VON NEUMANN ALGEBRAS IN PHYSICS: A ROUGH SKETCH OF

ALGEBRAIC QUANTUM FIELD THEORY

ULRICH PENNIG

Abstract. These are the notes from a talk I gave during a seminar about von Neumann
algebras. It deals with the role of type III-factors in algebraic quantum field theory, ex-
plains the KMS condition and the characterization of type III-factors via their modular
conjugation group.

1. Introduction
We give a short introduction into the main aspects of quantum field theory, in particular
algebraic quantum field theory via nets of von Neumann algebras. We assume no familiarity
with the concepts of mechanics and quantum mechanics, which we will briefly sketch in the
beginning. We pay particular attention to the case, where the local algebras are type III1 -
factors, which is supported by most examples from physics. In this case states of the net
can be changed locally to eigenstates without disturbing them on the causal complement. A
thought experiment by Fermi will be discussed, which shows that quantum mechanics alone
does not have these features and is therefore a strictly non-relativistic theory [7, 1].
In a second part we will explain the KMS condition and how it can be used to characterize
the modular group. This also reveals a nice characterization of type III-factors as those
with modular conjugation, which is outer for at least one state and at some value t ∈ R.
The material for this part is taken from [4].
2. From classical mechanics to quantum field theory
2.1. Classical mechanics. In classical Hamiltonian mechanics a physical system with ener-
gy conservation is described by functions on the phase space X. The latter captures the
information about the space coordinates and the momenta of all particles. For n point
particles moving in R3 this would be R6n . But it may be more complex, e.g. the particles
may be confined to a manifold M , in which case the phase space can be identified with the
cotangent bundle T ∗ M . The collection of all possible space coordinates is often called the
configuration space of the system. We will assume X = R6n and denote the points by (xji , pji )
with xji , pji ∈ R the ith space coordinate and ith momentum coordinate of the jth particle.
The dynamics of the system are described by
(1) the energy or Hamilton function on the phase space, often denoted by H : X → R.
(2) the Hamiltonian equations
∂H ∂H
ẋji = j , ṗji = − j
∂pi ∂xi
where the dot denotes the time derivative as is common in physics.
Date: 2011-03-22.
1
 (3) initial conditions for these differential equations.
Example 2.1. Let us assume we have one particle with mass m moving along the real axis
p2
with dynamics described by H(x, p) = 2m . Then we get
p
ẋ = =: v , ṗ = 0 ⇒ x(t) = v · t + x0 ,
m
which describes the movement of a free particle with initial velocity v and initial coordi-
nate x0 .
Example 2.2. Assume the particles moves along the real axis, but this time with dynamics
p2 2
described by H(x, p) = 2m + x2 and initial momentum p(0) = 0. Then
 
p t
ẋ = , ṗ = −x ⇒ x(t) = x0 cos √ .
m m
This is the movement of the so called harmonic oscillator. The particle oscillates around a
fixed coordinate with starting point x0 .
2.2. Quantum mechanics. Quantum mechanics arose from observations that showed that
classical mechanics breaks down at an atomic level. For example the movement of an electron
around the nucleus seemed to be confined to certain discrete energy levels. A phenomenon
that could not be explained using Hamiltonian mechanics. Bohr, Schrödinger, Heisenberg
and many others worked out the basics of quantum mechanics. The idea is to replace the
phase space X by the Hilbert space H = L2 (M, C) of complex valued square integrable
functions on the configuration space M and to replace the energy function H by an operator
in the following way
(1) exchange xji by the multiplication operator with the coordinate xji .
(2) exchange the momentum coordinate by the differential operator ~i ∂x∂ j , where the latter
i
is understood as a densely defined unbounded operator on H.
Note that this procedure is far from being unique or even unambiguous due to the fact
that the variables xji and pji commute whereas the operators do not. Nevertheless, for most
physical systems there is a sensible way to write down a Hamiltonian operator H. The
dynamics are now described by the Schrödinger equation for a state ψ ∈ H:
 
∂
i~ ψ (x, t) = (Hψ)(x, t)
∂t
and an initial state ψ0 ∈ H. The quantity ~ is Planck’s constant. If H is time independent,
the solution is given by  i 
ψ(x, t) = e− ~ tH ψ0 (x) .
So, we need the eigenspace decomposition of the operator H to compute the right hand
side. The interpretation of ψ is now of a stochastical nature. If ψ is a unit vector, then the
norm-square |ψ(x, t)|2 is the probability density for the presence of the particle at the point
x and time t. Thus, if U ⊂ M is a (measurable) subset of the configuration space, I is a
time interval and χU , χI are the characteristic functions of these sets, then
Z
hψ, χU · χI · ψi = χU (x)χI (t)|ψ(x, t)|2 dxdt
U ×I
2
is the probability to find the particle inside of U during the interval I. Likewise, any other
observable quantity is represented by a self-adjoint operator and its expectation value is
given by hψ, A ψi. For an observable A the only values that will appear in a measurement
are its eigenvalues. Since these may well be discrete, it solves the problems sketched at the
beginning of this section.
Example 2.3. Let us review the example of the free particle, but now consider the interval
[0, 1] as the configurations space, i.e. the particle is contained in a box. The Hilbert space is
H = L2 ([0, 1], C). At the boundary of the box we impose the conditions ψ0 (0) = ψ0 (1) = 0.
The Hamiltonian for the free particle is
~2 d2
H=− .
2m dx2
Its eigenvalues En are indeed discrete and labeled by the natural numbers, precisely we have
~2 π 2
En = n2 · .
2m
So the energy in our toy model is quantized. If we think about ψ as a (probability) wave,
then our boundary conditions force this wave to reflect at the points 0 and 1, therefore it
interferes with itself and either annihilates or strengthens itself depending on its energy,
which is basically its frequency.
Let us take a closer look at the expectation value of an observable A. It can be rewritten
as
i i
hψ, A ψi = hψ0 , e ~ tH A e− ~ tH ψ0 i .
Thus, instead of seeing the wave functions as dynamical quantities, we may as well think
about the observables as our fundamental objects! The time evolution yields
i i
αt (A) = e ~ tH A e− ~ tH
mapping the observables into themselves and the expectation value becomes a linear func-
tional
ω(A) = hψ0 , A ψ0 i .
So, why not study the algebra generated by the observables with the expectation values being
linear functionals on the algebra and the time evolution inducing an automorphism of it?
This is the Heisenberg picture of quantum mechanics, whereas the one that focusses on wave
functions is called the Schrödinger picture. The way to get from an abstract to a concrete
algebra is via its representation as operators on Hilbert spaces. We can represent the space
coordinate X by a multiplication operator and the momentum P by a derivative. But the
question remains, if there are other interesting representations of the algebra generated by
X and P . A theorem of Stone and von Neumann tells us that this is not the case.
Theorem 2.4 (Stone-von Neumann). The representation of the algebra generated by P and
X is unique up to unitary equivalence.
For example, we may simply exchange the roles of P and X and represent P by multi-
plication and X by a differential operator, then the two pictures are related by the unitary
equivalence given by Fourier transform.
3
2.3. Quantum field theory. As powerful as quantum mechanics is to describe physical
systems at an atomic scale, there still are problems: It is limited to systems with finitely many
degrees of freedom. So it is not possible to quantize continuously distributed observables like
the energy density of an electromagnetic field and there is no description of thermodynamic
limits. Moreover, it is inherently non-relativistic, even worse we have separated the time
evolution from the spacial part. In a picture compatible with special relativity they should
be treated on an equal footing.
To overcome these difficulties is the goal of quantum field theory. As mentioned above,
in a field theory the observables themselves are densities on the configuration space. By
what we learned, we have to replace them by operator valued densities or operator valued
distributions. It is not enough to simply work with operator valued functions, since an energy
measurement at a single point would need infinitely many energy, so H would not be defined
everywhere. Let S(R4 ) be the space of Schwartz functions on the Minkowski space R4 . A
field φ would then be a linear functional
φ : S(R4 ) → U B(H)
into the unbounded operators on a separable Hilbert space H. Of course we omit many
details about domains of definition here. The test function from S(R4 ) can be seen as
characterizing a region over which the measurement is smeared out or averaged.
Wightman developed a set of axioms for those fields. The most important of these are
• transformation behavior: There is a unitary representation U of the Poincaré
group G on H and the fields are covariant in the sense that U (g)ψ(f )U (g)∗ = ψ(g · f )
for g ∈ G.
• locality: If the supports of f, h ∈ S(R4 ) are space-like to each other, the correspond-
ing field operators commute, i.e. [ψ(f ), ψ(h)] = 0. The reasoning here is of course
that measurements, which happen in space-like separated regions are not allowed to
influence each other by special relativity.
Indeed, this leads to a setup for quantum field theory nowadays known as axiomatic qft.
Due to the analytical difficulties, it is quite hard to deal with. But there is another approach
we will sketch in the next paragraph.
2.3.1. Algebraic quantum field theory. As we have seen in the last chapters, the physical
content of a theory and its dynamics are contained in the algebra of observables and its
automorphisms. But working with unbounded densely defined operators forces us to take
care about domains all the time. One way to circumvent these difficulties is to study the
algebra of bounded functions of observables, which yields an algebra of bounded operators.
This lead to the definition of a net of operator algebras as developed by Haag and Kastler.
An introduction into this subject can be found in [3]. Before getting into the details of their
definition, we take a closer look at sensible regions in space-time, in which our measurements
will take place.
Suppose we have a detector, which we switch on at a space-time point (t1 , x1 ). Any event
that can be influenced by this lies in the forward light cone that has its tip at this point.
We switch off our detector at (t2 , x2 ). Any event that could have influenced our detector
up to then lies in the backward light cone at (t2 , x2 ). The intersecting region is called the
causal closure of the two points or a double cone. The term causal closure refers to the fact
that these are all points lying space-like to all points, which are space-like to both of the two
4
 Figure 1. A double cone in three-dimensional Minkowski space

points. The set C of all open double cones in Minkowski space forms a basis for its topology
and will be of fundamental importance for the next definition.
Definition 2.5. Let H be a fixed separable Hilbert space and let vNa(H) be the set of all
von Neumann algebras inside B(H). A map
A : C → vNa(H) , O 7→ A(O)
will becalled a net of von Neumann algebras if it has the following properties
•isotony: Let O1 , O2 ∈ C. If O1 ⊆ O2 , then A(O1 ) ⊂ A(O2 ) as a unital subalgebra.
•locality: If O1 is space-like to O2 , then [A(O1 ), A(O2 )] = 0.
•covariance: There exists a strongly continuous unitary representation of the Poincaré
group G on H, such that U (g)A(O)U (g)∗ = A(g O).
• positivity: The generator of translations has spectrum lying in the forward light
cone.
Clearly, the locality and covariance postulate were motivated by the Wightman axioms.
Isotony seems plausible: It should not matter, if I measure an observable in a ”bigger
laboratory”. The positivity assumption ensures the positivity of the energy operator. This
is only the lowest common denominator among all possible axioms, there may be more, but
this is the basic setup.
An astounding observation is that in all (irreducible) examples from physics, the local
algebras A(O) turn out to be isomorphic to the unique hyperfinite type III1 -factor. (This
has not been proven directly from the axioms, but a slightly weaker property called the
Borchers property holds.)
2.4. Type III-factors. During the seminar we have met von Neumann algebras of type In
for n ∈ N ∪ {∞}, type II1 and type II∞ . Let us review their definitions: Let M be a factor.
• M is of type I if there exists a minimal projection in M .
• M is of type II1 if there is a normal normalized trace tr : M → C on M . This is
equivalent to saying that all projections are finite, which is equivalent to 1 ∈ M being
finite.
• M is of type II∞ if 1 ∈ M is infinite, but M contains at least one finite projection.
As we see, if a factor M is of type III, i.e. does not belong to the list above, all projections
in M are infinite. An important corollary of this is the following:
5
Corollary 2.6. If M is a type III-factor on a separable Hilbert space H, then all projections
in M are (Murray-von Neumann) equivalent to 1 ∈ M .
Proof. Let e ∈ M be a projection. By the above considerations it is infinite. Therefore
∃u ∈ M such that u∗ u = e and e1 := uu∗ < e. Let en = uen−1 u∗ . Observe that e2 = ue1 u∗ <
uu∗ = e1 , thus by induction

en+1 = uen u∗ < uen−1 u∗ = en if en < en−1 .

The ei form a decreasing sequence of projections. Let pn = en − en+1 . These are orthogonal
projections, which are all equivalent to p1 and with p∞ = e − e1 we have
∞
X
e = p∞ + pn .
n=1
P
Let fi be a maximal system of orthogonal projections with fi -Pp1 . Suppose f0 := 1− i fi >
0. Then there is a projection 0 6= pe ∈ M with pe ≤ 1 − i fi and pe - p1 , which is a
contradiction. Thus, we can represent 1 ∈ M as a sum of mutually orthogonal projections
fi ∈ M with fi - p1 . Therefore
X X X
1= fi - pn ≤ p n + p∞ = e .
i n n

Since we also have e ≤ 1, we get e ∼ 1 and we are done. 

Let M be a type III-factor with separable predual. As we have seen in the chapter
about Tomita-Takesaki theory, a choice of a cyclic and separating vector Ω ∈ H for M or
equivalently of a normal faithful state φ on M yields an unbounded positive operator ∆φ and
an anti-unitary operator Jφ with Jφ M Jφ = M 0 and ∆itφ M ∆−it
φ = M . The spectrum σ(∆φ )
depends on the choice of state φ. Thus, consider
\
S(M ) = σ(∆φ ) ⊂ R+ .
φ

Connes proved that S(M ) is in fact a multiplicaltive subgroup of R+ . Since it is also closed,
the only possibilities are S(M ) = {0, 1}, S(M ) = {0} ∪ {λn | n ∈ Z} for some 0 < λ < 1
or S(M ) = R+ . This yields a finer classification of type III-factors. A type III-factor M is
called
• of type III0 if S(M ) = {0, 1},
• of type IIIλ if S(M ) = {0} ∪ {λn | n ∈ Z},
• of type III1 if S(M ) = R+ .
In the last case Connes and Størmer proved the following result:
Theorem 2.7 (Connes-Størmer [2]). If M is a type III1 -factor with separable predual, then
we have: For every  > 0 and every two normal states φ and ψ there exists a unitary u ∈ M
with
kφ(u · u∗ ) − ψk <  .
6
2.4.1. Type III-factors and algebraic quantum field theory. Let A be a net of von Neumann
algebras with all A(O) isomorphic to a type III1 -factor. To any such net there is an asso-
ciated C ∗ -algebra A defined via
[ k·k
A= A(O) .
O
We can consider states of the net, by which we mean a continuous linear functional on A
which is normal on the local algebras A(O) ⊂ A. Let ω be such a state and let p ∈ A(O)
be a projection. By our hypothesis, there is a partial isometry v ∈ A(O) with v ∗ v = 1 and
vv ∗ = p. Now consider
ωv (a) = ω(v ∗ av) ,
which is again a state of the net. For a0 ∈ A(O)0 we have
ωv (a0 ) = ω(v ∗ a0 v) = ω(v ∗ va0 ) = ω(a0 ) .
In particular, this holds for a0 ∈ A(O0 ) ⊂ A(O)0 , if O0 is space-like with respect to O.
Moreover, ωv (p) = 1. The lesson to learn from this construction is that – due to type III –
we can alter the state ω locally inside the region O without affecting anything that lies space-
like to it. In our case ωv has become an eigenstate for p. By theorem 2.7 we can approximate
any state in any region O by any other state if the local algebras are of type III1 .
2.4.2. A thought experiment by Fermi. To see that these approximation properties and the
locality of states fail in the case of quantum mechanics, let us look at a thought experiment
by Fermi that shows its acausality. Suppose a and b are two atoms separated by a distance R.
Initially b is in an excited state and a is in its ground state. We would like to describe the
energy transfer from b to a. The Hilbert space for this problem would be
H = Ha ⊗ Hb ⊗ Hc ,
where Ha is the state space of a, Hb that of b and Hc captures all information about the
radiation field. The algebra is B(H). The initial state of the system can be described by
ω0 = ωa ⊗ ωb ⊗ ωc ,
where ωa is the ground state of a, ωb the excited state of b and ωc is the ground state of the
radiation field. As we have seen above, we get the state ωt at time t via
ωt = ω0 eitH · e−itH

for some Hamiltonian operator H. Suppose ωa is given by a vector ψa ∈ Ha . Let Fa =

ψa hψa , · i be the rank 1 projection onto the subspace generated by ψa and set
Ea = (1 − Fa ) ⊗ 1b ⊗ 1c ∈ B(H) .
The probability to find a in an excited state at time t is then given by P (t) := ωt (Ea ). If
quantum mechanics were a relativistic theory, P (t) should vanish for t < Rc . But
t 7→ Ea eitH φa
may be continued to an analytic function in the upper half plane. Thus, if it vanishes on an
interval, it vanishes on the full real line! We get ωt (Ea ) = 0 for all t ∈ R. The only option
left is that the change of state occurs instantaneously.
We have seen in the last section that states on nets of von Neumann algebras are much
more flexible: In a net one would consider regions Oa ,Ob centered around a and b that are
7
space-like to each other. The time evolved state ωt can be the ground state on A(Oa ) and the
excited state of the radiation field in A(Ob ) without any problems. There still is a projection
Ea , but this is not contained in the local algebra A(Oa ), but in a larger one. For a detailed
treatment of this problem see [7, 1].
3. The KMS-condition
Now that we have explained the importance of type III-factors in quantum field theory,
the question may arise, how to detect whether a particular algebra is of type III. In this
section we will sketch a way to attack this problem via modular theory. We start with a
formal calculation: Let M be a von Neumann algebra in B(H). Let Ω ∈ H be a cyclic and
separating vector and define φ(x) = hx Ω, Ωi. Denote the modular automorphism group by
σtφ (x) = ∆it x ∆−it . Suppose we could extend the domain of definition for the modular group
from R to the whole complex plane. Then
1
φ(xy) = hxy Ω, Ωi = hy Ω, x∗ Ωi = hy Ω, J∆ 2 x Ωi
1
= hy Ω, J∆− 2 ∆ x ∆−1 Ωi = h∆ x ∆−1 Ω, S y Ωi
φ
= hy ∆ x ∆−1 Ω, Ωi = φ(y σ−i (x))
If we set F (t) = φ(σtφ (x) y) we have F (t + i) = φ(y σtφ (x)) by the above calculation. We turn
this property into a definition.
Definition 3.1. Let αt be a strongly continuous 1-parameter automorphism subgroup of
Aut(M ) and let φ be a faithful normal state on M . αt satisfies the KMS condition (KMS
stands for Kubo, Martin and Schwinger [5, 6]) for φ if
• φ(αt (x)) = φ(x)
• for every x, y ∈ M there exists a continuous function F : C → C which is bounded
on the strip {z ∈ C | 0 ≤ Im(z) ≤ 1} and analytic in its interior such that for all
t ∈ R:
F (t) = φ(σtφ (x) y) , F (t + i) = φ(y αt (x)) .
The next theorem shows that our naive thinking above is not too far away from the truth
and that the KMS condition yields a nice characterization of the modular automorphism
group.
Theorem 3.2. If φ is a faithful normal state on a von Neumann algebra M , then σtφ is the
unique one parameter subgroup satisfying the KMS condition for φ.
Instead of proving this theorem, we give an important corollary.
Corollary 3.3. Let φ and M be as in theorem 3.2. For a ∈ M the following statements are
equivalent:
a) φ(ax) = φ(xa) for all x ∈ M .
b) σtφ (a) = a for all t ∈ R.
Proof. Let Ω ∈ H be the cyclic and separating vector obtained from the GNS construction
of φ, i.e. φ(x) = hx Ω, Ωi. To prove that a) implies b) we show first, that a Ω ∈ Dom(S ∗ ).
This follows from
a)
hS x Ω, a Ωi = hx∗ Ω, a Ωi = hΩ, x a Ωi = hΩ, a x Ωi = ha∗ Ω, x Ωi
8
and the Cauchy-Schwarz inequality. We also get S ∗ (aΩ) = a∗ Ω = S(aΩ) and therefore
1 1
J∆− 2 a Ω = J∆ 2 a Ω ⇒ ∆aΩ = aΩ .
But this implies σtφ (a)Ω
= ∆it a∆−it Ω = aΩ. Since Ω is separating, the claim follows.
To see that b) implies a) observe that
b)
φ(σtφ (x) · a) = φ(σtφ (xa)) = φ(xa) .
Therefore F (t) is constant. By the Schwarz reflection principle F (t) has an analytic contin-
uation into the strip with −i ≤ Im(z) ≤ i. Since it is constant along the real axis, it has to
be constant on the whole strip. But this implies a). 
We are now in the situation to compare the two different modular automorphisms obtained
from two different states.
Theorem 3.4. Let M be a von Neumann algebra and φ, ψ two normal faithful states on M .
Then there is a strongly continuous map u : R → U (M ) mapping t to ut such that
σtφ = Adut ◦ σtψ .
Proof. We will use a matrix trick to see this: Let Φ be the following normal faithful state on
M2 (C) ⊗ M = M2 (M )
1
Φ : M2 (C) ⊗ M → C , (xij )ij 7→ (φ(x11 ) + ψ(x22 )) .
2
Moreover let    
1 0 0 0
p = e11 = , q = e22 = .
0 0 0 1
Since Φ(p X) = Φ(X p), the modular automorphism σtΦ fixes p by corollary 3.3. Thus, σtΦ
acts on p M2 (M ) p and satisfies the KMS condition. By theorem 3.2 we get
σtΦ (x ⊗ e11 ) = σtφ (x) ⊗ e11 .
and similarly σtΦ (x ⊗ e22 ) = σtψ (x) ⊗ e22 . Now set
vt = σtΦ (1 ⊗ e21 ) ,
which satisfies vt vt∗ = σtΦ (q) = q and vt∗ vt = p. A small calculation shows that it can only be
of the form  
0 0
vt =
ut 0
for a unitary ut . The calculation
       
0 0 Φ x 0 ∗ Φ 0 0 0 0
= vt σ t v = σt =
0 ut σtφ (x)u∗t 0 0 t 0 x 0 σtψ (x)
implies the claim. 
The characterization of type III-factors alluded to in the beginning is now a corollary.
Corollary 3.5. Let M be a factor, such that σtφ is an outer automorphism for any φ and
t ∈ R, then M is of type III.
9
Proof. We have to exclude the other cases. In case of type In for n ∈ N ∪ {∞}, all au-
tomorphisms are inner. In case of type II1 , the trace provides a state with inner modular
conjugation. By theorem 3.4, every modular conjugation then acts via inner automorphisms.
In case of type II∞ we can choose a faithful normal state φ on B(H) and use tr ⊗ φ and
the fact that the modular conjugation on the whole algebra is the tensor product of the
conjugations of the two factors and therefore also inner. 
References
[1] Detlev Buchholz and Jakob Yngvason. There are no causality problems for Fermi’s
two-atom system. Phys. Rev. Lett., 73(5):613–616, 1994. 1, 8
[2] Alain Connes and Erling Størmer. Homogeneity of the state space of factors of type III1 .
J. Functional Analysis, 28(2):187–196, 1978. 6
[3] Rudolf Haag. Local quantum physics. Texts and Monographs in Physics. Springer-Verlag,
Berlin, 1992. Fields, particles, algebras. 4
[4] Vaughan F.R. Jones. Lecture notes Von Neumann Algebras. 2009. 1
[5] Ryogo Kubo. Statistical-mechanical theory of irreversible processes. I. General theory
and simple applications to magnetic and conduction problems. J. Phys. Soc. Japan,
12:570–586, 1957. 8
[6] Paul C. Martin and Julian Schwinger. Theory of many-particle systems. I. Phys. Rev.
(2), 115:1342–1373, 1959. 8
[7] Jakob Yngvason. The role of type III factors in quantum field theory. Rep. Math. Phys.,
55(1):135–147, 2005. 1, 8

10