Recent Results in NonAsymptotic Shannon Theory

1
Recent Results in Non-Asymptotic Shannon Theory

• Dror Baron
• Supported by AFOSR, DARPA, NSF, ONR, and Texas
  Instruments
• Joint work with M. A. Khojastepour, R. G.
  Baraniuk, and S. Sarvotham

2
we may someday see the end of wireline

• S. Cherry, Edholms law of bandwidth, IEEE
  Spectrum, vol. 41, no. 7, July 2004, pp. 58-60

3
But will there ever be enough data rate?

• R. Lucky, 1989
• We are not very good at predicting uses until
  the actual service becomes available. I am not
  worried we will think of something when it
  happens.
• There will always be new applications that gobble
  up more data rate!

4
How much can we improve wireless?

• Spectrum is limited natural resource
• Information theory says we need lots of power for
  high data rates - even with infinite bandwidth!
• Solution transmit more power BUT
• Limited by environmental concerns
• Will batteries support all that power?
• Sooner or later wireless rates will hit a wall!

5
Where can we improve?

• Algorithms and hardware gains
• Power-efficient computation
• Efficient power amplifiers
• Advances in batteries
• Directional antennas
• Communication gains
• Channel coding
• Source coding
• Better source and channel models

6
Where will the last dB of communication
gains come from?Network information theory
(Shannon theory)
7
Traditional point to point information theory
Channel
Decoder
Encoder

• Single source
• Single transmitter
• Single receiver
• Single communication stream
• Most aspects are well-understood

8
Network information theory
Channel
Encoder
Decoder

• Network of
• Multiple sources
• Multiple transmitters
• Multiple receivers
• Multiple communication streams
• Few results
• My goal understand various costs of network
  information theory

Encoder
Channel
9
What costs has information theory overlooked?
10
Channel coding has arrived
Channel
Decoder
Encoder

• Turbo codes Berrou et al., 1993
• 0.5 dB gap to capacity (rate R below capacity)
• BER10-5
• Block length n6.5104
• Regular LDPC codes Gallager, 1963
• Irregular LDPC Richardson et al., 2001
• 0.13 dB gap to capacity
• BER10-6
• n106

11
Distributed source coding has also arrived
Decoder
Encoder
x
y

• Encoder for x based on syndrome of channel code
• Decoder for x has correlated side information y
• Various types of channel codes can be used
• Slepian-Wolf via LDPC codes Xiong et al., 2004
• H(XY)0.47
• R0.5 (rate above Slepian-Wolf limit)
• BER10-6
• Block length n105

12
Hey! Did you notice those block lengths?

• Information theory provides results in the
  asymptotic regime
• Channel coding 8?0, rate RC-? achievable with
  ?!0 as n!1
• Slepian-Wolf coding 8?0, rate RH(XY)?
  achievable with ?!0 as n!1
• Best practical results achieved for n105
• Do those results require large n?

13
But we live in a finite world

• Real world data doesnt always have n106
• IP packets
• Emails, text messages
• Sensornet applications
• (small battery ! small n)
• How do those methods perform for n104? 103?
• How quickly can we approach the performance
  limits of information theory?

14
And we dont know the statistics either!

• Lossless coding (single source)
• Length-n input xBer(p)
• Encode with wrong parameter q
• K-L divergence penalty with variable rate codes
• Performance loss (minor bitrate penalty)
• Channel coding, distributed source coding
• Encode with wrong parameter q
• Fixed rate codes based on joint-typicality
• Typical set Tq for q is smaller than Tp for p
• As n!1, Pr(error)!1
• Performance collapse!

15
Main challenges

• How quickly can we approach the performance
  limits of information theory?
• Will address for channel coding and Slepian-Wolf
• What can we do when the source statistics are
  unknown?
• Will address for Slepian-Wolf

16
But first . . .What does the prior art indicate?
17
Underlying problem

• Shannon 1958
• This inverse problem is perhaps the more
  natural in applications given a required level
  of probability of error, how long must the code
  be?
• Motivation may have been phone and space
  communication
• Small probability of codeword error ?
• Wireless paradigm
• Given k bits, what are the minimal channel
  resources to attain probability of error ??
• Can retransmit packet ? fix large ?
• n depends on packet length
• Need to characterize R(n,?)

18
Error exponents

• Fix rate R
• Bounds on probability of error
• Random coding Prerror2-nEr(R)
• Sphere packing Prerror2-nEsp(R)o(n)
• Er(R)Esp(R) for R near C

19
Error exponents

• Fix rate R
• Bounds on probability of error
• Random coding Prerror2-nEr(R)
• Sphere packing Prerror2-nEsp(R)o(n)
• Er(R)Esp(R) for R near C
• Shannons regime
• This inverse problem is perhaps the more
  natural in applications given a required level
  of probability of error, how long must the code
  be?
• Fix R
• E(R)O(1)
• log(?)O(n) good for small ?

20
Error exponents

• Fix rate R
• Bounds on probability of error
• Random coding Prerror2-nEr(R)
• Sphere packing Prerror2-nEsp(R)o(n)
• Er(R)Esp(R) for R near C
• Wireless paradigm
• Given k bits, what are the minimal channel
  resources to attain probability of error ??
• Fix ?
• nE(R)O(1)
• o(n) term dominates
• Bounds diverge

21
Error exponents fail for RC-?/n0.5
22
How quickly can we approach the channel
capacity?(known statistics)
23
Binary symmetric channel (BSC) setup
s
xf(s)
y
sg(y)
Encoder f
Decoder g
zBer(n,p)

• s21,,M input message
• x, y, and z binary length-n sequences
• zBernoulli(n,p) implies crossover probability p
• Code (f,g,n,M,?) includes
• Encoder xf(s)21,,M
• Rate Rlog(M)/n
• Channel yx?z
• Decoder g reconstructs s by sg(y)
• Error probability Prg(y)?s?

24
Non-asymptotic capacity
25
Key to solution Packing typical sets

• Need to encode typical set TZ for z
• Code needs to cover z2Tz
• Need Pr(z2TZ)¼ 1-?
• Probability ? of codeword error
• What about rate?
• Output space 2n possible sequences
• Cant pack more than 2n/Tz sets into output
• M2n/Tz
• Minimal cardinality Tmin covers 1-?
• CNA1-log(Tmin)

output space
Tz
26
Whats the cardinality of Tmin?

• Consider empirical statistics nz?i zi, PZnz/n
• p
• Minimal Tmin has form Tmin,z PZ?(?)
• Determine ?(?) with central limit theorem (CLT)
• EPZp, Var(PZ)p(1-p)/n
• PzN(p,p(1-p)/n)

• Asymptotic
• ?p?
• LLN ??0

• Non-asymptotic
• ?p?p(1-p)/n0.5
• CLT ? ! ?(?)

27
Tight non-asymptotic capacity

• Theorem
• CNA(n,?)C-K(?)/n0.5o(n-0.5)
• K(?)?-1(?) p(1-p)0.5 log((1-p)/p)
• Gap to capacity is K(?)/n0.5o(n-0.5)
• Note o(n-0.5) asymptotically negligible w.r.t.
  K/n0.5
• Tightened Wolfowitz bounds up to o(n-0.5)
• Gap to capacity of LDPC codes 2-3x greater
• We know how quickly we can approach C

28
Non-asymptotic capacity of BSC
29
Gaussian channel results
s
xf(s)
y
sg(y)
Encoder f
Decoder g
zN(0,?2)

• Continuous channel
• Power constraint ?i(xi)2 nP
• Shannon 1958 derived CNA(n,?) for Gaussian
  channel via cone packing (non-i.i.d. codebook)
• Information spectrum bounds on probabilities of
  error indicate Gaussian codebooks are sub-optimal
• i.i.d. codebooks arent good enough!

30
Excess power of Gaussian channel
31
How quickly can we approach the Slepian-Wolf
limit?(known statistics)
32
But first . . .Slepian-Wolf Review
33
Slepian-Wolf setup
x
fX(x)
gX(fX(x),fY(y))
Encoder fX
Decoder g
y
fY(y)
gY(fX(x),fY(y))
Encoder fY

• x and y are correlated length-n sequences
• Code (fX,fY,gX,gY,n,MX,MY,?X,?Y) includes
• Encoders fX(x)21,,MX, fY(y)21,,MY
• Rates RXlog(MX)/n, RYlog(MY)/n
• Decoder g reconstructs x and y by gX(fX(x),fY(y))
  and gY(fX(x),fY(y))
• Error probabilities PrgX(fX(x),fY(y))?x?X and
  PrgY(fX(x),fY(y))?y?Y

34
Slepian-Wolf theorem

• Theorem SlepianWolf,1973
• RXH(XY) (conditional entropy)
• RYH(YX)
• RXRYH(X,Y) (joint entropy)

RY
Slepian-Wolf rate region
H(Y)
H(YX)
RX
H(X)
H(XY)
35
Slepian-Wolf with binary symmetric correlation
structure (known statistics)
36
Binary symmetric correlation setup

• y, z, and z are length-n Bernoulli sequences
• Correlation channel z is independent of y
• Bernoulli parameters p,q20,0.5), rp(1-q)(1-p)q
  
• Code (f,g,n,M,?) includes
• Encoder f(x)21,,M
• Rate Rlog(M)/n
• Decoder g(f(x),y)20,1n
• Error probability Prg(f(x),y)?x?

37
Relation to general Slepian-Wolf setup

• x, y, and z are Bernoulli
• Correlation z independent of y
• Focus on encoding x at rate approaching H(Z)
• Neglect well-known encoding of y at rate RYH(Y)

RY
our setup
H(Y)
H(YX)
RX
H(X)
H(Z)
38
Non-asymptotic Slepian-Wolf rate

• Definition RNA(n,?)min9 code (f,g,n,M,?)
  log(M)/n
• Prior art Wolfowitz,1978
• Converse result RNA(n,?) H(XY)KC(?)/n0.5
• Achievable result RNA(n,?) H(XY)KA(?)/n0.5
• Bounds are loose KA(?)KC(?)
• Can we tighten Wolfowitzs bounds?

39
Tight non-asymptotic rate

• Theorem
• RNA(n,?)H(Z)K(?)/n0.5o(n-0.5)
• K(?)?-1(?) q(1-q)0.5 log((1-q)/q)
• Redundancy rate is K(?)/n0.5o(n-0.5)
• Note o(n-0.5) decays faster than K/n0.5
• Tightened Wolfowitz bounds up to o(n-0.5)
• We know how quickly we can approach H(Z) with
  known statistics

RNA(n,?)
tight bound
H(XY)
n
40
What can we do when the source statistics are
unknown?(universality)
41
Universal setup

• Unknown Bernoulli parameters p, q, r
• Encoder observes x and ny?iyi
• Communication of ny requires log(n) bits
• Variable rate used
• Need distribution for nz
• Distribution depends on nx and ny (not x)
• Codebook size Mnx,ny

42
Distribution of nz

• CLT was key to solution with known statistics
• How can we apply CLT when q is unknown?
• Consider a numerical example
• p0.3, q0.1, rp(1-q)(1-p)q
• PXr, PYp, PZq (empirical true)
• We plot Pr(nznx,ny) as function of nz20,,n

43
Pr(nznx,ny) for n102
44
Pr(nznx,ny) for n103
45
Pr(nznx,ny) for n104
46
Pr(nznx,ny) for n104
where
47
Universal rate

• Theorem
• RNA(n,?)H(PZ)K(?)/n0.5o(n-0.5)
• K(?)f(PY)K(?)
• f(PY)2PY(1-PY)0.5/1-2PY

f(PY)!1
f(PY)!0
48
Why is f(PY) small when PY is small?

• Known statistics ? Var(nz)nq(1-q) regardless of
  empirical statistics
• PY!0 ? can estimate nZ with small variance
• Universal scheme outperforms known statistics
  when PY is small
• Key issue variable rate coding (universal) beats
  fixed rate coding (known statistics)
• Can cut down expected redundancy (known
  statistics) by communicating ny to encoder
• log(n) bits for ny will save O(n0.5)

49
Redundancy for PY¼0.5

• f(PY) blows up as PY approaches 0.5
• Redundancy is O(n-0.5) with enormous constant
• Another scheme has O(n-1/3) redundancy
• Better performance for PY0.5O(n-1/6)
• Universal redundancy can be huge!
• Ongoing research improvement of O(n-1/3)

50
Numerical example

• n104
• q0.1
• Slepian-Wolf requires nH2(q)4690 bits
• Non-asymptotic approach (known statistics) with
  ?10-2 requires nRNA(n,?)4907 bits
• Universal approach with PY0.3 requires 5224 bits
• With PY0.4 we need 5863 bits
• In practice, penalty for universality is huge!

51
Summary

• Network information theory (Shannon theory) may
  enable to increase wireless data rates
• Practical channel codes and distributed source
  codes approaching limits, rely on large n
• How quickly can we approach the performance
  limits of information theory?
• CNAC-K(?)/n1/2o(n-1/2)
• RNAH(Z)K(?)/n1/2o(n-1/2)
• Gap to capacity of LDPC codes 2-3x greater

52
Universality

• What can we do when the source statistics are
  unknown? (Slepian Wolf)
• PY
• PY¼0.5 H(PZ)O(n-1/3) can be huge!
• Universal channel coding with feedback for BSC
• Capacity-achieving code requires PY0.5
• Universality with current scheme is O(n-1/3)

Channel
Decoder
Encoder
feedback
53
Further directions

• Gaussian channel (briefly discussed)
• Shannon 1958 derived CNA(n,?) for Gaussian
  channel with cone packing (non-i.i.d. codebook)
• Gaussian codebooks are sub-optimal!
• Other channels
• CNA(n,?) C-KA(?)/n0.5 via information spectrum
• Gaussian codebook distribution sub-optimal
• Must consider non-i.i.d. codebook constructions
• Penalties for finite n and unknown statistics
  exist everywhere in Shannon theory!!
• www.dsp.rice.edu