Appendix C

Useful theorems for the Hilbert transform

C.1 Real part sufficiency

Given the FRF for a causal system, equations (4.17) and (4.18) show that the real
part can be used to reconstruct the imaginary part and vice versa. It therefore
follows that all the system characteristics are encoded in each part separately.
Thus, it should be possible to arrive at the impulse response using the real part or
imaginary part of the FRF alone.
From (4.4)
g(t) = geven(t) + godd(t) (C.1)
and from (4.7)
g(t) = geven(t)(1 + (t)) (C.2)
or
g(t) = geven(t)  2(t) (C.3)
where (t) is the Heaviside unit-step function. Finally

g(t) = 2F 1 fRe G(!)g(t) (C.4)

which shows that the real part alone of the FRF is sufficient to form the impulse
response. A similar calculation gives

g(t) = 2F 1 fIm G(!)g(t): (C.5)

C.2 Energy conservation

The object of this exercise is to determine how the total energy of the system, as
encoded in the FRF, is affected by Hilbert transformation. If the energy functional
is defined as usual by Z1
d Æ jf (Æ)j2 : (C.6)
1
546
 Commutation with differentiation 547

Then, Parseval’s theorem

Z 1 Z 1
dt jg(t)j2 = d! jG(!)j2 (C.7)
1 1
where G(! ) = Ffg (t)g, shows that energy is conserved under the Fourier
transform.
Taking the Hilbert transform of G(! ) yields

HfG(!)g = G~ (!) (C.8)

and an application of Parseval’s theorem gives

Z 1 Z 1
d! jG~ (!)j2 = dt jF 1 fG~ (!)gj2 : (C.9)
1 1
By the definition of the Hilbert transform
1
G~ (!) = G(!)  (C.10)
!
and taking the inverse Fourier transform yields

F 1 fG~ (!)g = g(t)(t) (C.11)

so
jF 1 fG~ (!)gj2 = jg(t)(t)j2 = jg(t)j2 (C.12)
as j(t)j2 = 1. Substituting this result into (4.26) and applying Parseval’s theorem
once more gives
Z 1 Z 1 Z 1
d! jG~ (!)j2 = dt jg(t)j2 = d! jG(!)j2 : (C.13)
1 1 1
Thus, energy is also conserved under Hilbert transformation.

C.3 Commutation with differentiation

Given a function G(! ), it can be shown that the Hilbert transform operator H
commutes with the derivative operator d=d! under fairly general conditions, i.e.
 
dG dG~
H d!
=
d!
(C.14)

Consider

1 1 1 1
 Z  Z  
dG~ d G( ) d G( )
= d = d (C.15)
d! d! i 1 ! i 1 d! !
548 Useful theorems for the Hilbert transform

assuming differentiation and integration commute. Elementary differentiation

yields
1 1
Z
dG~ G( )
= d :
! )2
(C.16)
d! i 1 (
Now 
dG

1 1
Z dG
H = d d (C.17)
d! i 1 !
and integrating by parts yields
dG  =1
1 1 1 1
Z  Z  
d d = 1 G( ) + d G( )
d 1
:
i 1 ! i ! = 1 i 1 d !
(C.18)
Now, assuming that the first term (the boundary term) vanishes, i.e. G(! ) has fast
enough fall-off with ! (it transpires that this is a vital assumption in deriving the
Hilbert transform relations anyway—see chapter 5), simple differentiation shows

1 1
  Z
dG G( )
H d!
= d
i 1 ( ! )2
(C.19)

and together with (C.16), this establishes the desired result (C.14).
An identical argument in the time domain suffices to prove
 
dg(t) d~g
H dt
=
dt
(C.20)

with an appropriate time-domain definition of H.

Having established the Fourier decompositions (4.79) and (4.84), it is
possible to establish three more basic theorems of the Hilbert transform.

C.4 Orthogonality
Considered as objects in a vector space, it can be shown that the scalar product of
an FRF or spectrum G(! ) with its associated Hilbert transform G ~ (!) vanishes,
i.e. Z 1
hG; G~ i = d! G(!)G~ (!) = 0: (C.21)
1
Consider the integral, using the Fourier representation (4.79), one has
H Æ F = FÆ 2, so
Z 1 Z 1 Z 1 Z 1
d! G(!)G~ (!) = d! dt e i!tg(t) d e i! ( )g( ):
1 1 1 1
(C.22)
 Action as a filter 549

A little rearrangement yields

Z 1Z 1 Z 1 
dtd d! e i!(t+ ) ( )g(t)g( ) (C.23)
1 1 1
the bracketed expression is a Æ -function 2Æ (t + ) (appendix D). Using the
projection property, the expression becomes
Z 1 Z 1
2 dt ( t)g(t)g( t) = 2 dt (t)g(t)g( t): (C.24)
1 1
The integrand is clearly odd, so the integral vanishes. This establishes the desired
result (C.21). An almost identical proof suffices to establish the time-domain
orthogonality:
Z 1
hg; g~i = dt g(t)~g(t) = 0: (C.25)
1

C.5 Action as a filter

The action of the Hilbert transform on time functions factors as

H = F 1Æ 2 ÆF (C.26)

as derived in chapter 4. This means that, following the arguments of [289], it can
be interpreted as a filter with FRF

H (!) = (!) (C.27)

i.e. all negative frequency components remain unchanged, but the positive
frequency components suffer a sign change. (It is immediately obvious now
why the Hilbert transform exchanges sines and cosines. Energy conservation also
follows trivially.) Each harmonic component of the original signal is shifted in
phase by =2 radians and multiplied by i 1 .
Now consider the action on a sine wave

Hfsin( t)g = F 1 Æ 2 ÆFfsin( t)g

 
= F 1 Æ 2 1 (Æ(! ) Æ(! + ))
2i
 
1
= F 1 (Æ (! ) + Æ(! + ))
2i
= i cos( t) (C.28)
1 With the traditional time-domain definition of the Hilbert transform, the filter action is to phase shift
all frequency components by = . 2
550 Useful theorems for the Hilbert transform

a result wich could have been obtained by phase-shifting by =2 and multiplying
by i. The same operation on the cosine wave yields,

Hfcos( t)g = i sin( t): (C.29)

Now suppose the functions are premultiplied by an exponential decay with a time
constant long compared to their period 2= . Relations similar to (C.28) and
(C.29) will hold:

Hfe t sin( t)g  ie t cos( t) (C.30)

Hfe t cos( t)g  ie t sin( t) (C.31)

for sufficiently small . This establishes the result used in section 4.7.
The results (C.28) and (C.29) hold only for > 0. If < 0, derivations of
the type given for (C.28) show that the signs are inverted. It therefore follows that

Hfsin( t)g = i( ) cos( t) (C.32)

Hfcos( t)g = i( ) sin( t): (C.33)

These results are trivially combined to yield

Hfei tg = ( )ei t : (C.34)

C.6 Low-pass transparency

Consider the Hilbert transform of a time-domain product m(t)n(t) where the
spectra M (! ) and N (! ) do not overlap and M is low-pass and N is high-pass,
then
Hfm(t)n(t)g = m(t) Hfn(t)g (C.35)
i.e. the Hilbert transform passes through the low-pass function. The proof given
here follows that of [289].
Using the spectral representations of the function, one has
Z 1Z 1
m(t)n(t) = d! d M (!)N ( )e i(!+ )t : (C.36)
1 1
Applying the Hilbert transform yields
Z 1Z 1
Hfm(t)n(t)g = d! d M (!)N ( )Hfe i(!+ )tg (C.37)
1 1
and by (C.34)
Z 1Z 1
Hfm(t)n(t)g = d! d M (!)N ( )(! + )e i(!+ )t g: (C.38)
1 1
 Low-pass transparency 551

Now, under the assumptions of the theorem, there exists a cut-off W , such that
M (!) = 0 for ! > W and N ( ) = 0 for < W . Under these conditions, the
signum function reduces (! + ) = ( ) and integral (C.38) factors:
Z 1  Z 1 
Hfm(t)n(t)g = d! M (!)e i!t d N ( )( )e i t :
1 1
(C.39)
The result (C.35) follows immediately.