Radioengineering
Vol. 9, No. 2, June 2000
‘Theoretical Background of an Optimal Multrate Subband Coding of Cyclostationary Signals 5
D.KULA,
FUNDAMENTALS OF AN OPTIMAL MULTIRATE
SUBBAND CODING OF CYCLOSTATIONARY
SIGNALS
Daniel KULA!
Dept. of Electrical and Computer Engineering,
‘The University of lowa
210 CC Towa City, TA 52242
USA.
E-mail: daniel kula@ceskymobil.cz.
Abstract
A consistent theory of optimal subband coding of
zero mean wide-sense cyclostationary signals, with N-
periodic statistics, is presented in this article. An Mf-
channel orthonormal uniform filter bank, employing
Neperiodic analysis and synthesis filters, is used to
form the subband coder. A dynamic scheme involving
N-periodic bit allocation is used, while an average
variance condition is applied to evaluate the output
distortion. In three lemmas and final theorem, the
necessity of decorrelation of blocked subband signals
and requirement of specific ordering of power spectral
densities are_proven.
Keywords
Coding Gain, Decimation, Filter Bank, Polyphase
Representation, Power Spectral Density Matrix,
Maltirate System, Decorrelation, Subbabd Coding
1. Introduction
A structure of the Multirate Subband Coder for Wide-
sense Cyclostationary Signals, that is to be treated
hereinafter is depicted in Fig. 1. Set of fillers
H,(k,z7!)and F,(k,z~"), represent the Analysis Bank
and the Synthesis Bank respectively. The index k indicates
time varying nature of these blocks, since each of them in
fact consists of a sequence of NM LTI (Linear Time-
Invariant) filters, where 1 is assigned to the periodicity of
cyclostationarity of an LPTV (Linear Periodically Time-
Varying) structure. For the sake of simplicity we assume
that period of time variation of the structure in Fig. 1 is
the same as that of input signal, Blocks to the left of the
analysis bank are M-fold decimators that discard all but
' Also with: Dept. of Radioclectronics, FEL VUT
Brno, Purkyiiova 118, 612 00 Brno, Czech Republic,
every M-th sample. Blocks to the right of the synthesis
bank are M-fold interpolators that raise the sampling rate
by a factor of M, by inserting (M-1) zero samples between
two consecutive samples of an incoming steam, Blocks
Q, stand for A-D converters, communication channel and
D-A converters all together.
sw “iy
maa} Eo Fe
Face)
F
mar fe te} am |
fans 0
Fig, 1: Multirate Subband Coder for WSCS Signals
Ps)
Solely for the purpose of following derivation, they
are assumed to be a source of additive, uncorrelated, zero
mean quantizing noise q, (K) (see Fig.2), such that
(=) +9) . a)
and the variance of which is
a3 =02"%0% @
Here 02 is the variance of i-th subband signal and c is a
constant determined by the signal distribution.
The WSCS (Wide Sense CycloStationary) signal to
be coded is denoted x(&), subband signals are v,(k)'s (for
F running from 0 to (4-1) ), £(k) is the reconstruction of
2(), synthesized in the output of the synthesis bank.
Matrices £(k,2"') and R(é,z) in Fig.2 are polyphase
representations of the analysis and synthesis filter banks,
respectively. They are introduced to simplify mathematical
treatment of the coder. Since both analysis and synthesis
filter banks are LPTV structures, these polyphase matrices
are also LPTV, as denoted by time index k.
The goal to be achieved is to bring the synthesized
output 2(k) as close to x(k) as possible, by minimizing
distortion introduced by quantizers, ic. maximizing so
called Coding Gain of the coder. This is done by suitable
choice of the analysis and synthesis filter banks, as well as
by proper allocation of bits 5, to individual quantizers O,
The overall bitrate
pnw o
=
is kept constant for all k,6 “Theoretical Background of an Optimal Mutrate Subband Coding of Cyclostattonary Signals, Radicengineering
0. KULA Vol. 9, No. 2, June 2000
ye
sath soe v9 tt cy
re low | I sou coun % =| I
ai
o io
tt ow | | “pansy
Far rime gs vacate) Lv" 8
: son vie 3 wa fame
tet? + | amen] gy [nome a zey [kt 5
est vat ae! key |v now | Rey Su
— vem
vormnany women Lm A
Fig, 2: Polyphase Decompostion of Multirate Coder 5
Pte Pes xoeuven) wom) I Linon
i Guam
A definition of the Coding Gain, that follows Pane 1 Ne
naturally from the theory of WSS signals [2],[4] is ee 50 Ba nq ———
E(R(K) ~ xP)
O= FEO)"
a
where %(K) in the numerator is a reference output of the
direct quantizer depicted in Fig3, and where the
denominator can be expressed after using AM-GM
inequality [2] as
co? ys
Joa = an ¥ (fer) oO
Expression (5) is the function to be minimized to
achieve optimum performance of the coder. Especially for
‘M>2 it is not a trivial task, and the rest of this article is
mostly devoted to the solution to this problem,
Aw _, Quantizer |_xt)
Fig. 3: Direct b-bit Quantizer (PCM)
To overcome problems brought to the analysis of the
proposed structure by LPTV nature of the filter banks, an
equivalent Blocked LT! system is to be analysed, as
depicted in Fig.4, Here, the parallel (time-demultiplexed)
input signals x(K), subband signals v,(ie) together with
associated quantizing noise signals q,(k) and parallel
reconstructed output £(X), can be described by NM-fold
vectors X(k), F(k), O(k) and X(k), respectively.
Matrices E(z) and R(z") are so called Blocked
Polyphase Representations of the analysis and synthesis
bank respectively, while the decimators, interpolators and
delay-chains displayed in Fig.2, lay outside the depicted
structure [5].
Fig. &: Blocked Version of Multiale Subband Coder Using
Polyphase Representations of the Fier Banks
2. Mathematical concept
In following definitions, lemmas and theorem, our
effort is to prove that optimal performance of the LPTV
Multirate Subband Coder, i.e. maximised Coding Gain as
defined in (4), (5), is achieved if subband signals
vy, (k) are mutually decorrelated and diagonal elements of
the PSD (Power Spectral Density) matrix of the vector
V7 (k) obeys a specific ordering for all @&, This ordering
shall be unique,
Definition 1:
Consider a sequence 4 = {a,}"" , such that
G24, 2.2 yyy 20, where MN=MxN
Define a set of integers S = {1,2,3,....., MN}, its
paritioning —{5,,5y,...,Sy}, such that
card|s,]=M, vi €{1,2,...,.N}
Define a Characteristic Sum
(AS, S35
and a Minimum Sum
Sv)
J*(A)= min J(A,S),S,,
where the sequence-argument
{s*(a}={5,*(.S, (D,-Sy A}, ®
represents The Best Ordering of a sequence A, that may
not be unique.Radloengineering
Vol. 9, No. 2, June 2000
‘Theoretical Background of an Optimal Mullirate Subband Coding of Cyclostationary Signals 7
O.KULA
Lemma 1:
Consider sequence 4, and J, J*, (S*4)} as in (6)
through (8). For k €S,*(A) , eS, *(A), where
klef{l2...MN} Rel, 1,7 ef12,..,Npiey
Let introduce:
Tle, . °
aeS*tAlgrk
and
(10)
If a, >a,,then G, Sa,,
Proof: By contradiction
Consider
HAS S888 5859) (U)
Sr= {54 -}UG
Syt={5,*(4)- JUS
Then write:
J'(A)= I*(A) = (Gg) + @,G,,)" =
y=
(aya) ~ (GF my
uM
Gx
x
wary
Assuming a, >a, and also G,>@,, we conclude
that J'(A) = J *(A) <0 Since J *(A) is
The Minimum Sum, this establishes a contradiction.
Hence, for a, >a,
Gy SH, (12)
must hold.
Definition 2: Majorization
Lets take sequence 4 = {a,}""
and A= {2},
i
with elements labeled such that
4,24, 2...2ay 20, AA Bw.AzAy BO.
Suppose that A and A obey relation:
i
Da < DA, torisksN, (13)
with equality at A=. Then we say that A majorizes 4, or
A is majorized by A , noting A< A
Definition 3: Schur-concave function
A real-valued function P(x,,x,,.
ona subset JC R" is said to be strictly Schur-concave
cn this subset whenever X’ ~ ¥ implies
O(X)2 OF), (ay
in terms of equal elements,
»X,), defined
with equality if.\=
Theorem 1;
Let ®(x,,%,
where x, >x,>..>x, 20, defined on JC R" and
+) is
X,) be a real-valued function,
twice differentiable on its interior. D(.x,,x,,..
symmetric on 3. Denote [2]
WX
®uy(X) = e ) and
POX
aan= Brg
Then P(X) is strictly Schur-concave on 3 if:
1. Qq)(X)is inoreasing in k
2. Puy(X) = Mery (X) implies:
PAD ~ Peper 4) Rois AVF Ressean(A) <0
Proof of this theorem is available in (3)
Lemma 2:
The Minimum Sum as presented in (7) is strictly
Schur-concave function and hence for all A and A,
A~4,,, 20, is a," 0. Also follows from Fig.
4, that PSD matrix of 7 (k) obeys
$5, = Fes, okey”
ey
Let's return to the denominator of the Coding Gain (5)
that concerns the subband o,,(/) for
ie{0,...M-}} , 7 {0,...,N—l}. Obviously the ordering
of the sequence
variances
(js, cone lok MV- |
Me
plays an important part in minimizing (5).
Lemma 3:
Consider sequence Eas above and Z(w)as in
(20). Then under (18):
ag
5, ifs (oe) locks va «22
Proof:
Suppose that the elements of S,(@) are not
ordered in a consistent way at all @, e.g. around some
frequency @, (S,(@,)), >(Sp(@,)},» but around
another frequency @; (S,(@,)), <(Sp(a,)},,- Then at
cach @ there exists a permutation matrix P() such that
forall @,
S'7 = P(@)S;(@)P'(@) 23)
obeys
(sr(@,)), 2(S-(o)), - (24
Define
ae
{iff lo