June 26, 2014

Chapter 4

The Hilbert Transform

Additional Reading:
Bracewell (1986), pp. 267–273
Bendat and Piersol (1986), Chapter 13
von Storch and Zwiers (1999), Section 16.2

The Fourier transform considered in detail in chapters 2 and 3 is one of many transforms that
provide useful interpretations of a process y(t). One that is particularly useful for analysis of prop-
agating signals measured at different locations along the propagation path is the Hilbert transform,
which is defined in terms of a convolution (chapter 3). It will be seen that the Fourier transform
provides a simple interpretation of the Hilbert transform. The Hilbert transform properties sum-
marized in this chapter are useful later in empirical orthogonal function analysis of propagating
signals (see chapter 11).

4.1. Definition and Properties of the Hilbert Transform

Additional Reading:
Bendat and Piersol (1986), pp. 484–492

The Hilbert transform of a continuous, infinite (possibly complex) process y(t) can be defined
as
∞
y(t0 ) 0
Z
1
yH (t) = dt (4.1.1a)
π −∞ t − t0

= h(t) ∗ y(t) , (4.1.1b)

where h(t) ∗ y(t) represents the convolution (see chapter 3) of y(t) with the filter weighting function
h(t) given by
1
h(t) ≡ . (4.1.2)
πt
A convenient notation for the Hilbert transform of y(t) is

H [y(t)] = yH (t) . (4.1.3)

4–1

c D. B. Chelton June 26, 2014

It is shown in section 4.3 that the Hilbert transform is very useful for the study of both traveling-
wave and standing-wave processes y(t).
Since the Hilbert transform can be expressed as a convolution, it represents a filtered version
of the process y(t). The filtering properties of this convolution and physical interpretation of the
Hilbert transform are greatly simplified by the use of the Fourier transform. The effects of the
Hilbert transform in the frequency domain can be determined from the Fourier transform of yH (t),
which is easily obtained by applying the convolution theorem (see section 3.1.2 of chapter 3) to
(4.1.1b),
yH (t) ⊃ YH (f ) = H(f )Y (f ) , (4.1.4)

where Y (f ) and H(f ) are the Fourier transforms of y(t) and h(t), respectively. It can be shown
(see Bracewell, 1986, p. 130) that
1
⊃ H(f ) = − i sgn(f ) , (4.1.5a)
πt
where (
+1 if f ≥ 0
sgn(f ) ≡ (4.1.5b)
−1 if f < 0
The Fourier transform of yH (t) is therefore
(
−i Y (f ) if f ≥ 0
YH (f ) = − i sgn(f ) Y (f ) = (4.1.6)
+i Y (f ) if f < 0.

The Fourier transform of y(t) can be expressed in polar coordinates as

Y (f ) = AY (f ) eiφY (f ) . (4.1.7)

where the amplitude AY (f ) and phase φY (f ) are obtained in the usual way (see section 1.2.1 of
chapter 1) from the real and imaginary parts of Y (f ) by
 1/2
2 2
AY (f ) = Re [Y (f )] + Im [Y (f )]
 
−1 Im [Y (f )]
φY (f ) = tan
Re [Y (f )]

Since ±i = exp(±iπ/2) in polar coordinates (see section 1.2.1 of chapter 1), the filter transfer
function (4.1.5a) of the Hilbert transform can be written as

H(f ) = GH (f ) eiφH (f ) , (4.1.8a)

where the gain GH (f ) and phase φH (f ) are given by

GH (f ) = 1 for all f (4.1.8b)

(
−π/2 if f ≥ 0
φH (f ) = (4.1.8c)
+π/2 if f < 0

4–2

c D. B. Chelton June 26, 2014

The gain and phase of the filter transfer function of the Hilbert transform are shown in Figure
4.1.1. With (4.1.8), the Fourier transform of the Hilbert transform yH (t) can be represented as
(
Y (f ) e−iπ/2 = AY (f ) ei[φY (f )−π/2] if f ≥ 0
YH (f ) = (4.1.9)
Y (f ) e+iπ/2 = AY (f ) ei[φY (f )+π/2] if f < 0.

It is thus seen that the amplitude of each frequency component of the original process y(t) is
unchanged in yH (t). The Hilbert transform yH (t) obtained from the inverse Fourier transform of
YH (f ) is therefore equivalent to filtering the original process y(t) by shifting each positive frequency
component by −π/2 in phase (i.e., lagging each sinusoidal component in time by one quarter of its
period) and each negative frequency component by +π/2 in phase (i.e., leading each sinusoidal com-
ponent in time by one quarter of its period). The Hilbert transform is therefore also referred to as
the quadrature function. An alternative interpretation is that the Hilbert transform is proportional
to the time derivative of the process y(t). It will become apparent in section 4.2 that the physical
significance of the Hilbert transform yH (t) is that, together with y(t), it contains information about
the time-lagged behavior of the process y(t).
Analogous to the Fourier transform pair, the original process y(t) can be recovered from the
Hilbert transform yH (t) by an appropriate “inverse Hilbert transform.” The form of the inverse
Hilbert transform becomes apparent from the application of two Hilbert transforms in succession
of a process y(t). Define this double Hilbert transform to be yHH (t). From (4.1.6), the Fourier
transform of yHH (t) is

yHH (t) ≡ H [H [y(t)]] = H [yH (t)] ⊃ YHH (f ) = − i sgn(f ) YH (f )

(
−i YH (f ) if f ≥ 0
= (4.1.10a)
+i YH (f ) if f < 0.

Substituting for YH (f ) from (4.1.6) yields

 
YHH (f ) = − i sgn(f ) − i sgn(f ) Y (f ) = −Y (f ) . (4.1.10b)

The double Hilbert transform time series yHH (t) obtained by the inverse Fourier transform of
YHH (f ) is therefore the original time series multiplied by −1,

H [H [y(t)]] = − y(t) . (4.1.11a)

The original process y(t) is thus related to the Hilbert transform yH (t) by

y(t) = −H [H [y(t)]]

= −H [yH (t)] . (4.1.11b)

4–3

c D. B. Chelton June 26, 2014

The inverse Hilbert transform that recovers y(t) from yH (t) is therefore seen to be

H−1 [yH (t)] = y(t)

= −H [yH (t)]

1 ∞ yH (t0 ) 0
Z
= − dt (4.1.12a)
π −∞ t − t0

1 ∞ yH (t0 ) 0
Z
= dt . (4.1.12b)
π −∞ t0 − t

From (4.1.12a), it is apparent that the inverse Hilbert transform can be expressed as a convo-
lution analogous to (4.1.1),
H−1 [yH (t)] = h0 (t) ∗ yH (t) , (4.1.13)

where the inverse filter function h0 (t) is given by

1
h0 (t) = − . (4.1.14)
πt
Just as the signs of the complex exponentials in the Fourier transform pair can be switched (see
section 2.1 of chapter 1), the signs of the Hilbert transform pair (4.1.1) and (4.1.12) can be switched.
In other words, it is possible to switch the filter function definitions (4.1.2) and (4.1.14). The
convention above is adopted here, which is also used by Bendat and Piersol (1986). The reverse
convention is used by Bracewell (1986).
Some properties of the Hilbert transform that are easily deduced from (4.1.6) are:

i) Addition Theorem: The Hilbert transform of the sum of two processes is equal to the sum
of the individual Hilbert transforms of the two processes,

H [y(t) + z(t)] = yH (t) + zH (t) . (4.1.15)

This is easily seen from the convolution representation (4.1.1b) of the Hilbert transform
and the fact that convolution is distributive with respect to addition (see section 3.1.1 of
chapter 3).
ii) Multiplication Theorem: The Hilbert transform of a constant times y(t) is equal to the
same constant times yH (t),
H [a y(t)] = a yH (t) . (4.1.16)

This is trivially seen from the convolution representation (4.1.1b) of the Hilbert transform
and the scale preservation property of convolution (see section 3.1.4 of chapter 3).
iii) Shift Theorem: The Hilbert transform of y(t − a) is yH (t − a),

H [y(t − a)] = yH (t − a) . (4.1.17)

This is easily proven by substitution of t00 = t0 − a in the definition (4.1.1a).

4–4

c D. B. Chelton June 26, 2014

iv) Scale Theorem: The Hilbert transform of y(at) is yH (at),

H [y(a t)] = yH (a t) . (4.1.18)

This is easily proven by substitution of t00 = at0 in the definition (4.1.1a).

v) Even and Odd Function Theorem: The Hilbert transform of an even process is odd and
the Hilbert transform of an odd process is even. These properties are easily derived by
expressing the Fourier integral representation of yH (t) in terms of the sum of cosine and
sine components.
vi) Energy is preserved in the Hilbert transform time series. This can be proven by noting
from (4.1.6) that the complex conjugate of the Fourier transform YH (f ) of the Hilbert
transform is
+i Y ∗ (f ) if f ≥ 0
(
∗ ∗
YH (f ) = +i sgn(f ) Y (f ) = (4.1.19)
−i Y ∗ (f ) if f < 0
Combining (4.1.19) and (4.1.6) yields

YH∗ (f )YH (f ) = Y ∗ (f )Y (f ) for all f .

Rayleigh’s Theorem for the energy of y(t), i.e., for the integral of the product of y ∗ (t) and
y(t) (see section 1.4.2 of chapter 1), can therefore be expressed as
Z ∞ Z ∞
y ∗ (t)y(t) dt = Y ∗ (f )Y (f ) df
−∞ −∞
Z ∞
= YH∗ (f )YH (f ) df . (4.1.20a)
−∞

But Rayleigh’s Theorem for the energy of the Hilbert transform time series yH (t) is
Z ∞ Z ∞
∗
yH (t) yH (t) dt = YH∗ (f )YH (f ) df . (4.1.20b)
−∞ −∞

It is apparent from the equivalence of the right sides of (4.1.20a) and (4.1.20b) that the
energy of yH (t) is that same as that of y(t).
vii) Real Function Theorem: The Hilbert transform of a real process is also a real process.
This can be proven by applying the Real Function Theorem for Fourier transforms (see
section 2.2.7 of chapter 2), which states that, if y(t) is real, then its Fourier transform
Y (f ) is Hermitian, i.e.,

y ∗ (t) = y(t) ⇐⇒ Y ∗ (f ) = Y (−f ) . (4.1.21)

Substitution of (4.1.21) into (4.1.19) thus yields

(
∗
+i Y (−f ) if f ≥ 0
YH (f ) =
−i Y (−f ) if f < 0

= YH (−f ) .

4–5

c D. B. Chelton June 26, 2014

The Fourier transform of yH (t) is thus also Hermitian, which implies that yH (t) is real.

viii) Orthogonality Property: The integral of the product of a process y(t) and the complex
conjugate of its Hilbert transform is zero,
Z ∞
∗
y(t)yH (t) dt = 0 . (4.1.22)
−∞

This can be proven from the Generalized Rayleigh’s Theorem for the integral of the prod-
∗
uct of y(t) and yH (t) (see section 2.2.10 of chapter 2), which yields
Z ∞ Z ∞
∗
y(t)yH (t) dt = Y (f )YH∗ (f ) df
−∞ −∞
Z ∞
= i sgn(f ) Y (f )Y ∗ (f ) df , (4.1.23)
−∞

where (4.1.19) has been used to express YH∗ (f ) in terms of Y ∗ (f ). Since the product
Y (f )Y ∗ (f ) is an even function of frequency and sgn(f ) is an odd function of frequency,
the right side of (4.1.23) is an odd function of frequency. The integral is therefore zero,
which yields (4.1.22).

4.2. The Hilbert Transform of Periodic Processes

Additional Reading:
Bendat and Piersol (1986), pp. 488–489

4.2.1. The Hilbert Transform of Monochromatic Sinusoidal Processes

As a simple example of application of the Hilbert transform, consider the even function con-
sisting of a monochromatic cosine with unit amplitude and positive frequency f0 ,

y(t) = cos (2πf0 t) . (4.2.1)

Recall from section 2.3.4 of chapter 2 that the Fourier transform of the cosine process (4.2.1) is the
symmetric function
 
1
Y (f ) = F [cos(2πf0 t)] = δ(f + f0 ) + δ(f − f0 ) , (4.2.2)
2

i.e., a pair of Dirac delta functions centered at f = ±f0 . Since f0 is positive, the Fourier transform
(4.1.6) of the Hilbert transform of y(t) is
 
i
YH (f ) = δ(f + f0 ) − δ(f − f0 ) = F [sin(2πf0 t)] . (4.2.3)
2

4–6

c D. B. Chelton June 26, 2014

The right side of (4.2.3) is the Fourier transform of sin(2πf0 t) (see section 2.3.4 of chapter 2). The
Hilbert transform of y(t) obtained from the inverse Fourier transform of YH (f ) is thus the odd
function

yH (t) = sin (2πf0 t) (4.2.4a)

= cos (2πf0 t − π/2)

= cos (2πf0 [t − T0 /4])

= y(t − T0 /4) , (4.2.4b)

where T0 = f0−1 is the period of the monochromatic process y(t). The Hilbert transform pair (4.2.1)
and (4.2.4a) is shown in Figure 4.2.1.
Analogous to the derivation of (4.2.3), consider the monochromatic sine process with unit
amplitude and a positive frequency f0 ,

y(t) = sin (2πf0 t) . (4.2.5)

The Hilbert transform of this process is

yH (t) = − cos (2πf0 t) (4.2.6a)

= sin (2πf0 t − π/2)

= sin (2πf0 [t − T0 /4])

= y(t − T0 /4) . (4.2.6b)

The Hilbert transform pair (4.2.5) and (4.2.6a) is shown in Figure 4.2.2.
The results (4.2.4) and (4.2.6) for the Hilbert transforms of a cosine and a sine, both with
unit amplitude and zero phase, can be generalized to the case of a sinusoid with arbitrary phase.
Consider the sinusoidal process with unit amplitude and phase φ of the form

y(t) = cos(2πf0 t + φ) . (4.2.7a)

With the sum angle formula cos(α ± β) = cos α cos β ∓ sin α sin β, this is equivalent to

y(t) = A cos(2πf0 t) − B sin(2πf0 t) , (4.2.7b)

where

A = cos φ (4.2.8a)

B = sin φ (4.2.8b)

4–7

c D. B. Chelton June 26, 2014

With the solutions (4.2.4a) and (4.2.6a) for the Hilbert transforms of the cosine and the sine and
the Addition and Multiplication Theorems (4.1.15) and (4.1.16), the Hilbert transform of (4.2.7b)
is
yH (t) = A sin(2πf0 t) + B cos(2πf0 t) . (4.2.9)

Substitution of (4.2.8a) and (4.2.8b) for the constants A and B and use of the sum angle formula
sin(α ± β) = sin α cos β ± cos α sin β transforms (4.2.9) into

yH (t) = sin(2πf0 t + φ) (4.2.10a)

= cos(2πf0 t + φ − π/2)

= cos(2πf0 [t − T0 /4])

= y(t − T0 /4) . (4.2.10b)

The result (4.2.10a) can also be obtained directly from (4.2.7a) by the Shift Theorem (4.1.17).
The Hilbert transform of any monochromatic sinusoid with positive frequency f0 thus consists
of the original sinusoid lagged in time by one quarter of the period T0 = f0−1 of the monochromatic
process. As shown by (4.1.9) for the case of a general process y(t), these simple examples of
monochromatic processes illustrate the fact that the Hilbert transform is equivalent to filtering the
positive frequencies of the process by a quadrature phase shift of −π/2.

4.2.2. The Hilbert Transform of an Arbitrary Periodic Process

The Hilbert transforms (4.2.4a) and (4.2.6a) of monochromatic cosine and sine functions with
positive frequency can be used to derive an expression for the Hilbert transform of an arbitrary
periodic process with period T . Recall from section 1.5.1 of chapter 1 that such a time series can
be written as a Fourier series expansion of the form
∞
X
y(t) = Aj cos(2πfj t) + Bj sin(2πfj t) , (4.2.11)
j=0

where the frequencies fj = j/T are the Fourier frequencies, which are all non-negative in the one-
sided sum in (4.2.11). Each cosine and sine in the Fourier series expansion (4.2.11) has a Hilbert
transform given by (4.2.4a) and (4.2.6a), respectively, except with f0 replaced with each of the
Fourier frequencies fj . By the Addition and Multiplication Theorems (4.1.15) and (4.1.16), the
Hilbert transform of the arbitrary periodic process is thus
∞
X
yH (t) = −Bj cos(2πfj t) + Aj sin(2πfj t) . (4.2.12)
j=0

From the trigonometric identities

sin(2πfj t) = cos(2πfj t − π/2)

− cos(2πfj t) = sin(2πfj t − π/2)

4–8

c D. B. Chelton June 26, 2014

and inspection of (4.2.11) and (4.2.12), it is apparent once again that the Hilbert transform yH (t)
is just the original process y(t) with each positive frequency component fj shifted −π/2 in phase
but unaltered in amplitude.

4.3. The Analytic Signal

Additional Reading:
Bracewell (1986), pp. 268–271
Bendat and Piersol (1986), pp. 486–487

A useful feature of the Hilbert transform becomes apparent by considering the complex process
z(t) constructed from a process y(t) and its Hilbert transform yH (t) by

z(t) = y(t) + i yH (t) . (4.3.1)

The process z(t) is called the analytic signal associated with y(t). The Fourier transform of z(t) is
easily seen from (4.1.6) to be (
2Y (f ) if f ≥ 0
Z(f ) = (4.3.2)
0 if f < 0
The process z(t) is thus equivalent to suppressing the negative frequency components of y(t) and
amplifying the non-negative frequency components by a factor of two.
It was shown in section 4.2 that the Hilbert transform yH (t) is the same as the original
process y(t) except with each positive frequency component shifted −π/2 in phase but unaltered
in amplitude. The analytic signal (4.3.1) thus contains both co- and quadrature information about
the process y(t). It is therefore well suited to describing traveling-wave signals y(x, t) as a function
of spatial location x and time t. This is illustrated in example 1 below. Before determining the
analytic signal for a traveling wave signal, however, consider first the simple cosine process (4.2.1)
for which the Hilbert transform is (4.2.4a). The analytic signal associated with the cosine with unit
amplitude and zero phase is thus

z(t) = cos(2πf0 t) + i sin(2πf0 t) = ei2πf0 t . (4.3.3)

| {z } | {z }
y(t) yH (t)

The last equality follows from Euler’s formula in section 1.3.1 of chapter 1.
The analytic signal associated with the simple sine process (4.2.5) with unit amplitude and
zero phase is determined similarly from its Hilbert transform (4.2.6a), which gives
 
z(t) = sin(2πf0 t) + i − cos(2πf0 t)
| {z } | {z }
y(t) yH (t)
= −i [i sin(2πf0 t) + cos(2πf0 t)]

= −i ei2πf0 t . (4.3.4a)

4–9

c D. B. Chelton June 26, 2014

Expressing −i in polar coordinate as −i = e−π/2 (see section 1.2.1 of chapter 1) transforms the
expression (4.3.4a) for the analytic signal associated with the sine process into

z(t) = ei2πf0 t−π/2 . (4.3.4b)

For the the general sinusoid (4.2.7a) with arbitrary phase φ for which the Hilbert transform is
(4.2.10a), the analytic signal is

z(t) = cos(2πf0 t + φ) + i sin(2πf0 t + φ) = ei2πf0 t+φ . (4.3.5)

| {z } | {z }
y(t) yH (t)

Thus, while the Euler representation of a sinusoid consists of the sum of two complex exponentials,
the analytic signal associated with a sinusoid consists of only one complex exponential with a phase
shift φ. The phase shift is φ = 0 for a cosine and φ = −π/2 for a sine.
The analytic signal associated with an arbitrary periodic process y(t) with period T is easily
derived by noting that each cosine and sine in the Fourier series decomposition (4.2.11) has an
analytic signal of the form (4.3.3) and (4.3.4a), respectively, except with f0 replaced with each of
the Fourier frequencies fj = j/T, j = 0, 1, . . . , ∞. The analytic signal associated with the periodic
process is therefore the sum of the analytic signals associated with each cosine and sine, i.e.,
∞
X
z(t) = Aj ei2πfj t − i Bj ei2πfj t
j=0

∞
X
Aj − i Bj ei2πfj t .


= (4.3.6)
j=0

As shown from the following two examples, the analytic signal can describe both standing-
and traveling-wave processes. It is especially useful for describing the latter (see Example 1).

Example 1: The Analytic Signal Associated with a Traveling-Wave Process

As an example application of the analytic signal, consider a monochromatic, traveling-
wave, real process y(x, t) in one spatial dimension (see section 1.3.2 of chapter 1) with
wavenumber k0 and frequency f0 , which can be represented as

y(x, t) = A(x, t) cos [2π(k0 x − f0 t)] . (4.3.7a)

Initially, the case of constant amplitude A will be considered. The traveling wave of the
form (4.3.7a) can be written as

y(x, t) = A cos [Φ(x, t)] ,

where
Φ(x, t) = 2πk0 (x − ct)

4 – 10

c D. B. Chelton June 26, 2014

is the phase of cosine and

c = +f0 /k (4.3.8)
is the speed of constant phase, i.e., the phase speed of the wave. The wave therefore
propagates in the positive x direction if f0 and k0 both have the same sign and in the
negative x direction if f0 and k0 have opposite signs. If f0 < 0, (4.3.7a) can be written as
y(x, t) = A cos [2π(k0 x + |f0 |t)]

= A cos [−2π(−k0 x − |f0 |t)]

= A cos [2π(k00 x − |f0 |t)] .

This expression has the same form as (4.3.7a), except that k0 is replaced by k00 = −k0 .
Without loss of generality, the frequency f0 in (4.3.7a) can therefore be considered pos-
itive and the wavenumber can be either positive or negative. The direction of the wave
propagation then depends only on the sign of the wavenumber.
The Hilbert transform of the monochromatic traveling wave is easily determined
from the Hilbert transforms (4.2.4a) and (4.2.6a) of the cosine and sine by following the
same procedure used in section 4.2. After applying the sum angle formula cos(α ± β) =
cos α cos β ∓ sin α sin β, the traveling-wave cosine function (4.3.7a) can be written as
y(x, t) = a cos(2πf0 t) + b sin(2πf0 t) , (4.3.7b)
where the amplitudes of the cosine and sine functions are
a = A cos(2πk0 x)

b = A sin(2πk0 x) .
The amplitude A(x, t) can, in general, vary spatially and temporally. The amplitudes
a and b can therefore also vary spatially as well as temporally. Since f0 ≥ 0 with the
convention adopted above, the Hilbert transform of (4.3.7b) is obtained by substituting
(4.2.4a) and (4.2.6a) for the Hilbert transforms of the cosine and sine functions for positive
frequencies into (4.3.7b), which yields
yH (x, t) = − b cos(2πf0 t) + a sin(2πf0 t) . (4.3.9a)
Substituting for a and b from the above expressions and using the sum angle formula
sin(α ± β) = sin α cos β ± cos α sin β in reverse transforms this expression into
 
yH (x, t) = A sin 2π(f0 t − k0 x)
 
= −A sin 2π(k0 x − f0 t) (4.3.9b)
 
= A cos 2π(k0 x − f0 t) + π/2
 
= A cos 2π(k0 x − f0 t + 1/4)


= A cos 2π k0 x − f0 [t − T0 /4] (4.3.9c)

= y(t − T0 /4) ,

4 – 11

c D. B. Chelton June 26, 2014

where T0 = f0−1 is again the period of the traveling wave signal. As for the cases of the
simple sinusoids considered in section 4.2, the Hilbert transform of a sinusoidal traveling
wave is thus the same sinusoidal traveling wave, except lagged in time by one quarter
of the period T0 of the monochromatic process. This is just another illustration of the
general quadrature relation (4.1.9) for the Hilbert transform.
The analytic signal associated with the traveling-wave process is obtained by substi-
tuting (4.3.7) and (4.3.9b) into the definition (4.3.1) of the analytic signal and again using
Euler’s formula in section 1.3.1 of chapter 1 to obtain

z(x, t) = A cos [2π(k0 x − f0 t)] − i A sin [2π(k0 x − f0 t)]

= A e−i2π(k0 x−f0 t) . (4.3.10)

The analytic signal z(x, t) is thus a complex representation of the monochromatic traveling
wave. From the definition (4.3.1) of the analytic signal, it is clear that the traveling-wave,
real process y(x, t) defined by (4.3.7) is obtained by simply taking the real part of (4.3.10).
For the application to the traveling wave signal (4.3.7) considered in this example,
the analytic signal (4.3.10) expressed in polar coordinates as in (4.3.5) has the form

z(x, t) = A eiφx (x) eiφt (t) , (4.3.11)

where the spatial and temporal phases are

φx (x) ≡ − 2πk0 x (4.3.12a)

φt (t) ≡ 2πf0 t . (4.3.12b)

Since k0 and f0 are constant for the monochromatic traveling wave under consideration
here, φx (x) and φt (t) vary linearly with x and t, respectively. The wavenumber k0 and
frequency f0 can thus be determined from the spatial and temporal phases of the analytic
signal z(x, t) by

1 dφx
k0 = − (4.3.13a)
2π dx
1 dφt
f0 = . (4.3.13b)
2π dt

The phase speed (4.3.8) of the traveling-wave signal (4.3.7) is thus

f0 dφt /dt
c = = − . (4.3.14)
k0 dφx /dx

The phase propagation is in the positive x direction if the slopes of the linear variations of
spatial phase φx and temporal phase φt have opposite signs and in the negative x direction
if the two slopes have the same sign.

4 – 12

c D. B. Chelton June 26, 2014

The amplitudes and phases of the analytic signal (4.3.11) associated with an example
traveling wave of the form (4.3.7) with spatially and temporally constant amplitude A = 1
are shown in Figure 4.3.1 for positive frequency f0 and positive wavenumber k0 . The
unit amplitude can be inferred from the product of the spatially and temporally constant
amplitudes Ax (x) and At (t) in the upper left and upper right panels, respectively. For this
example, the wavenumber calculated by (4.3.13a) from the (negative) slope of the spatial
phase φx (x) in the lower left panel is k = 10−2 m−1 , which corresponds to a wavelength
of λ = k0−1 = 100 m. The frequency calculated by (4.3.13b) from the (positive) slope of
the temporal phase φt (t) in the lower right panel is f0 = 0.05 sec−1 , which corresponds
to a period of T = f0−1 = 20 sec. The opposite signs of the slopes of both linearly varying
phases indicate that the wave propagates in the positive x direction. The phase speed
(4.3.14) of the propagating wave is c = 5 m s−1 .
The above formalism can easily be modified for the case of a traveling wave with
spatially inhomogeneous and temporally nonstationary amplitude A = A(x, t) when the
spatial and temporal modulations of the wave amplitude are separable so that A can be
expressed as
A(x, t) = Ax (x) At (t) . (4.3.15)

Then the analytic signal (4.3.11) becomes

z(x, t) = Ax (x) eiφx (x) At (t) eiφt (t) . (4.3.16)

The amplitudes and phases of the analytic signal associated with an example traveling
wave of the form (4.3.7) with spatially and temporally varying amplitude (4.3.15) are
shown in Figure 4.3.2 for the case of a spatial modulation Ax (x) that increases linearly
with increasing x and a temporal modulation At (t) that decreases linearly with increasing
t. The amplitude A(x, t) of the wave at x = 0, for example, decreases from 2 m to 1 m
over the 90-sec period of the data record shown. During the same 90-sec time period, the
amplitude A(x, t) at x = 300 m decreases from 4 m to 2 m. The wavenumber, frequency
and phase speed calculated by (4.3.13a), (4.3.13b) and (4.3.14), respectively, are the same
as for the example of a traveling wave with constant amplitude A shown in Figure 4.3.1.

Example 2 : The Analytic Signal Associated with a Standing-Wave Process

As a second application of the analytic signal, consider a monochromatic, standing-
wave, real process y(x, t) in one spatial dimension (see section 1.3.2 of chapter 1) with
frequency f0 ,
y(x, t) = a(x, t) cos(2πf0 t) , (4.3.17)

where the amplitude a(x, t) of the standing wave can in general vary spatially and tem-
porally. Nodes of the standing wave occur at locations x where a(x, t) = 0 at all times t.
As in Example 1 above, it will be assumed that f0 ≥ 0. The example considered here is
a spatial modulation in the form of a sinusoidal standing wave with wavenumber k0 ,

a(x, t) ≡ A(x, t) cos(2πk0 x) (4.3.18)

4 – 13

c D. B. Chelton June 26, 2014

with amplitude A(x, t) that may vary spatially and temporally.

From (4.2.4a), the Hilbert transform of the standing-wave process (4.3.17) with pos-
itive frequency f0 is
yH (t) = a(x, t) sin(2πf0 t) . (4.3.19)

The analytic signal associated with the standing-wave process is then determined by sub-
stituting (4.3.17) and (4.3.19) into the definition (4.3.1) of the analytic signal to obtain
 
z(x, t) = a(x, t) cos(2πf0 t) + i sin(2πf0 t)

= a(x, t) ei2πf0 t . (4.3.20)

Although somewhat cumbersome, the analytic signal (4.3.20) associated with the
standing-wave process defined by (4.3.17) and (4.3.18) can be interpreted in terms of
spatial and temporal phase in the same manner as the analytic signal associated with
the traveling-wave process considered in Example 1 above. The amplitude (4.3.18) of the
spatially sinusoidal standing wave can be expressed as

a(x, t) = A0 (x, t) sgn [cos(2πk0 x)] , (4.3.21a)

where
A0 (x, t) ≡ A(x, t) cos(2πk0 x)
 
(4.3.21b)

is a spatially and temporally varying non-negative amplitude and the sign of a(x, t) is
accounted for by the function sgn [cos(2πk0 x)] defined as in (4.1.5b),
(
+1 if cos(2πk0 x) ≥ 0
sgn [cos(2πk0 x)] = (4.3.21c)
−1 if cos(2πk0 x) < 0.

Expressing ±1 in polar coordinates as

+1 = e0
−1 = eπ

(see section 1.2.1 of chapter 1) transforms (4.3.21a) into

a(x, t) = A0 (x, t) eiφx (x) , (4.3.22a)

where (
0 if cos(2πk0 x) ≥ 0
φx (x) = (4.3.22b)
π if cos(2πk0 x) < 0.
The analytic signal (4.3.20) with amplitude given by (4.3.22a,b) then becomes

z(x, t) = A0 (x, t) eiφx (x) eiφt (t) , (4.3.23)

4 – 14

c D. B. Chelton June 26, 2014

where the temporal phase φt (t) is defined as

φt (t) = 2πf0 t . (4.3.24)

The temporal phase thus increases linearly in time t with a slope of dφt /dt = 2πf0 . The
(positive) frequency of the standing-wave oscillations is therefore

1 dφt
f0 = . (4.3.25)
2π dt
As defined by (4.3.21b), the non-negative amplitude A0 (x, t) of the analytic signal
consists of the amplitude A(x, t) modulated by the absolute value of the sinusoidal spatial
variation | cos(2πk0 x)| with wavenumber k0 and thus has zeros at the nodes and maxima
at the anti-nodes of the standing wave. There are two nodes and two anti-nodes per period
of the spatially varying amplitude. The spatial phase (4.3.22b) for the standing wave is
a square wave that transitions abruptly between values of 0 and π at the nodal points
where the amplitude of the standing wave vanishes. The sinusoidal spatial variability of
the amplitude of the standing wave is fully described by (4.3.22a) with A0 (x, t) defined by
(4.3.21b) and φx (x) defined by (4.3.22b).
The amplitudes and phases of the analytic signal associated with an example of a
spatially sinusoidal standing wave of the form (4.3.17) with amplitude a(x, t) given by
(4.3.18) for constant amplitude A = 1 are shown in Figure 4.3.3. The non-negative
amplitude A0 (x, t) defined by (4.3.21b) can then be partitioned as A0 (x, t) = A0x (x)A0t (t),
where A0x (x) = |cos(2πk0 x)| and A0t (t) = 1. The spatially varying amplitude A0x (x) in
the upper left panel has extrema values of max A0x (x) = 1 and the temporal amplitude
 

in the upper right panel has a constant value of A0t (t) = 1. The constant unit amplitude
A of the standing wave can be inferred from the product max A0x (x) A0t (t) = 1. The
 

distance between the nodes of the sinusoidal standing wave, i.e., the distance separating
the locations x where A0x (x) = 0 in the upper left panel, and where the spatial phase
φx (x) in the lower left panel transitions abruptly between 0 and π, is seen to be 50 m.
The wavelength is therefore 100 m. The frequency f0 calculated from the (positive) slope
of the linearly varying temporal phase φt (t) in the lower right panel is f0 = 0.05 cycle per
sec, corresponding to a period of period T = f0−1 = 20 sec.
As in the case of the traveling wave considered in Example 1 above, the preceding
formalism for the analytic signal associated with a standing wave can easily be modified
for the case for which the amplitude A(x, t) is spatially inhomogeneous and temporally
nonstationary. When the spatial and temporal modulations are separable so that A(x, t)
can again be expressed as in (4.3.15) as A(x, t) = Ax (x) At (t), the analytic signal (4.3.23)
becomes
z(x, t) = A0x (x) eiφx (x) A0t (t) eiφt (t) , (4.3.26)
where φx (x) and φt (t) are still given by (4.3.22b) and (4.3.24), respectively, and

A0x (x) = Ax (x) cos(2πk0 x)

 
(4.3.27a)

A0t (t) = At (t) (4.3.27b)

4 – 15

c D. B. Chelton June 26, 2014

is the total spatial modulation of the amplitude of the standing-wave process.

The amplitudes and phases of the analytic signal associated with an example of
a spatially sinusoidal standing wave of the form (4.3.17) with spatially and temporally
varying amplitude are shown in Figure 4.3.4 for the case of a spatial modulation Ax (x)
that increases linearly with increasing x and a temporal modulation At (t) that decreases
linearly with increasing t. The amplitude A(x, t) of the wave at the anti-node at x = 0,
for example, decreases from 2 m to 1 m over the 90-sec period of the data record shown.
During the same 90-sec time period, the amplitude A(x, t) at the anti-node x = 300 m
decreases from 4 m to 2 m. The wavenumber determined from the nodes of the amplitude
A0x (x) and the frequency f0 calculated by (4.3.25) are the same as for the example of a
standing wave with constant amplitude Ax (x) = 1 shown in Figure 4.3.3.

4.4. Practical Considerations

Additional Reading:
Bendat and Piersol (1986), p. 488
von Storch and Zwiers (1999), Section 16.2.4
Barnett, T. P., 1983: Interaction of the monsoon and Pacific trade wind system at interannual
time scales. Part 1: The equatorial zone. Mon. Wea. Rev., 111, 756–773.
Horel, J. D., 1984: Complex principal component analysis: Theory and examples. Climate Appl.
Meteor., 23, 1660–1673.
Marple, S. L., 1999: Computing the discrete-time “analytic” signal via FFT. IEEE Trans. Signal
Proc., 47, 2600–2603.
Merrifield, M. A., and R. T. Guza, 1990: Detecting propagating signals with complex empirical
orthogonal functions: a cautionary note. J. Phys. Oceangr., 20, 1628–1633.
Oppenheimer, A. V., and R. W. Schafer, 1975: Digital Signal Processing. Prentice Hall, 360 pp.

White, W. B., S. E. Pazan and M. Inoue, 1987: Hindcase/Forecast of ENSO events based upon
the redistribution of observed and model heat content in the western tropical Pacific, 1964–86.
J. Phys. Oceanogr., 17, 264–280.

The preceding formalism for the Hilbert transform is for a continuous, infinite process
y(t). Practical applications are limited to discrete samples y(tn ) at N observation times
tn = (n − 1)∆t, n = 1, . . . , N of the continuous process y(t) at sample interval ∆t over a finite
record length T = (N − 1)∆t. The exact Hilbert transform (4.1.1) must therefore be estimated
from the sample record. Since the Hilbert transform can be equivalently represented in terms of
Fourier transforms according to (4.1.4), any sample estimate of yH (t) constructed from a finite
number of discrete samples is influenced by the same effects of aliasing, leakage and smearing of the
frequency content of y(t) that affect discrete, finite record length estimates of the Fourier transform.
These effects are discussed in detail later in chapter 5 in the context of sample estimates of the
Fourier transform.

4 – 16

c D. B. Chelton June 26, 2014

In applications to real data, two approaches can be used. The most straightforward approach
(see, for example, Barnett, 1983; Horel, 1984; Merrifield and Guza, 1990; Marple, 1999) is to
represent the discretely sampled time series over a finite record length T by a discrete Fourier
transform. It is shown later in chapter 6 that this is equivalent to truncating the infinite Fourier
series representation of a periodic time series with period T considered in section 4.2.2. An estimate
of the Hilbert transform can then be constructed from the amplitudes of the sine and cosine terms
in the discrete Fourier transform using, respectively, the relations (4.2.6a) and (4.2.4a) with f0
replaced with fj for each of the N Fourier frequencies fj = j/T, j = 0, 1, . . . , N − 1.
An alternative approach (see, for example, Barnett, 1983; White et al., 1987) is to approximate
the infinite, continuous convolution (4.1.1) by a finite, discrete convolution of the form

M
X
ŷH (tn ) = y(tn − m∆t) ĥ(m∆t) , (4.4.1)
m=−M

where ∆t is the discrete sample interval, M ∆t is the half span (to be specified) of the convolution
filter and ĥ(m∆t) are the weights of a simple filter that has the desired properties of the true filter
function (4.1.2) and (4.1.8), i.e., the amplitude response (4.1.6) of unity with phase shifts of −π/2
for positive frequencies and +π/2 for negative frequencies. Oppenheimer and Shafer (1975) suggest
the discrete filter function

 2 sin2 (πm∆t/2) if m 6= 0


ĥ(m∆t) = π m (4.4.2)

0 if m = 0.
When M is large, the discrete convolution (4.4.1) very nearly approximates the true convo-
lution. In practice, however, M is limited by the record length available. The Hilbert transform
time series estimated by (4.4.1) often exhibits severe “ripple” contamination within M ∆t of the
beginning and end of the data record. The ripple contamination in the remaining portion of the
data record is usually less severe, although this depends on the detailed temporal characteristics of
the process y(t), the sample interval ∆t, and the choice of the half-span M ∆t. In practice, the half
span M ∆t can be selected by initially trying a small value of about M = 5 and iteratively increasing
M until there is little detectable difference in the estimated Hilbert transform in the middle of the
data record where there are no end effects. A value of M ≤ 15 usually yields acceptable results.
The data points within M ∆t of the beginning and end of the data record should be omitted from
subsequent analysis.
When there is energetic low-frequency variability in the finite record length time series y(t),
the discrete Hilbert transform obtained by either of the above two methods often exhibits ripples
throughout the time series. This is especially the case when there is a large offset between the
beginning and end of the data record, as often happens when a time series is dominated by unre-
solved low-frequency variability. It is then helpful to detrend the time series prior to computing the
discrete Hilbert transform by either of the above two methods. This can be achieved by high-pass
filtering the time series or by removing the linear trend defined by the first and last points in the
data record. The detrimental effects of an offset between the beginning and end of a data record
(called “spectral leakage”) are discussed extensively later in chapter 5.

4 – 17
Figure 4.1.1. The gain GH (f ) and phase φH (f ) of the filter transfer function of the Hilbert
transform.
Figure 4.2.1. A cosine process y(t) = cos(2πf0 t) and its Hilbert transform process yH (t) =
sin(2πf0 t) for a positive frequency f0 .
Figure 4.2.2. A sine process y(t) = sin(2πf0 t) and its Hilbert transform process yH (t) =
− cos(2πf0 t) for a positive frequency f0 .
Figure 4.3.1. The spatial amplitude Ax (x) and phase φx (x) and the temporal amplitude At (x)
and phase φt (t) of the analytic signal associated with a traveling wave with 100-m wavelength,
20-sec period and spatially and temporally constant amplitude A = Ax (x)At (t).
Figure 4.3.2. The same as Figure 4.3.1, except for the case of amplitude A(x, t) = Ax (x)At (t)
with spatial component Ax (x) that increases linearly with increasing x and temporal component
At (t) that decreases linearly with increasing t.
Figure 4.3.3. The spatially varying amplitude Ax (x) and phase φx (x) and the temporal amplitude
At (x) and phase φt (t) of the analytic signal associated with a spatially sinusoidal standing wave
with 100-m wavelength, 20 sec period and spatially and temporally constant amplitude A.
Figure 4.3.4. The same as Figure 4.3.3, except for the case of amplitude A(x, t) = Ax (x)At (t)
with spatial component Ax (x) that increases linearly with increasing x and temporal component
At (t) that decreases linearly with increasing t.