.e. incompressible
a

non-viscous) fluid.
hope that, in the future, the debt of mechanics to the
I

concepts of group theory will be more explicitly recognized.

150
 CHAPTER V

Virtual Mass and Groups

1. Introduction: Virtual Mass of Sphere

Qualitatively, the idea of virtual mass is a familiar one.

For example, let a light paddle be dipped into still water
and then suddenly given a rapid acceleration broadside. It
is a matter of common experience that the apparent inertia (i.e.
resistance to acceleration) of the paddle is greatly increased
by the water around it. This increased inertia is what is
called the "virtual mass" of the paddle, the difference between
the real mass and the virtual mass being called the "induced
mass."
The idea of virtual mass was first given an exact mathe
matical interpretation by Green and Stokes, a little over a
century ago.1 The essence of their discussion went somewhat
as follows.
Consider a sphere of mass m and radius a moving with
speed v through an incompressible non-viscous fluid of density
p (throughout the whole of the present chapter, we shall con
sider only Dirichlet flows of such a "perfect fluid "). Without
loss of generality, we may choose the axis of spherical coordi
nates as the direction of motion. Relative to the fluid at
infinity, the velocity potential U is well-known to be the
dipole potential given in spherical coordinates by

,,»
M
,,
U =
— a3v
—
— 2?cos 6

In fact, one easily verifies that the normal derivative d U/dr =

v cos 8 gives the normal component of motion of the surface
of the sphere (Chapter I). At a general point of the fluid,
1
G. Green, Mathematical Papers, p. 315 (1833); G. Stokes, Mathe
matical and Physical Papers, Vol. 1, p. 17, (1843). See [9], pp. 123-124,
for a bibliographical discussion. The physical idea of virtual mass (also
called "added" or "carried" mass), was expressed in a very crude way in
1786, by the Comte DuBuat (see L. G. Dubuat, Principles d'hydraulique,
1816 ed., Vol. 2, p. 222).
151

r
 VIRTUAL MASS AND GROUPS [CH. V

the radial and angular components of velocity are

a3v cos 6 dU ah sin 6

Hence the total kinetic energy of the fluid is

T = I I I xAp(n\ + «,2)r2 sin d drd6d<t>

'd0

W-il' [— cos 0 — cos' 6]%

l
L3r3J0
irpv2a3 1 27rpa3 , 1 , ,
3 2 3 2

This gives the following classical [Link] kinetic energy

of the fluid is equal to that which would be possessed by a
particle moving with the same speed as the sphere, and whose
mass ml is equal to half the mass of displaced fluid.
Moreover it is evident that, in a non-viscous fluid, rotation
of the sphere has no effect on the surrounding fluid; hence
the moment of inertia of the sphere should be unaffected.
This suggests that (neglecting gravity) a sphere in such a fluid
should be dynamically equivalent to a heavier sphere in
vacuo, whose "virtual mass" m* = m + m' is the mass m
of the sphere plus an "induced mass" m' equal to half the
mass of displaced water, but whose moment of inertia 2ma2/5
is unchanged. This will be proved rigorously in Section 15,
where it will be shown that the entire dynamic action of
any Dirichlet flow can be deduced from its kinetic energy.

2. Physical Validity of Conclusion

Although it is not to that the physical facts

be supposed
are as simple in real fluids as those predicted by the preceding
theoretical argument, there are various observable experi
152
§2] PHYSICAL VALIDITY

mental phenomena for which "virtual mass" gives the best

known explanation.
Thus, as already stated, an object at rest in a dense fluid
does experience an extra resistance to acceleration. Gen
erally speaking, resistance to acceleration from rest agrees with
the mathematical theory.2 Various special cases should be
mentioned.
One classical application is to the initial acceleration
experienced by a spherical hydrogen balloon, suddenly released
from a mooring. Suppose the mass of the balloon is J^o that
of the displaced air, if suddenly released. A person unaware
of virtual mass would probably make the following erroneous
The hydrostatic buoyancy by Archimedes'

is,
computation.
Law, lOg times the mass of the balloon; hence (he would
compute) the initial acceleration should be lOg. In the case
of spherical hydrogen balloon immersed in water, an acceler
a

ation of at least 1000<7 would be erroneously expected .

However, using the theory of virtual mass, easy to

it
is
predict the correct initial acceleration. The virtual mass m*
of the balloon = that of the displaced air; hence
%
is .5
is

+
.1

the actual acceleration In water, would be about 2g.

it

5gr/3.
more subtle application arises weightless sphere
A

is
if
a

immersed in liquid, and the whole liquid suddenly given

is
a

an acceleration a. What the acceleration a* of the sphere,

is

relative to an outside observer? This problem may be

treated as follows. The acceleration a equivalent, for an
is

observer in the fluid, to an apparent gravitational field of

strength a. Hence, by the reasoning just given, the initial
acceleration a* — of the sphere relative to an observer in the
a

fluid, should satisfy a* — = 2a, or a* = 3a.

a

This prediction was confirmed by T. E. Caywood and the

author [19], where other references are given; proved essen
it

tial in interpreting experimental data on fluid flow patterns.

Surprisingly few dynamical experiments have been performed to
'

test this hypothesis; see however P. Hirsch, Zeits. ang. Math. Mech.,
3

(1923), 93-107, and Thesis, Goettingen, 1918, Aer. Res. Council Reps. 590,
612, 1166; G. Cook, Phil. Mag., 39 (1920), 350-352. There also support
is

ing photographic evidence. See Blasius, Zeits. Math. u. Physik, 36 (1908),

20-37; Tollmien, Handbuch Exp. Physik, Vol. Part (1931), 272-279;
4,

S. Goldstein and L. Rosenhead, Proc. Camb. Phil. Soc., 56 (1938), 20-37.

153
 VIRTUAL MASS AND GROUPS [CH. V

3. Induced Mass Tensor

In the case of the sphere, induced mass can be treated as a

scalar; but for most bodies, we must treat induced mass as a
matrix or "tensor."
The most convenient mathematical basis for such a treat
ment is furnished by computations for the total kinetic energy
of the fluid involved. We consider Dirichlet motions of a
perfect fluid bounded internally and/or externally by a rigid
body ("solid") 2, and constrained to move with 2. Let
U1, U2, U3 be the velocity potentials induced by translations
of 2 from a given reference position parallel to the three
coordinate axes; and let U*, Us, U1 be the velocity potentials
for rotation about these axes, with unit angular velocity.
Then if the solid has instantaneous velocity-components
<ji, . . . qt with respect to the six degrees of freedom of
motion just mentioned, the velocity-potential U in the fluid
satisfies U = qkUk, by the principle of superposition for a
perfect fluid. Here repeated indices denote summation. It
follows that the total kinetic energy T of the fluid satisfies

(2) 2T = pffFUVUdR = qhqkf!fpVUhVUkdR = qhqkThk,

where dR = dxidx^dx^ differential of volume. The

is the
symmetric matrix \\Thk\\ is called the "induced mass tensor";
its coefficients depend on the choice of axes in the body; we
shall discuss this dependence further in Section 5.
The coefficients Thk have a simple dynamical interpretation.
By Green's Second Identity ([7], p. 212), we have

f
fu»

(3) Thk =
Pf fvUWdR
=
Pf (jPj dS.
But now dUk/dn equal (Chapter (8)) to the normal
I,
is

velocity-component of under unit velocity in the qk direc

2

tion. shall introduce, for use now and later, the convenient
I

notation dSk = (dUk/dn)dS, so that

(4) dSi = dx2dx3, dS2 = dx3dxi, dS3 = dxidxt,
dSi = x2dS3
—
xadS2, dSs = x3dSi — xidS3,
—
dSt = xidS2 x2dSi.
154
§4] CANONICAL FORM

It is evident from the definition itself that

(3') Thk = pttUWS* = pffU*dSh.

It is also evident that if p denotes the scalar pressure, then
ffpdSi, ffpdS2, ffpdS3 are the components of force exerted
by Z on the fluid, while ffpdSt, ffpdSs, ffpdSt are the corre
sponding components of moment.
Now consider unit acceleration from rest in the ^A-direction.
It is easy to calculate that, initially, U = 0 and dU/dt = Uh:
U(x;t) differs from tUh(x) by infinitesimals of the second
order in t. We now apply Bernoulli's equation for pressure
in an accelerated fluid,

(3*) P + \ pVUVU + ^Jp = const.,

neglecting hydrostatic buoyancy forces (Chapter III, Section

12). This shows that the initial hydrodynamic pressure ph is
everywhere equal to p times the "acceleration potential"
Uh. Hence, substituting in (3'), Thk = ffphdSk, and Thlc is
the k-component of force for unit acceleration from rest in the
qh-direction.
Incidentally, because of the symmetry in h,k of Thk evident
from the kinetic energy definition (2), we obtain immediately
the following Reciprocity Principle. •The A>component of
force under unit ^-acceleration is equal to the ^-component
of force under unit fc-acceleration.
The preceding formulas refer to the induced or "carried"
mass. The total inertia or virtual mass, defined as the internal
(rigid body) inertia of 2 plus the induced mass of the external
fluid, is clearly another symmetric tensor (matrix), with just
the same properties. We shall now discuss the algebraic sim
plification of the tensor of total virtual inertia.

4• Reduction to Canonical Form

As we have already seen (Section 1), the sphere is a very

simple case. If axes are taken through the center of the
sphere, then Tn = T^ = TS3 = m*, Tit = Tii = Tet =
2ma2/5, and all Ta [i j* j] off the main diagonal are zero.
Another familiar special case concerns a rigid body in vacuo.
155
 VIRTUAL MASS AND GROUPS [CH. V

Ifthe origin is taken at the center of gravity, then all Tn [i j± j

and i or j
= 1, 2, 3] vanish, while Tn = Tn = T33 = m, the
mass of the body. Again, by choosing Cartesian axes as the
principal axes of inertia, we can make all T„ [i j* and j
i, =j 4, 5, 6] vanish. Hence we can characterize the inertia
tensor by four scalars Tn, Tu, Tii, Tu, which can be reduced
to two by a change of length and time units. Further, Tu +
Tu ^ TM and cyclically; the case of an ellipsoid is perfectly
general.
In the preceding two cases, the matrix ||7\,|| could be
reduced to diagonal form by proper choice of axes. It is
interesting to see how far one can simplify the matrix of inertial
coefficients 7\„ by appropriate choice of Cartesian axes and
a "center of virtual inertia" in the general case of a fluid of
positive density. This is a simple exercise in the theory of
quadratic forms.
The example of the paddle (Section 1) shows that the "vir
tual inertia" for translation can be different in different direc
tions; sidewise acceleration is much easier than broadside
acceleration. However, since any quadratic form is equiva
lent3 to a diagonal form under the rotation group, we can
always rotate axes so as to get Tu = T23 = Tu = 0, and
"virtual inertias" of Tn, Tn, T33 in "principal directions"
of translation.
Except in the degenerate case of a weightless plane lamina,
where one of the Tu [i — 1, 2, 3] vanishes, we can achieve a
further simplification by proper choice of the origin as a
"virtual center of inertia." Let u>i, m>s, w3 denote rotation
at one radian per second about one set of axes parallel to the
principal directions of translation; let X, Y, Z denote the
translations at unit speed in the principal directions; and let
w'i, w'2, w3 denote rotations about axes translated through the
vector (x,y,z). Then
—
w[ = wi + yZ zY,
= w2 + zX - xZ,
(5) w't
= w3 + xY - yX.
w'3

The algebraic theorems assumed here are proved for example in [37],
3

Chapter IX, Theorems 19, 20. For the results see [9], Section 126; the
general case was treated by Clebsch, Math. Annalen, (1870), 238.
3

156
§4] CANONICAL FORM

The substitution of w[ for wi does not change the interaction

energy Tu between X and Wi, since there is none between
X and Y or Z. On the other hand, it increases 7\6 = Tn
by zTn, increases Tie = Tei by xTn, decreases T2t by zTn,
and so on.
Hence we can make Tu = T24 by proper choice of z, and
similarly Tie = ^34 and T26 = T3i. Thus we can reduce the
matrix of inertial coefficients to the simplified form of Fig. 21.

Translation Rotation
a 0 0 P 1 n
o p 0 { <r r
0 0 7 ij f r
p * i? r44 i45 ^46

* ' f 7*45 J 55 Tse

v r r ?46 r 66 Tee
Fig. 21.

In this canonical form, we have fifteen arbitrary constants,

which can be reduced to thirteen by a change of length and
That
is,

time unit. the general case involves thirteen dimen-

sionless ratios, and two choices of unit.
If body has three perpendicular planes of symmetry, like
a

an ellipsoid, we can take these as coordinate planes. Reflec

Z,

tion in (say) the (x,z/)-plane leaves wi, w2 fixed but reverses

the signs of X,
Y,

w3; also leaves kinetic energy unchanged.

it

Hence T3i — —
Tu = and similarly
0,

T32 = T3e — Tu — * = ^46 = Tu — Tm = T$e =

0

42

Repeating the argument for the other coordinate planes, we

see that the matrix of inertial coefficients can be diagonalized.
Thus we have six arbitrary constants, and four dimensionless
ratios. (There are two in the case of rigid body in vacuo.)
a

Another interesting case that of symmetry in three per

is

pendicular axes, though not in the planes through them. This

typified by the helix x = =
g

cos sin
is

\r\
z,

z,

1,
y
r

Z, r

2ir. Symmetry in the z-axis leaves w3 fixed but

^
\z\

reverses the signs of X,

Y,

wi, w2. Hence as above,

Tl3 = = ^43 = Tm = Tie = = = —

0.

Tie
T

23 * 26 * 46

157
 VIRTUAL MASS AND GROUPS [CH. V

Repeating the argument for the other coordinate planes, we

see that all inertial coefficients off the principal diagonal
vanish, except the "helical products of inertia" p, a, r. We
thus have nine coefficients of inertia,4 which we can reduce
after changes of length and time scales to seven essentially
independent parameters.

5. Induced Mass and Momentum

In Section 3, we interpreted the coefficients of induced mass5

in terms of kinetic energy and the force required for acceler
ation from rest, respectively. It is possible also to assign
simple momentum interpretations to these coefficients.
The convenience of the definition (3) of the Thk as kinetic
energy integrals was, that these integrals converge at infinity.
To see this, we recall that (Chapter I, Section 6) VU = 0(r-3)
in space, whence ///(VC/Vf/)di2 = 0(r-•r2dr) is finite. The
kinetic energy integral also converges at infinity in the case
of plane Dirichlet flows; in this case fff(VUVU)dxdy =
0(r-Vdr) is finite.
We shall now interpret the Thk as momentum integrals.
It has been noted by various authors6 that the momentum
integrals diverge in the usual sense. Hence to interpret the
Thk in terms of momentum, some care must be taken. Now
for the details.
The Thk of (3') are integrals over the boundary S of 2,
which can be written in a new notation as

(3") Thk = p
f fsUdSK,
U = Uh being harmonic and dSK = Y . KdSi expressing the
differential of flux of a vector field K = (K i(x), K2(x), K3(x))
* The only references which I have found on the reductions of Section
4 are Kelvin's papers [71], esp. pp. 384-390, where "helicoidal symmetry"
is discussed.
* The reductions to canonical form described in Section 4 apply equally

to the induced mass and to the virtual mass tensor — naturally, with
different characteristic axes in general. I treated the case of virtual
mass in Section 4, because the familiar special case of a rigid body in
vacuo is so suggestive.
* [9], p. 161; H. Tollmien, Zeits. ang. Math. Mech., 18
(1938), 154.
158
§5] INDUCED MASS AND MOMENTUM

through S. Thus in the case k = 1 of translations parallel

to the Xi-axis, K = (1,0,0) ; in the case of rotations about the
Xi-axis, K = (0,^3, —x2), and so on. This convenient notation
defines a useful class of Stieltjes integrals over surfaces in
space, whenever ff\dSK\ is finite. Note that whenever K is
the velocity-field of a rigid motion, div K = 0. This con
dition, and the condition that U is a harmonic function
"regular" at infinity ([7]), p. 217) are all that we require for
the argument below.
We define a "JC-line" as an integral of dxt = Kidt. Thus
if K represents translation parallel to the xi-axis, the K-lines
are the parallels to this axis; if K corresponds to screw motion
about the xi-axis, the if-lines are helices about this axis, etc.
We then apply the Divergence Theorem to the vector field
UK, over regions R bounded by surfaces S + S' + S", where
S" consists of Jf-lines. Since dSK = 0 on the surface S" con
sisting of iC-lines, and since

div (UK) = K~+U

dxi
div K,
we get from (3"), using div K = 0,

<6) T»->inxK,%)a->iiuds-
Here the relation between the signs of the double and triple
integrals involved is correct, provided surface integrals are
taken into R. In particular, if R is an infinite region, if S'
is removed to infinity, and if the integral over R converges,
we get

We shall now consider various cases of (6)-(6') corresponding

to special induced mass coefficients.
If k = 1, then K = (1,0,0) = grad xi = Vxi in various
notations. We let S" be an infinitely long cylinder parallel
to the rr•i-axis and containing 2. Then since the momentum
integral
'

(6.1) Thl = Tlh=p

f [ UHx^tx, = p

159
f f f (j£) dR
 VIRTUAL MASS AND GROUPS [CH. V

converges at infinity, we get the following result. The induced

mass coefficients corresponding to translation along the xi-axis
are equal to the xi-momentum inside any infinite cylinder parallel
to the xi-axis and containing 2, under motion with unit h-velocity.
It is simplest to take R as the cylinder circumscribed about 2.
This result, as regards Tn, was obtained by Theodorsen [77].
Again, if k = 4, then K = (0,x3, — x2) = xtfxz — x$x3.
We let <S" be any solid of revolution containing 2 and having
the xi-axis as axis of symmetry. Then the boundary S + S"
of R consists of S and X-lines (circles), so that (6') reduces to

(6.2)

where 6 =tan-1 x3/x2. Hence Tht is the moment of momentum

of the fluid in R, about the a^-axis.
The other Thk are easily obtained from (6.1)-(6.2) by cyclic
permutations of axes. Also, the case of screw motion about
the £i-axis, K = aVxi + 0(x^x3 — x^xz) is easily obtained
by superposition of the two preceding formulas. We let
S" be a circular cylinder circumscribed about 2. Hence the
induced inertia is the screw momentum pff(adUh/dxi +
pdUh/d6)dR of the fluid inside S"; this is the same for all
circular cylinders parallel to the zi-axis which contain 2.

6. The Case of Rankine Bodies

The induced mass for translation can be expressed in still

another way. Apply Green's Second Identity to UVxi — xiVU,
recalling that V2t/ = V2xi = 0. Hence if S" is a large sphere
containing 2, and R is the region between 2 and S", we get
from (3'), writing Uh = U,

(7) Tu.„jjsu(%;)dS

160
§6] RANKINE BODIES

moreover on S", —d/dn = d/dr. The integrals over <S" can

easily be evaluated asymptotically, if we expand U in the
form

(8) -U =
l^ + 0(r->),
^ =
2^ + 0(0.
Since the area of S" is 4nr2, the terms 0(r~3) resp. 0(r-4) con
tribute negligibly to (7). By symmetry, the terms in m, m3
contribute nothing. To evaluate the remainder, we use

<t>,
spherical coordinates, writing xi = r sin dS = 27rd(sin <t>).
The integral over S" differs by 0(r-1) from

<t>)
sin2 sin2 <£}d(sin
+
{2

/w/2
0

= 2irp/xi[sin3 = 4*7>/ii.
<t>]\

Hence, passing to the limit as —>

« we get
r

(9) Tu = TH = 4tPMJ -p xy
(j£) dS,
ff
where the dipole moment of Uh in the xi-direction. We
/i*
is

note that the implication (7) — valid for any potential

is

(9)
>

(dU/dn)dS = 0.
U,

function regular at infinity, with

f
J

This generalizes result of Sir Geoffrey Taylor [76], who

a

considered the case = In this case (3U1/dn)dS dSi,

is
1.
h

and ffxidSi = ///diZ = vol (2), integrating by parts over

the interior of S. Hence,
2

(9') Tn = 4*pn\ - vol (2).

p

Taylor explains the presence of the term vol (2) in (9'), by

observing that fluid-filled hollow in will increase its
2
a

induced mass by times the volume of the hollow, without

p

changing the dipole moment at infinity of U1.

He also notes that in the case of Rankine body, =
n\
a

2x(1*)e*
the total dipole moment of sources for translation along its
is

axis, 4xp times the dipole moment of the source-and-sink

is

distribution defining
it,

minus the mass of fluid displaced.

The formula for the sphere, with = a3/2 as in Section
1,
fi{

(1), special case:

is
a

„ /4xa3\ 2irpa3
rn 2.pa3-p^j=-3-
m
,

161
 VIRTUAL MASS AND GROUPS [CH. V

The theoretical virtual mass, under the assumption of

Dirichlet flow, is also known for elliptic cylinders, ellipsoids
and spheroids under translation and rotation (see [9], pp. 85,
88, 153, 159, and St. Neumark, Zeits. ang. Math. Mech., 16
(1936), 117-120). The virtual mass for translation is also
known for tori, spherical bowls, sphere pairs, and lenses.7
Finally, Polya has shown that circular cylinders have the
least "average" virtual mass for translations perpendicular
to their axis, and studied the corresponding conjecture for
spheres.

7. More Applications of the Induced Mass Concept

The concept of induced mass has several applications

besides those to acceleration from rest. I shall now describe
some of these further applications. In general, the concept
must be applied in a quasi-empirical manner; the flows
involved do not approximate Dirichlet flows.
One application is to the correction for measured Cd of
bodies in wind-tunnels with diverging walls. Such walls have
been used to compensate for the pressure drop occurring with
parallel walls; they clearly produce a decelerating air stream,
and hence a pressure rise. But now, in Newtonian mechanics,
a deceleration of the frame of reference is equivalent to the
superposition of an apparent gravity field. Hence the effect
of the deceleration of the air stream with Dirichlet flows,
is,

the same as the body were accelerated in the direction of

if

deceleration, and acted on by "buoyancy" force correspond

a

ing to the pressure gradient producing the deceleration. After

little reflection, one sees that all forces are cancelled out
a

except for the "induced mass" force corresponding to the

For tori, see W. M. Hicks, Phil. Trans., 172 (1881), 609, and F. W.
7

Dyson, ibid., 184• (1892), 42. For spherical bowls, see A. B. Basset,
Proc. Lond. Math. Soc., 16 (1885), 286. For lenses, see [75], p. 279.
For sphere-pairs, see W. M. Hicks, Phil. Trans., 171 (1880), 455-492, and
A. B. Bassett, Proc. Lond. Math. Soc., 18 (1887), 369-377. For Polya's
results, see Proc. Nat. Acad. Sci., 33 (1947), 218-221. G. Szego, "On
the virtual mass of nearly spherical solids," Duke Jour. Math., 16 (1949),
209-223, has gven an interesting discussion of inequalities satisfied by
the coefficients of inertia. See also M. Schiffer and G. Szego, Trans.
Am. Math. Soc., 67 (1949), 130-205.
162
§7] THE INDUCED MASS CONCEPT

relative acceleration of the body ; hence a correction of roughly

this magnitude must be made.8
Again, especially for calculations involving wing "nutter,"
it is important to be able to estimate the aerodynamic forces
on an oscillating airfoil — or equivalently, those on an airfoil
in gusty air. These forces have been estimated theoretically
and measured in wind-tunnels, and the two estimates have
been found to agree9 within twenty per cent.
The effect of "induced mass" on the periods of vibration
of ships is also of practical importance. This has been esti
mated by various authors.10
It also seems necessary, in the theory of the stability of
liquid-filled chemical shell, to consider their virtual inertia.
In this case the fluid is bounded by external instead of internal
walls ; the case of ellipsoidal cavities is susceptible of complete
theoretical treatment, to the extent that viscosity and turbu
lence can be neglected.
Likewise, when a sphere or other missile enters water, it
may be supposed to lose more energy than what is required to
maintain its motion at constant speed, as it has to set the
water in motion. This leads one to define an "impact
induced mass," corresponding to this additional loss. Such
a loss has been measured.11

8 See M. Munk, N.A.C.A. Report 114 (1921); G. I. Taylor [77], esp.

pp. 260-283. Thus the correction for the pressure gradient dp/dx in a
wind-tunnel is (V + Tn/m)dp/dx, where V is the volume of the body,
and m is the mass of air displaced.
9
See W. Gracey, N.A.C.A. Report No. 707; W. Scruton, Reps, and
Mem. Aer. Res. Comm., No. 1931; W. T. Jones, ibid., Nos. 1946, 1947,
1953, 2026, 2117; W. Muller, "Zur Theorie der Krafte bei der beschleunig-
ten Bewegung eines Korpers in der reibungsloser Flussigkeit," Ing.
Archiv, 14 (1944), 332-350; L. I. Sedov, N.A.C.A. Tech. Memo. 1156
(1947), 1-62; Appl. Mech. Rev., 1 (1948), No. 309; Y. T. Yu, Jour. Appl.
Phys., 13 (1942), 66-69. A complete resume of the mathematical theory
and literature has been given by E. Reissner, Bull. Am. Math. Soc., 55
(1949), 825-850.
10
See F. Lewis, Amer. Soc. Nav. Arch., Nov. (1929); J. L. Taylor, Phil.
Mag., 9 (1930), 161-183; M. Haskind, Prikl. Mech., 10 (1946), 475-480.
"See W. Bauer, Annalen der Physik, 82 (1927), 1014-1016. The
theoretical status of this is quite vague, since the theoretical kinetic
energy of cavity motion is infinite. A theoretical basis for estimating
deceleration in the initial moments of impact has been given by Th.
von Karman, N.A.C.A. Tech. Note 321 (1929); see also [75].
163
 VIRTUAL MASS AND GROUPS [CH. V

Finally, though cavity flow is quite different from Dirichlet

flow,12 so that the discussion of Sections 1-7 certainly does
not apply, the measured cavity drag coefficient of a freely
moving sphere in water varies with its specific gravity in a
way which can most easily be explained by assuming that its
mass is increased by an "induced mass" m' equal to about
.4 times the mass of displaced water. If m is the mass of the
sphere, then m' = Am, whence if a denotes the deceleration,
we should have

D = (1 + Aa)ma instead of D = ma.

Hence the apparent drag ma = D/(l + Aa), and the apparent

drag coefficient should be 1/(1 + Aa) times that observed under
conditions of no deceleration. This relation is observed
experimentally. 12*

8. Inertial Lagrangian Systems

Consider a dynamical system consisting of a rigid body 2,

and a perfect fluid without free surfaces bounded internally
and/or externally by 2. A classical chapter of hydrodynamics
([9], Chapter VI) concerns the behavior of such a system.
It is now well-known that this model does not correspond to
physical reality, except in the case of acceleration from rest.
Nevertheless, it has great historical importance, and to this
day no better model of comparable simplicity is known.
I therefore feel justified in presenting a new mathematical
ultimately in terms of the ideas of group
it,

treatment of
theory, whose importance for other hydrodynamical questions
has already been shown (Chapters III, IV).
Clearly has six degrees of freedom of motion, which may
2

be described by coordinates g% Again, q(t)

if

. . g«.
.

given, then under very general conditions there one

is

is

and only one velocity potential (see [7]), pp. 213, 217, 311)

An interesting recent note of S. Luneau (Comptes Rendus, 227

12

(1948), 823-825 and 229 (1949), 927-928), shows that the virtual mass
of disc, accelerated from rest, can be very large. This presumably
is
a

due to the entrainment of a large "wake."

12* Unpublished experimental data of A. May indicate that the
"virtual
mass" term 0.4<r above should be 0.12<r.
164
§8] INERTIAL LAGRANGIAN SYSTEMS
U = (ftf/'Cq) which satisfies V2C7 = 0 and the values of
dU/dn given by the motion of 2. Hence the fluid kinetic
energy satisfies

(10) T
-Mfff P(VUVU)dR =
H £ r<,(q)M;

moreover the combined kinetic energy of fluid and solid

satisfies a similar equation with different coefficients. By
choosing Qi, ■ ■ ■ , qt near the reference position q = 0, as
in Section 3, we can make the r*y(0) equal to the coefficients
of induced (resp. virtual) mass already considered. But in
general, 7V,.(q) will vary with q unless translations alone are
considered.
Further, if S is just buoyant, gravity may be neglected;
this is also approximately true if the Froude number F = v2/gd
is very large.13 We shall consider this case, which involves
no potential energy.
In the classical literature ([73], p. 238; [78], p. 320; [90],
loc. cit.) it is usually assumed without proof that the system
just defined may be treated as a "Lagrangian system," so
that the "generalized forces" Q, satisfy

(ID
{11) Q.=±(*I\_W
^ dt\dqj dQi

I shall prove this rigorously in Section 15. For the present,

I note merely that it is plausible, in view of the idea that a
perfect (i.e. incompressible non-viscous) fluid moves under
"workless constraints." But it is certainly not obvious; thus
there are an infinite number of degrees of freedom for the
possible fluid configurations.
I also note that the "generalized external forces" Q, have
a very simple meaning in the present case. We have chosen
coordinates near 0 as in Section 3, so that qi, qz, g3 measure
translations along Cartesian axes through the center of gravity,
while #4, ?6, ?6 measure rotation about these axes. Hence,
at q = 0, Qi, Q2, Q3 are ordinary force components along the
axes, while Qt, Qi, Qt are moments about the axes. Under
II As in Chapter III,
Theorem 4, the effect of buoyancy is simply to
superpose a constant vertical force.
165
 VIRTUAL MASS AND GROUPS [CH. V

any transformation of coordinates, the variational expression

(11) transforms like a contravariant vector.
To sum up, then, we assume that our dynamical system is a
Lagrangian system without potential energy, so that the forces
are given by (10) and (11). Such a system may suggestively
be called an inertial Lagrangian system. By proper choice of
coordinates, we can furthermore make the T„• be the ordinary
"virtual mass" coefficients, and the Qi ordinary forces and
moments, at any particular instant.

9. Geometrical Interpretation

As is well known,14 the behavior of any inertial Lagrangian

system can be suggestively visualized in terms of the motion
of a particle of unit mass on a Riemannian variety V, in which
the differential ds of arc-length is

ds* = (tidqdq;.
(12)
^ Ta
For on V (assumed frictionless), the kinetic energy is clearly
given by (10).
It is also clear that the "natural" trajectories, which occur
in the absence of external forces, are geodesies. Specifically,
Q =0 in (11) if and only if fTdt is a minimum. This obvious
application of Euler's equations is the simplest case of the
Principle of Least Action, the variational formulation of
dynamics.
Thus if V = V2 is a frictionless surface x = x(qi,qz) in
ordinary space, we see that the force of constraint is per
pendicular to F2; hence in the absence of external forces the
normal to the particle path is normal to V2\ this condition
is well-known to characterize geodesies. More generally,
consider an arbitrary trajectory y on F2. The force-com
ponent of constraint, normal to V2, is usually neglected. The
remaining force is in the plane tangent to W It has two
components: the component s tangent to y, which can be
computed using T = K2** as s = Qi<ji + Q2q2, and the
14 The kind of geometrical interpretation discussed in Section 9 is
treated in great detail by J. L. Synge, "Geometry of Dynamics," Phil.
Trans., A226 (1926), 31-106.
166
§10] GROUP-THEORETIC INTERPRETATIONS

component normal to y in the plane tangent to V2, which is

t>2 = times the geodesic curvature.

s2
Corresponding formulas hold in any Riemann space V. In
particular, transforms like a (contravariant) vector, and

Q
its normal component times the vector geodesic curvature.

v2
is
Hence the dynamical problems of inertial Lagrangian systems
are equivalent to geometrical problems.

10. Group-theoretic Interpretations

shall now show that in the case of rigid body in an infinite

a
I

ideal fluid, we also have the following important group-

theoretic Relativity Principle. Relative to the body, all posi
tions seem equivalent. We can express this formally as
follows.
The different positions = a, c, • • • of the body in
b,
q

space correspond one-one to the different rigid motions

o, 7, • carrying the body from a fixed reference position
g,

to a, c, • • • Hence we can identify the points of the

b,
0

"configuration-space" of Section with the elements of

9,
V

the Euclidean group. Moreover a any particular rigid

is
if

motion, the position a<7 appears just the same to an observer

in position a, as a does to an observer at all Cartesian
0,

coordinate systems being equivalent. Hence the "group-

translation" a—> aa cannot affect the kinetic energy metric
(12).
Now let a vary; regarding as the group manifold of the
V V

Euclidean group, we see that has "simply transitive"16

a

group of "isometries" (i.e. of motions leaving the metric ds2

invariant). shall call such manifold, a Riemannian
a
I

group manifold; we can always regard the isometries as

left-translations.
Moreover another group-theoretic remark possible. By
is

a "steady motion" meant in dynamics motion which

is

is
a

independent of time, relative to body axes. Clearly, an

acceleration relative to steady motion increases Qi =
q

"
By this meant that, given <r, nV, there exists one and only one
it
is

a with aa = x. We are here assuming some slight acquaintance with

the left-translations of an abstract group.
167
 VIRTUAL MASS AND GROUPS [CH. V

d/dt(Tijqj) + 2(dThk/dqi)qhqk by Taq). Hence to obtain the

forces for an arbitrary motion, we can simply superpose those
for an acceleration q from rest, discussed in Sections 1-6, onto
those for steady motion. Thus if we can determine the forces
on a rigid body under steady motion in an "ideal" (i.e. incom
pressible non- viscous) fluid, we can determine the forces under
any motion. We therefore concentrate on this problem.
The only kinds of steady motion which are geometrically
possible for a rigid body in Euclidean space are well-known16
to be translation, rotation, and screw motion with a fixed
pitch — all at constant velocity.
By definition, if a(t) is a steady motion the displacement a
required to get from a(t) to a(t + h) is a function <r(h) of h
alone. Hence

a(0)a(h + h') = a(h + h') = a(h)a(h') = a(0)o(h)a(h').

Left-cancelling a(0), have a(h + h') = <r(h)cr(h') ; hence
we
o(h) is a one-parameter subgroup under canonical parameters
—
and a(t) is an image thereof under an isometry, namely, the
group translation a(t) —> a(0)a(t) = a(t).
Moreover since the total kinetic energy is constant in
steady motion, clearly s = v is constant in the associated
Riemannian variety V. Hence by Section 9, under steady
motion, the force vector Q is a constant v2 times the vector
geodesic curvature of the one-parameter subgroup a(h).
Hence the force on a rigid body under steady motion in an ideal
fluid is proportional to the vector geodesic curvature of the corre
sponding one-parameter subgroup of the Euclidean group V, for
a suitable "left-invariant" metric of V. Moreover this left-
invariant metric is determined at all points by the "inertial
coefficients" jT,,(0) already discussed in Sections 3-6.

11. Concepts from Lie Theory

Ishall now derive a formula for the geodesic curvature of

the one-parameter sub-groups of an arbitrary Riemannian
group manifold. This result is incidentally of interest for
" See for example J. S. Ames and F. D. Murnaghan, Theoretical
Mechanics, p. 87; or J. L. Synge and B. A. Griffith, Principles of Me
chanics, p. 282.
168
§11] CONCEPTS FROM LIE THEORY

the geometry of Lie groups — which illustrates once again the

essential unity of all mathematics.
It would be impossible for me to give enough of the theory
to Lie groups, in this short space, to make clear all details
of the derivation. However, I should at least like to give
enough so that you can form a clear idea of what the final
formula means, at least in the case of the Euclidean group.
If a rigid body is moving with unit velocity parallel to the
xi-axis, the velocity of any molecule of the body is (1,0,0).
Hence if F(xi,xz,x3) is any function defined over space, the
rate of change of the value of the function, as observed from
the molecule, is dF/dxi. The operator d/dxi defined in this
way is called the Lagrangian symbol expressing the infini
tesimal transformation associated with the rigid motion of
translation parallel to the xi-axis.
If the rigid body were rotating with unit angular velocity
(one radian per second) around the xi-axis, a molecule at
(xi,X2,X3) would have velocity (0, — x3,x2). Hence the rate
of change in F(xi,x2,x3), as observed from the molecule, would
be x2 dF/dx3 — x3 dF/dxt. The infinitesimal transformation
associated with rotation around the Xi-axis is therefore
expressed by the Lagrangian symbol (linear differential oper
—
ator) x2d/dxs x3d/dxt.
Associated with the six degrees of freedom of motion of a
rigid body, we therefore have the six infinitesimal transforma
tions expressed by

E - d
v x2d X3d
OX3 dx2

(13) E - d
E = — xid
dXi dx3

— ^—
xld x2d
E3 = •T , T?
JO'S
OX3 dx2 dxi

Or by the corresponding vector fields (velocity fields) (1,0,0),

(0,1,0), (0,0,1), (0, -*,,*,), (x3,0,-x1), (-x2,x,,0).
The effect of the velocity-field (infinitesimal transforma
tion) E„ operating for a time t, is denoted exp (<E<) ; thus
exp (2E4) would denote rotation about the xi-axis through
two radians. If t < 0, exp (<E<) is taken to denote the inverse
169
 VIRTUAL MASS AND GROUPS [CH. V

of the transformation exp ( — <E<). Hence we have, for all

real t, u, the identity.

exp (<E.) exp (wE.) = exp +

(14) ({t w}E<).

By canonical for (say) the Euclidean group, we

parameters
mean a parametric representation of small rigid motions by
vectors, such that the rigid motion associated with the vector
t = (<i • • • ,<«), is the finite transformation

(15) exp (tjti + • • •

+ UEt),

which expresses what happens of the velocity-field tjEi +

•
+ teEe operates for a unit time.
■ •

Finally, the Poisson bracket or commutator [E<, E3] of two infini

tesimal transformations E* and Ey is defined as the double
limit

(16) lim(,u-,0
f- exp (-<E.) exp (-«E,) exp (<E<) exp (mE,).

It is well known that this limit is the differential operator17

which can be easily computed. Thus with the Euclidean group,

we have for example

[El EJ = [E,, EJ = 0;
(18) [Elt Ed = -E„
[Ei,• Ed = -Ee,
From these identities, all other Poisson brackets of Ei, . . . ,
E« can also be computed by cyclic permutations of the sub
scripts and use of the obvious identity [E„ E,] = — [E,•, E,],
which implies in particular [E;, EJ = 0.
It
may be of interest to note that in the case of infinitesimal
rotations E4, E6, E6, [E,•, E,] is simply the outer or vector
product E, X E,. Again, if E< and E, (or equivalently,

17
1 use throughout the convention that EiE, means "First Ev, then
E,•." The reverse convention is used by many authors.
170
§12] ONE-PARAMETER SUBGROUPS

exp (tEi) and exp (mE,)) are permutable, so that E<Ej = E,E„
then [E<, E,] = 0, and conversely.
It will be noted that in (18), we have always

[E.•.E,] =
(19)
£tcVE*
for suitable constants cjjf. It is a fundamental theorem of
Lie, that relations like (19) hold for any finite-parameter
group. The constants

ckj
are called the structure constants
the group, and determine the group to within isomorphism.
hope that the explanations which have just given will
I

I
make the results derived below understandable, even though
the derivations cannot be understood except by those already
familiar with Lie Theory.

12. Curvature One-parameter Subgroups

of

Now let be an arbitrary r-parameter Riemannian group

G

G,
manifold, and let be any one-parameter subgroup of
C

generated by the infinitesimal transformation E.

One can always ([59], p. 118) introduce canonical parameters
into near the identity, with basis = Elf E2, •••,£»
G

E
a

of infinitesimal transformations, so that = (qt, ■ •

,qH)
if

•
q
G,

any sufficiently small vector element of

is

(20) = exp (q1El + ■ ■ ■

+ gnEn).
q

From this generalization of (15), one can derive the following

generalization of (14),
=
+

Xq M)q,
(X

(21) Mq
•

where • denotes the group product of

and in G.
r
s

C r

Now consider the geodesic curvature of at = relative to

0,
is q

the metric ds2 = Tijdqrfqj. By ection this proportional to

2

9,
S

^ n * (dT\ - dT IT - dThk
d

di d

. . .
1
x

Qi mk
(22)
dim) dq-r (TikQk)
-dqT

—-
2

_, _ dTik
qk— .
- -— dThk dTa -—,
dTU
1

= Tikqk =
.

.
.
+,

qh qhqk

=
ji

since =for = •
and qk = for all
k.
2,
•
•
1,

r,
0

0
&

This disguised Christoffel tensor (cf. p. 39); by

is

[70],
a

171

X~"
 VIRTUAL MASS AND GROUPS [CH. V

Section 10, the constant of proportionality is v2 = Tn, which

we can reduce to unity by appropriate choice of time scale.
Ultimately, it is Q and not the geodesic curvature which really
interests us, so that the constant of proportionality need not
trouble us.
We now compute the partial derivatives occurring in the
last formula of (22). To do this, we note that by definition,
the infinitesimal vector dq' issuing from exp (<E,) is equivalent
under left-translation to the infinitesimal vector dq issuing
from the identity 0, if and only if

(23) exp (iEi + dq') = {exp (<E,)} • dq.

But the Schur-Campbell-Hausdorff series18 expresses the

right hand side of (23) in the form

{exp (<£,)) dq = exp (ffi, + dq + Ht[Eit dq] + • •

■),
•
(24)

where the omitted terms are quadratic in Since dq and

t.
dq' are equivalent as regards their infinitesimal part, we infer

(24') dq = dq' - ^[E,•, dq'] + •

•
•

Now writing dq = dgiEi + • • • dg^En and dq' = dq'JLi

+

+
• • •
+ dq'nEn, we get by definition, as in (19),
[E,•, dq'] = dq'j&i, E,] = dq'^Eh.

Substituting back in the vector equation (24'), and equating

corresponding components, we obtain the fundamental relation
= - Y%t4 +
■
■
■

(25) dqh dq'h dq\

,

neglecting quadratic and higher terms in

t.

But by definition of Riemannian group manifold, ds2

is

invariant under left-translations. Hence by (23) and (25)

= dqhThk(0)dqk
dq'hThk(tE.i)dq'k
= (dqL - V2t4dq'i)nk(0)(dqk - Mttfdq'i)
= dq'hThk(0)dq'k - c^T^Wdg',},
+

V2t{cWiTHk(0)dq'k

18 This classical formula proved under extremely general conditions

is

as Theorem 14, on p. 92 of my paper "Analytical groups," Trans. Am.

Math. Soc., 43 (1938), 61-101.
172
§13] CLASSICAL APPLICATIONS

neglecting quadratic and higher terms in t. Equating coeffi

cients, by interchanging the "dummy indices" of summation
j, h and I, k in curly brackets, we get
rM«E.) =
r„(0) - H*{cf r*(0) + cj»r«(0)} + • • •

Now differentiating with respect to and writing both sum

t,
mation indices as j, we get the formulas

_l{c«r,*
^L*
= atO.

+
(26) C?rw}

Now substituting (26) into (22), we get

4Q<
=
-2c«r„ - 2c^Tii + c?Tn + <%TU.

Moreover, by the well-known anti-symmetry [E^, Ej] =

— [Ei, EJ of structure constants, — = and cj1 =

cj'

0.
c}1
Substituting back and dividing by four, we get our final
formula19

(27) o,• = cpr„.

Obviously, we get the corresponding formula

H^hT»'= -T,-^Thi
(27') Q<
=

in the case of the E*-axis. Hence steady motion particle

of
a

on any Riemannian group manifold under EA requires the

G

external force (27')• Alternatively, cfThj/Thh (summed over j

but not expresses the geodesic curvature of the one-param
h)

eter subgroup exp (<E») on G. normal ortho

If

we choose
a

gonal basis Ei, En, relative to the metric ds2 at this

0,

•
•
•

curvature simply 4*.

is

13. Classical Applications

From the single general formula (27'), and the special

commutation relations (18) for the Euclidean group, one can
derive the external force required to maintain steady trans
lation or steady rotation in an ideal fluid. (The force exerted
(27) was obtained in 1945 by John Breakwell and myself,
Formula
19

using independent methods. See Abstract 52-7-242, Bull. Am. Math.

Soc., 68 (1946), 617.
173
 VIRTUAL MASS AND GROUPS [CH. V

by the fluid on the body is of course the reverse of this force.)

Thus under steady translation Ei along the Zi-axis, the force
required is

(28) (o,o,o,•o,7,1.,-rll).

Under steady rotation Eu with angular velocity of one radian

per second about the Zi-axis, the force is similarly,

(29) (olr«'-r«)o,r«,-r«i).
These are the classical formulas of Kirchhoff and Kelvin;20
it will be noted that (28) reduces to 0 (steady motion under
no constraints), if the kinetic energy tensor ) r«9,oy for

translation has been diagonalized. In other words, steady

translation without external forces is (theoretically) possible
along the characteristic axes of the translation kinetic energy
tensor, and no others.21
Although the preceding formulas are highly theoretical,
and steady translation in the absence of external forces is not
physically possible, formula (28) does provide the classical
explanation for the broadsiding tendency of a flat plate. Using
(28), it is not hard to show that the stable steady translation
is along the characteristic axis maximizing the kinetic energy
tensor for translation. This result is in qualitative accord
with experiment.
We may note as another application of our general formula
(27'), the fact that the component of generalized force is
zero, in any direction corresponding to an infinitesimal trans
formation which is permutable with the steady motion in
question. For if [Eh, E,] = 0, then Q< is zero.
This provides a group-theoretic motivation for the d'Alem-
bert Paradox: steady translation along an axis gives rise
(theoretically) to no translation thrust at all, but only to a
couple, since all translations are permutable. Since transla
tions and rotations about the same axis are permutable, transla-

10
See [9], p. 169, formula (4), for (28); 29 is implicit in [9], Section 125.
See also the references given there.
11
The stability of such translations has recently been considered by
H. D. Ursell, Proc. Camb. Phil. Soc., 37 (1941), 150-167.
174
§14] GENERALIZATIONS

tion also gives rise to no torque about the axis of translation ;

the moment axis is perpendicular to the axis of translation.
The same consideration leads immediately to what may be
called the Propeller Paradox. Screw motion about an axis
can (in the classical theory) give rise to no axial thrust or
torque about the axis of the motion! Hence for a propeller,
or other object possessing n-fold rotational symmetry about
that axis (n > 1), all force components are (theoretically)
zero.22

14• Generalizations

The preceding formulas have obvious analogs, for imaginary

rigid bodies moving in an ideal fluid in a non-Euclidean
geometry (i.e. n-dimensional spherical, elliptic, or hyperbolic
space). In fact, analogs can be stated without difficulty,
for rigid motions in any "homogeneous" space, or space
possessing a continuous group of rigid motions. (By defi
nition, any Riemannian group manifold is such a space.)
It is of course doubtful whether such analogs have even
the limited physical significance of the classical formulas.
Further, it should be realized clearly that the generalization
to non-Euclidean spaces has not been completely established,
because the validity of the assumption that the dynamical
system involved is "Lagrangian" will not be established
except for the case of Euclidean spaces.
Nevertheless, it may be of interest to state some of these
analogs, as illustrating the effect of curvature of space (if it
exists) on the reactions of an unlimited ideal fluid to steady
motion.
Since any motion is permutable with itself, steady trans
lation of a solid through an ideal fluid filling non-Euclidean
space may be expected to cause no drag on the solid, and
uniform rotation about an axis should cause no torque about
that axis. Moreover, since translation along any axis is
permutable with rotation about that axis, neither translation,

" Incidentally, since the exterior of a propeller is simply connected,

we cannot remedy the situation by introducing "circulation" about
the propeller blades — without considerable "fudging."
175
 VIRTUAL MASS AND GROUPS [CH. V

rotation, nor screw motion about any axis in a non-Euclidean

geometry should cause any thrust along axis or torque about
That the Propeller Paradox remains valid in non-

is,
that axis.
Euclidean geometry.
On the other hand, no longer true that all translations

it
is
are permutable, in non-Euclidean geometry. Hence should

it
be possible for translation to give rise to cross-force in non-

a
Euclidean space, even in perfect fluid without circulation!

a
It thus no longer true that translation gives rise to no
is

cross-force, but only to pure couple — indeed one can

if
a
define "pure couple" in non-Euclidean geometry. Hence
a

impossible to extend the full d'Alembert Paradox to

it
is

non-Euclidean geometry.
more plausible application to rigid body which
A

is

is
a
constrained to move parallel to two fixed planes, bounding
an otherwise infinite perfect fluid. Since, in this case, we are
dealing with the "Euclidean group" of rigid motions of the
plane, the d'Alembert Paradox does apply, and the force of
reaction under steady motion (theoretically) pure moment.
is

a
Still another application can be made to a fluid bounded
by two coaxial circular cylinders, and rigid body moving
a

parallel to, or under rotation around, the common axis of the

cylinders. There will be no cross-force.
The case of translation parallel to the axis of a cylindrical
pipe or canal, of any cross-section, follows from the same
principle.23

Proof that System Lagrangian'

is

15.

shall now prove, for the case of finite body in space, the
a
I

validity of the assumption that the forces Q<, defined by the

variational equations (11) of Lagrange, are in fact the com
ponents Q* of pressure thrust resp. moment in the ordinary
sense. These are, of course, defined mathematically as
integrals over the body surface

Qf =
(30) fpN4S.
f

"This case was treated by U. Cisotti, Bendic. Palermo, 28 (1909),

307-352.
176
§15] PROOF THAT SYSTEM IS LAGRANGIAN

Here Ni denotes the normal component of surface displace

ment under translation resp. rotation, corresponding to the
1th generalized coordinate, and p is defined by the Bernoulli

equation

it]
p + + = p.(o,
(300 5p[v£w
for accelerated motion in an ideal fluid (Chapter (6")),

I,
provided that hydrostatic buoyancy forces are neglected as
usual.
The proof made difficultby the facts that the total mass
is

involved infinite, and that the "configuration space" of the

is

fluid has an infinite number of dimensions — corresponding

to the number of degrees of freedom of fluid motion. The only
discussions of the point in the literature are due to Lamb,
([74], and [9], Sections 135-136) and seem not fully satisfac
tory.24 shall therefore present new and very elegant
a
I

variational proof due, with minor modifications, to Mr. John

Breakwell. In presenting this proof, shall follow the sug
I

gestive notation for variational differentials generally used

8

in dynamics, even though the derivative notation pre

is
ferred by most modern authors writing about the calculus of
variations.26
Two general theorems will be used, without proof. The
first simply Euler's identity for the first variation
is

Tdt =
fh

fh

(31) QiSqi(t)dt,
8

Jin Jto

with the Qt as in (11). The second the fact that, in (31),

is

all variations 5g,(<) are possible, subject only to 8qi(to) =

Sqi(h) = This may be expressed by saying that the
0.

configuration space "holonomic" for the rigid body. It

is

Q,

follows that, in order to prove the identity = Q* in com-

** Thus in [74], the integrals involved are not convergent, and the
exact assumptions underlying the discussion of [9] seem somewhat
obscure.
26
M. Morse, The Calculus Variations in the Large, New York, 1934;
of

G. A. Bliss, Lectures on the Calculus Variations, Chicago, 1946.

of

177
 VIRTUAL MASS AND GROUPS [CH. V

paring (11) and (30), it is sufficient to prove the first identity of

J* Q?5q<(t)dt I J* Tdt =

jTh
(32) dt
fliVUVUdmY
&

|s
f
f
The second identity obvious, writing dm for the mass-

is
differential pdR in (2).
One can prove (32) by manipulating its right-hand term
as follows. The manipulation permissible since VUVU =

is
0(l/r6) as before, whence the quadruple integrals over space-
time are absolutely convergent, and interchange of the order
of integration allowed. First, using "Lagrangian" coordi
is

nates moving with the fluid, we get immediately

-/>{///,*H
m />{a///^H= dm Ui8u,dt\, where ut = -—
I
I

I
\
I

Integrating by parts, we get for each fluid particle

—
'

UiSuidt =
'

(34) UiSxi ciiSxidt,

/
I

where a* = du,/dt denotes acceleration. By the equations of

motion (Chapter (3)), with the buoyancy force of gravity
I,

—
neglected, a* = dp/p dxt, hence

div (p&x) =
W div (&c)
LS

(^-)
+

Sxi
p

(340
= = -paiSxi'
I&x<(teJ
since div (Sx) in an incompressible fluid. Similarly,
=
0

div (U8x) SuiSxi. Substituting back from (34) in the last

=

expression of (33), and from (34') and the other divergence

formula in (34), we get

fff div (p5x)di}•

+

f*

(33') dR[div (pUSx)]'^ <ffi

f[
f
{

Since = 0(l/r6), 0(r), and dR = 4rr2dr, the quadruple

8x =
p

integral absolutely convergent as before. Hence we can

is

178
§16] CONCLUSIONS

reverse the order of integration, and then apply the Divergence

Theorem,26 to get

05) i
£t* =
[JfaPm*ndsl + /;* {ffgP*[Link]
Here Sxn denotes the component of dx normal to the common
boundary S of the solid and liquid. Since the solid and liquid
are in contact, 5xn = 2iV,•5g<, in the notation of (30). In
particular, 5xn -
0 at U and and the first term on the

ti,
right-hand side of (35) vanishes, since Sqi(to) = 8g,(2i) = 0.
From Sxn = 2Ni8qi, we get also

= pNtBqidS
fapSxndS =^Qf&qi,
£
f

ff
by (30). Substituting back from these results into (35), we
get
Tdt =
fh
/""

Q?8qi(t)dt.
+

(35')
0
5

This proves (32), hence Q? = Qit and hence the validity of

my contention that the dynamics of rigid body in perfect

a
a

fluid can be subsumed under the geometry of continuous

groups.

16, General Conclusions

What the moral of the preceding results? there

If
is

is
a
moral, philosophical, and therefore very old — yet often
it
is

forgotten. It that there is hidden order in Nature, to be found

is

only by patient search.

In the first two chapters, stressed the need for patient
I

search, by showing the paradoxes to which purely dialectical

arguments led, and the complexity of the real facts. In this,
simply followed AristotleVmaxim. "Let us first understand
I

the facts, and then we may seek for the causes."27 Moreover
" To justifythis, we need the theorem, due to Lagrange, that the fluid
particles in contact with the rigid body form an invariant set ([12],
Vol. p. 97). Also, we need to remark that the surface integrals, over
1,

large spheres, of USxn and pSxn, tend to zero, since 5xn = 0(J«dt) =
0(l/r>).
Borrowed from E. T. Whittaker's brilliant monograph on the
21

philosophy of science, Space and Spirit, Edinburgh, 1946. Aristotle's

views, which are often misstated, are given on p. 27.
179
 VIRTUAL MASS AND GROUPS [CH. V

I continued in the same vein for much of the third chapter, by

exposing the pitfalls of the usual theories of modeling.
But already in the third chapter I began to show how
group theory provided a hidden order in fluid mechanics.
This theme was continued with greater and greater emphasis
in the last two lectures, revealing the group concept as a
unifying principle in innumerable questions of fluid mechan
ics — as indeed it has proved to be already in other branches
of physics.
Although the moral is old, the illustrations I have given are
new. Almost none of the material may be found in the
standard treatises on fluid mechanics. Thus I have tried to
uncover one aspect of the hidden order in Nature.

180
 Bibliography
Chapter I and General References

[1] G. Birkhoff, "Reversibility paradox and two-dimensional airfoil

theory," Am. Jour. Math., 68 (1946), 247-256.
[2] G. Cisotti, "Una notevole eccezione del teorema di Kutta-Joukow-
ski," Rendiconti Acad. Lincei, 5 (1927), 16-21; also ibid., 7 (1928)
17-19 and 538-543; and H. Pistolesi, ibid., 12 (1930), 409-411.
[3] C. Cranz, Handbook of Ballistics (translated), Vol. 1, His Majesty's
Stationary Office, London, 1921.
[4] P. Curie, Oeuvres Scientifijues, Paris, 1908, esp. pp. 119, 215.
[5] W. Durand and others, Aerodynamic Theory, 5 vols., Berlin, 1935.
[6] S. Goldstein and others, Modern Developments in Fluid Dynamics,
2 Vols., Oxford, 1938. Revised three volume edition in preparation.
[7] O. D. Kellogg, Potential Theory, Springer, 1929; reprinted by Dover.
[8] Z. Kopal, Tables of Supersonic Flow Around Yawing Cones, Mass.
Inst, of Technology (1947), esp. pp. xvi-xvii; further details in
Report No. 2 of M.I.T. Center of Analysis, Dept. of Electrical
Engineering.
[9] H. Lamb, Hydrodynamics, 6th ed., Cambridge, 1932. See also
the Appendix of the second German edition, Lehrbuch der Hydro-
dynamik, Leipzig, 1931.
[10] L. M. Milne-Thomson, Theoretical Hydrodynamics, Oxford, 1938;
second edition in preparation.
[11] C. W. Oseen, Neuere Methoden und Ergebnisse in der Hydrodynamik,
Leipzig, 1927.
[12] L. Prandtl and O. G. Tietjens, Hydro- and Aeromechanics, McGraw-
Hill, 1934, 2 Vols.
[13] Rayleigh, J. W. Strutt Lord, Scientific Papers, 6 Vols., Cambridge
University Press, 1899-1920.
[14] Hunter Rouse, Fluid Mechanics for Hydraulic Engineers, McGraw-
Hill, 1938.

Chapter II
[15] J. Ackeret, "Experimentelle und theoretische Untersuchungen ilber
Kavitation," Techn. Mech. und Thermodyn., 1 (1930), 1-21 and
63-72.
[16] A. Betz and E. Petersohn, "Anwendung der Theorie der freien
Strahlen," Ingenieur Archiv, 2 (1931), 190-211.
[17] G. Birkhoff, "Remarks on streamlines of discontinuity," Revista
Ciencias Lima, 50 (1948), 105-116.
[18] G. Birkhoff, Modeling of Entry into Water, Applied Math. Panel,
National Defense Research Council, May 1945 (declassified).
[19] G. Birkhoff and T. E. Caywood, "Water entry flow patterns,"
Jour. Applied Physics, 2O (1949), 646-659.
[20] G. Birkhoff, M. Plesset, and N. Simmons, "Wall effects in davity
flow," Quar. Applied Math., 8 (1950).
181
 BIBLIOGRAPHY

[21] G. Birkhoff, D. P. MacDougall, E. Pugh, and Sir Geoffrey Taylor,

"Explosives with lined cavities," Jour. Applied Physics,. 19 (1948),
563-582.
[22] M. Brillouin. "Les surfaces de glissement de Helmholtz et la resist
ance des fluides," Annates de chimie et physique, 23 (1911), 145-230.
[23] R. M. Davies, "The influence of atmospheric pressure on the phe
nomena accompanying the fall of small scale projectiles into a
liquid," 1944 (unclassified mimeographed report).
[24] D. Gilbarg and R. A. Anderson, "Influence of atmospheric pressure
on . . . entry of spheres into water," Jour. Applied Physics, 19
(1948), 157-169.
[25] G. Greenhill, "Theory of a streamline past a plane barrier," Reps.
and Memos., 19, Adv. Comm. Aero. (Great Britain), 1910.
[26] H. von Helmholtz, "Uber discontinuierliche Flussigkeitsbeweg-
ungen," Monatsber Berliner Akad. Wiss., April 23, 1868, 215-228;
reprinted in Phil. Mag., Nov. 1868.
[27] Th. von Karman, "The engineer grapples with non-linear problems,"
Bull. Amer. Math. Soc., 46 (1940), 616-683.
[27'] G. Kirchhoff, "Zur Theorie freier Flussigkeitsstrahlen," Crelle's
Jour, fur Math., 70 (1869), 289-298.
[28] J. Leray, "Les problemes de representation conforme de Helmholtz,
theorie des sillages et des proues," Commentarii Math. Helv., 8
(1935), 149-180 and 250-263.
[29] T. Levi-Civita, "Scie e leggi de resistenzia," Rendiconti Circolo Mat.
Palermo, 23 (1907), 1-37.
[30] N. Levinson, "The asymptotic shape of the cavity . . . ," Annals
of Math., 47 (1946), 704-730.
[31] A. E. Love, Enzykl. Math. Wissenschaften Vol. iv, part 3, p. 97.
[32] D. Riabouchinsky, Proc. Land. Math. Soc., 19 (1921), 206-215.
See also ibid., 25 (1926), 185-194, and E. G. C. Poole, ibid., 25
(1926), 195-212.
[33] R. Southwell and G. Vaisey, "Fluid Motions Characterized by
'Free' Streamlines," Phil. Trans., 240 (1946), 117-161.
[34] A. Weinstein, "Conformal representation and hydrodynamics,"
Proc. First Canadian Math. Congress (1945), Univ. of Toronto Press,
1947, pp. 355-364.
[34'] A. Weinstein, "Non-linear problems in the theory of fluid motion
with free boundaries," Proc. First Symposium Applied Math., pub
lished by Am. Math. Society, 1949.
[35] A. M. Worthington, A Study of Splashes, Longmans, 1908.

Chapter III
[36] J. Bertrand, "Note sur la similitude en mecanique," Jour, de I'Sc.
Polytechnique, 19 (1848), 189-197.
[37] G. Birkhoff and S. MacLane, Survey of Modern Algebra, New York,
1941.
[38] P. W. Bridgman, Dimensional Analysis, 2d ed., New Haven, 1931.
[39] J. M.
Burgers, E. Dobbinga, L. Troost, and M. J. Bossen, Model-
proeven en Kengrootheden op Stromingsgebied, Amsterdam, 1947.
[40] E. Buckingham, "On physically similar systems," Physical Review, 4
(1914), 354-376,
182
 BIBLIOGRAPHY

[41] E. Buckingham, "Model experiments and the forms of empirical

equations," Trans. Am. Soc. Mech. Eng., 37 (1915), 263-296.
[42] C. Camichel and L. Escande, "Similitude hydrodynamique et
technique des modeles reduits," Paris, 1938 (scarce).
[43] Norman Campbell, "Dimensional analysis," Phil. Mag., 47 (1924),
481^94.
[44] A. Craya, "Similitude des modeles fluviaux a fond fixe," La Houille
Blanche (Grenoble), 3 (1948), 335-346.
[45] Mrs. T. Ehrenfest-Afanassjewa, "Dimensionsbegriff und Bau
physikalischer Gleichungen," Math. Annalen, 77 (1915), 259-276.
[46] Mrs. T. Ehrenfest-Afanassjewa, "Dimensional analysis . . . theory
of similitudes," Phil. Mag., 1 (1926), 257-272.
[47] J. B. J. Fourier, Thiorie analytique de la chaleur," 1822, Eng. trans
lation by A. Freeman, Cambridge, Univ. Press, 1878.
[48] W. Froude, "On experiments with H. M. S. Greyhound," Trans.
Inst. Nov. Arch., 15 (1874), 36-59.
[49] A. Martinot-Lagarde, "Analyse dimensionelle. Applications a la
mecanique des fluides," Lille, 1946 (ozalid copies); printed in 1948
by the Groupement des Recherches Aeronautiques.
[50] A. W. Porter, The Method of Dimensions, Methuen, 1933.
[51] Osborne Reynolds, Collected Papers, esp. Vol. I, pp. 257-390, and
Vol. II, pp. 51-105.
[52] A. E. Ruark, "Inspectional Analysis: A method which supplements
dimensional analysis," Jour. Elisha Mitchell Sci. Soc., 51 (1935),
127-133.
[53] R. C. Tolman, "The principle of similitude," Physical Review, 3
(1914), 244-255.

Chapter IV
[54] K. Bechert, "Differentialgleichungen der Wellenausbreitung in
Gasen," Annalen der Physik, 39 (1941), 357-372.
[55] K. Bechert, "Ebene Wellen in idealen Gasen mit Reibung und
Warmeleitung," Annalen der Physik, Jfi (1941), 207-248.
[56] G. Birkhoff, "Dimensional analysis and partial differential equa
tions," Electrical Engineering, 67 (1948), 1185-1188.
[57] G. D. Birkhoff, "The principle of sufficient reason," Rice Institute
Pamphlet 28 (1941), No. 1, pp. 24-50.
[58] H. Blasius, "Grenzschichten in Flussigkeiten mit kleiner Reibung,"
Zeits. Math. Phys., 56 (1908), 1-37.
[59] C. Chevalley, Theory of Lie Groups, Princeton, 1946.
[60] R. Courant and K. Friedrichs, Supersonic Flow and Shock Waves,
New York, 1948.
[61] L. E. Dickson, "Differential equations from the group standpoint,"
Annals of Math., 25 (1924), 287-378.
[62] L. Infeld, "The factorization method and its application to differ
ential equations in theoretical physics," unpublished. See also Phys.
Rev. 59 (1941), 737-747
[63] L. Prandtl,"Neuere Untersuchungen fiber die stromende Be-
wegungen der Gase und Dampfe," Phys. Zeits., 8 (1907), 23-31.
[64] R. M. Redheffer, "Separation of Laplace's equation," Doctoral
Thesis, Massachusetts Institute of Technology (1948). See also
183
 BIBLIOGRAPHY

N. Levinson, B. Bogert, and R. Redheffer, "Separation of Laplace's

Equation," Quar. Appl. Math., 7 (1949), 241-262.
[65[ A. Rosenblatt, "Solutions exaotes des equations du mouvement des
liquides visqueux," Mem. Sci. Math., 72 (1935), 63 pp. •
[66] L. I. Sedov, "On unsteady motions of a compressible fluid," Doklady
U.R.S.S., j7 (1945), 91-93, and 52 (1946), 17-20.
[67] K. P. Staniukovitch, "On automodel solutions of equations of hydro
dynamics possessing central symmetry," Doklady U.R.S.S., 48
(1945), 310-312, 60 (1948), 1141-1144, and 64 (1949), 29-32, 179-181,
and 467-470.
[68] G. I. Taylor and J. W. Maccoll, "The air pressure on a cone moving
at high speeds," Proc. Roy. Soc., A139 (1944), 278-311.
[69] W. Tollmien, "Zum Ubergang von Unterschall in Uberschall-
stromungen," Setts. Ang. Math. Mech., 17 (1937), 117-136.
[70] E. T. Whittaker, Analytical Dynamics, 4th ed., Cambridge, 1937.

Chapter V

[71] Kelvin, Sir W. Thomson, Lord, Proc. Roy. Soc. Edinburgh, 7 (1869),
384-390 and 668-672. See also Math, and Phys. Papers, Vol. 4,
69-75, 93-97, and 101-114.
[72] Kelvin, Lord, " Hydrokinetic solutions and observations," Phil.
Mag., 4^ (1871), 362-380.
[73] G. Kirchhoff, "Uber die Bewegung eines Rotationskorpers in einer
Flussigkeit," Crelle's Jour., 71 (1869), 237-273.
[74] H. Lamb, "On the forces experienced by a solid moving through a
liquid," Quar. Jour., 19 (1883), 66-70.
[75] M. Shiftman and D. C. Spencer, "The flow of an ideal incompressible
fluid about a lens," Quar. Appl. Math., 3 (1947), 270-288.
[76] G. I. Taylor, "The energy of a body moving in an infinite fluid, with
an application to airships," Proc. Roy. Soc., A120 (1928), 13-22,
and 260-283.
[77] Th. Theodorsen, "Impulse and momentum in an infinite fluid,"
Th. von Karman Anniversary Volume (1941), 49-58.
[78] W. Thomson and P. G. Tait, Principles of Natural Philosophy,
2 Vols., esp. section 313ff.

184
 INDEX
Accolade, 50 Galilei-Newton group, 92
Approximation paradox, 31, 63, 93 Group, 91, 148, 167
Asymptotic paradox, 4
Hadamard paradox, 42
Bernoulli equation, 9, 71, 134 Helical products of inertia, 158
Borda mouthpiece, 45 Hodograph, 42, 55, 143
Boundary layer, 19, 126
Brillouin paradox, 55 Induced mass, 152ff.; induced ten
sor, 154ff.; induced tensor and
Canonical parameters, 170-171 momentum, 158
Cauchy-Kowalewski Thm., 96, 132 Inertial, 80, 82, 97, 146; inertial
Cavitation number, 54, 106 Lagrangian system, 166
Cavity charges, 67 Infinitesimal transformation, 169
Cavity induced mass, 164 Inspectional analysis, 90, 94, 98
Cisotti paradox, 18 Inverse methods, 140
Commutator, 170
Complex potential, 16, 42
Joukowsky flow, 16, 98
Conical flow, 120, 130
Conversion factor, 80
Kinematic viscosity, 91
Cusped cavity, 58, 61
Kirchhoff flow, 40ff.
Kopal's paradoxes, 25
D'Alembert paradox, 10, 19, 23,
Kutta-Joukowsky Thm., 16
174
Deep seal, 62
Lagrangian coordinates, 123, 178
Determinism principle, 95
Lagrangian system, 165
Dimension, 80
Lie group, 168
Dimensional analysis, 78, 80
Lined cavity charge, 67
Dimensionally homogeneous, 80, 85
Dirichlet flow, 10, 98
Discontinuity, 41 Mach model, HOff., 134
Dividing streamline, 20 Mach number, 81
Down-nosing, 63, 109 Magnus effect, 35
Drag, 13 Model, 77
Drag coefficient, 13
Dynamic similitude, 90 Navier-Stokes equations, 7, 91, 138
Non-viscous fluid, 7
Earnshaw paradox, 22
Eiffel paradox, 31 Paradox, 3
Equation of continuity, 6, 95, 123 Permutable, 171
NEquations of motion, 7 Pi Theorem, 82ff.
Equation of state, 6, 95 Poiseuille flow, 28
Euclidean group, 169 Poisson bracket, 170
Eulerian variables, 6 Polytropic, 120
Prandtl-Meyer, 119, 140
Fatness paradox, 19 Propeller paradox, 175
Free streamline, 41
Froude model, 102ff. Reentrant jet, 56
Froude number, 63, 94 Reversibility paradox, 14, 121
Fundamental units, 78, 80, 89 Reynolds model, 99, 105
185
 INDEX

Reynolds number, 8, 81, 98 Taylor-Maccoll, 121

Riabouchinsky flow, 55 Topological oversimplification, 21
Riemannian group manifold, 167 Turbulence, 28, 82, 100, 104
Rising bubble paradox, 34 Turbulence paradox, 27

Shaped charge (see lined cavity Units, 78

charge) Unit-free, 80, 83, 85
Shock wave, 23
Similitude, 90 Velocity potential, 9
Singular point paradox, 18 Virtual mass, 151ff.
Spiral flow, 124 Viscosity, 6
Steady motion, 167 Viscous fluid, 26
Stokes paradox, 33 Von Neumann paradox, 24
Structure constant, 171 Vorticity, 96
Surface seal, 62
Symmetric solutions, 116 Wake, 19, 40
Symmetry paradox, 5, 29 Wall corrections, 64

186
OTHER DOVER BOOKS

Archimedes. WORKS (including "The Method of

Archimedes"). Edited by T. L. Heath. 506pp. 5 3/8
x 8. Clothbound $4.95
Paperbound $1.95

Bateman, H. PARTIAL DIFFERENTIAL EQUATIONS

OF MATHEMATICAL PHYSICS. Index. 29 illus. xxii
+ 522pp. 6x9. Clothbound $4.95

Bridgman, P. W. THE NATURE OF PHYSICAL THE

ORY. Index, xi + 138pp. 5 3/8 x 8. Clothbound $2.50
Paperbound $1.25

Brillouin, Leon. WAVE PROPAGATION IN PERIODIC

STRUCTURES. Second revised edition. Index, xii
+ 259pp. 5 3/8 x 8. Clothbound $3.75
Paperbound $1.85

Carslaw, H. S. INTRODUCTION TO THE THEORY

OF FOURIER'S SERIES AND INTEGRALS. Third
revised edition. Index. 39 illus. xiii + 368pp. 5 3/8
8. Clothbound $4.50
Paperbound $1.95

Deimel, Richard F. MECHANICS OF THE GYRO

SCOPE: Dynamics of Rotation. 75 diagrams. 208pp.
5 3/8x8. Clothbound $3.50
Paperbound $1.60

Descartes, Rene*. THE GEOMETRY. The complete

translation by D. E. Smith and M. L. Latham, vii +
246pp. 5 3/8 x 8. Clothbound $2.95
Paperbound $1.50
Einstein, Lorentz, Minkowski, and Weyl. THE PRIN
CIPLE OF RELATIVITY. An English translation of
eleven of the most important original papers on the
general and special theories of relativity. Notes by
Sommerfeld. Translated by Perrett and Jeffrey, viii +
216pp. 53/8 x 8. Clothbound $3.50
Paperbound $1.50

Emmons, Howard W. GAS DYNAMICS TABLES FOR

AIR. 3 illus. 10 graphs. 4 tables. 46pp. 61/8 x 9%.
Paperbound $1.75

Fry, William J., Taylor, John M., and Henvis, Bertha W.

DESIGN OF CRYSTAL VIBRATING SYSTEMS. Second
revised edition. 126 diagrams, viii + 182pp. 6 1/8 x 9V4.
Clothbound $3.50

Galilei, Galileo. DIALOGUES CONCERNING TWO

NEW SCIENCES. Translated by Henry Crew and Al
fonso de Salvio. Introduction by Antonio Favaro. Index.
Bibliography. 126 diagrams, xxi + 300pp. 5 3/8 x 8.
Clothbound $3.50
Paperbound $1 .50

Hay, G. E. VECTOR AND TENSOR ANALYSIS. 208pp.

5 3/8 x 8. Clothbound $2.75
Paperbound $1.60

Heaviside, Oliver. ELECTROMAGNETIC THEORY.

Biographical introduction by Ernst Weber. Unabridged
one-volume edition of the three-volume work. Biblio
graphy, xxx + 386pp. 8 x 12%. Clothbound $7.50

Heisenberg, Werner. THE PHYSICAL PRINCIPLES

OF THE QUANTUM THEORY. Translated by Eckart
and Hoyt. Index, viii + 184pp. 5 3/8x8.
Clothbound $2.75
Paperbound $1.25
Helmholtz, Hermann L. F. SENSATIONS OF TONE.
Unabridged reissue of the last English edition, with
a new introduction by Henry Margenau. Index, xix +

576pp. 6 1/8 x 9%. Clothbound $4.95

Ince, E. L. ORDINARY DIFFERENTIAL EQUATIONS.

18 illus. viii + 558pp.
Fourth revised edition. Index.
5Vj x 9. Clothbound $4.95

Jahnke, Eugene and Emde, Fritz. TABLES OF FUNC

TIONS WITH FORMULAE AND CURVES. (Funktionen-
tafeln.) Fourth revised edition. Text in German and
English. Index. 212 illus. xv + 382pp. SlA x 8V2
Clothbound $3.95
Paperbound $1.95

Jeans, James. THE DYNAMICAL THEORY OF GASES.

Republication of the fourth revised edition. 444pp. 5

3/8 x 8. Clothbound $3.95

Paperbound $2.00

Jessop, H. [Link] Harris, F. C. PHOTOELASTICITY:

PRINCIPLES AND METHODS. Index. 164 diagrams,
vii + 184pp. 61/8 x 9%. Clothbound $3.75

Kellogg, Oliver Dimon. FOUNDATIONS OF POTEN

TIAL THEORY. Index, ix + 384pp. 5 3/8 x 8.
Clothbound $3.95
Paperbound $1.90

Klein, Felix. ELEMENTARY MATHEMATICS FROM

AN ADVANCED STANDPOINT. Vol. I: Arithmetic,
Algebra, Analysis. Translated from the third German
edition by E. R. Hedrick and C. A. Noble. Index .
125 illus. xiv + 274pp. 5 3/8 x 8. Paperbound $1.60
Klein, Felix. ELEMENTARY MATHEMATICS FROM
AN ADVANCED STANDPOINT. Vol. II: Geometry.
Translated from the third German edition by E. A.
Hedrickand [Link]. Index. 141 illus. ix + 214pp.
5 3/8x8. Paperbound $1.60

Kober, H. DICTIONARY OF CONFORMAL REPRE

SENTATIONS. 447 diagrams, xvi + 208pp. 61/8 x 9%.
Clothbound $3.95

Kraitchik, Maurice. MATHEMATICAL RECREATIONS.

Second revised edition. 181 illus. Over 40 tables.
328pp. 5 3/8 x 8. Paperbound $1.60

Lamb, Horace. HYDRODYNAMICS. Sixth revised edi

tion. 83 illus. xviii + 738pp. 6x9. Clothbound $5.95

Lorentz, H. A. THE THEORY OF ELECTRONS AND

ITS APPLICATIONS TO THE PHENOMENA OF
LIGHT AND RADIANT HEAT. Unabridged, unaltered
reissue of the second edition. Index. 9 figures. Appen
dix. 352pp. 5 3/8 x 8. Paperbound $1.80

Mach, Ernst. PRINCIPLES OF PHYSICAL OPTICS. .

Unabridged reissue of the English translation of the

last German edition. 10 plates. 280 diagrams, x +
324pp. 5 3/8 x 8. Clothbound $3.50
Paperbound $1.75

MacRobert, T. M. SPHERICAL HARMONICS. An

Elementary Treatise Harmonic Functions
on with
Applications. Second revised edition. Index. 20 dia
grams, vi + 372pp. 5% x &1/,. Clothbound $4.50

Maxwell, James Clerk. ELECTRICITY AND MAGNET

ISM. An unabridged republication of the third edition.
Two volumes bound as one. Vol. I: xxxii + 506pp.
Vol. II: xxiv + 500pp. 5 3/8 x 8. Clothbound $4.95

UNIVERSITYOF MICHIGAN

9015 01436 0039