Iterative Quantization. Using Codes On Graphs

Transcription

1 Iterative Quantization Using Codes On Graphs Emin Martinian and Jonathan S. Yedidia 2 Massachusetts Institute of Technology 2 Mitsubishi Electric Research Labs

2 Lossy Data Compression: Encoding: Map source x = [x x 2... x n ] to bits. x Decoding: Map bits to y = [y y 2... y n ]. y x Encoder Decoder Goal: Construct mapping so that y near x. y

3 Lossy Compression Performance: Minimum rate is R(D) = inf I(x; y) p(y x):e[d(x,y)]=d

4 Lossy Compression Performance: Minimum rate is R(D) = inf I(x; y) p(y x):e[d(x,y)]=d Gap to R(D) for mean square distortion measures: Asymptotically high rates ECSQ gap =.53 db

5 Lossy Compression Performance: Minimum rate is R(D) = inf I(x; y) p(y x):e[d(x,y)]=d Gap to R(D) for mean square distortion measures: Asymptotically high rates ECSQ gap =.53 db Moderate rates, Gaussian source ECSQ gap = db

6 Lossy Compression Performance: Minimum rate is R(D) = inf I(x; y) p(y x):e[d(x,y)]=d Gap to R(D) for mean square distortion measures: Asymptotically high rates ECSQ gap =.53 db Moderate rates, Gaussian source ECSQ gap = db 256-state Trellis Coded Quantization gap = db

7 Lossy Compression Performance: Minimum rate is R(D) = inf I(x; y) p(y x):e[d(x,y)]=d Gap to R(D) for mean square distortion measures: Asymptotically high rates ECSQ gap =.53 db Moderate rates, Gaussian source ECSQ gap = db 256-state Trellis Coded Quantization gap = db How do we approach R(D) with low complexity?

8 Compressing a source with erasures x 0 y 0 * Pr[x = ] = s, Pr[x = 0] = Pr[x = ] = ( s)/2

9 Compressing a source with erasures x 0 d(0,)=0. * d(*, )=0 y 0 d(,0)=0 Pr[x = ] = s, Pr[x = 0] = Pr[x = ] = ( s)/2 Quantizing 0 or 0 causes non-zero distortion

10 Performance for erasures models R(D = 0) PSfrag replacements Capacity Pr[x = ] Pr[x = ]

11 Performance for erasures models R(D = 0) Capacity PSfrag replacements Pr[x = ] Pr[x = ]

12 Coding Complexity Random Linear Code O(n 3 ) PSfrag replacements O(n) Block Length

13 Coding Complexity Random Linear Code O(n 3 ) Codes On Graphs O(n) PSfrag replacements Block Length Block Length

14 Data Compression Error Correction Duality Data Compression Code x Source Source Encoder Decoder y Error Correction Code Channel y x Channel Channel Encoder Decoder

15 An LDPC Quantizer 0 Source * Symbols 0 To Quantize * *

16 An LDPC Quantizer 0 0 * *

17 An LDPC Quantizer 0 0 * *

18 An LDPC Quantizer 0 * 0 * *

19 An LDPC Quantizer 0 * 0 * *

20 Problems with LDPC Quantizers: Theorem: LDPC quantizers fail if parity check density < o(log n). Proof Idea: Imagine f n checks have degree d

21 Problems with LDPC Quantizers: Theorem: LDPC quantizers fail if parity check density < o(log n). Proof Idea: Imagine f n checks have degree d For each low degree check P r[check unsatisfied] 2 ( e)d

22 Problems with LDPC Quantizers: Theorem: LDPC quantizers fail if parity check density < o(log n). Proof Idea: Imagine f n checks have degree d For each low degree check P r[check unsatisfied] 2 ( e)d E[bad checks] f n 2 ( e)d

23 Problems with LDPC Quantizers: Theorem: LDPC quantizers fail if parity check density < o(log n). Proof Idea: Imagine f n checks have degree d For each low degree check P r[check unsatisfied] 2 ( e)d E[bad checks] f n 2 ( e)d Successful quantization requires d log n

24

25 Dual LDPC Quantizer 0 Uncompressed * 0 Bits * * Compressed Bits

26 Dual LDPC Quantizer 0

27 Dual LDPC Quantizer 0 0

28 Dual LDPC Quantizer 0

29 Dual LDPC Quantizer 0 * 0 * *

30 Dual LDPC Quantizer 0 * 0 * * 0 0

31 Optimal Decoding/Quantization Duality: * 0 * 0/ * Any data encoded with C and received with erasures e can be decoded any source with erasures e can be quantized by C.

32 Message-Passing Rules for Erasure Decoding: = # 0 # 0

33 Erasure Decoding For LDPC Code: *

34 Erasure Decoding For LDPC Code:

35 Erasure Quantization For Dual LDPC Code: * *

36 Erasure Quantization For Dual LDPC Code: * * *

37 Erasure Quantization For Dual LDPC Code: * * *

38 Erasure Quantization For Dual LDPC Code: * * * 0

39 Erasure Quantization For Dual LDPC Code: * * 0

40 Erasure Decoding vs. Erasure Quantization: * 0 * = # 0 # 0

41 Erasure Decoding vs. Erasure Quantization: * 0 * = 0 0 * * * * 0 0 # 0 # 0

42 Message-Passing Rules for Erasure Quantization: = # 0 0 # 0 0

43 Erasure Quantization For Dual LDPC Code: 0 * * * *

44 Erasure Quantization For Dual LDPC Code: * 0 * * * *

45 Erasure Quantization For Dual LDPC Code: * * 0 * * * *

46 Erasure Quantization For Dual LDPC Code: * * 0 * * * * 0

47 Erasure Quantization For Dual LDPC Code: * * 0 * * * *

48 Erasure Quantization For Dual LDPC Code: * * 0 * * * *

49 Erasure Quantization For Dual LDPC Code: * 0 * * * *

50 Erasure Quantization For Dual LDPC Code: 0 * 0 * *

51 Iterative Decoding/Quantization Duality: Theorem: The code C can iteratively decode erasures e iff the dual code C can iteratively quantize erasures e.

52 Iterative Decoding/Quantization Duality: Theorem: The code C can iteratively decode erasures e iff the dual code C can iteratively quantize erasures e. Corollary: Duals of capacity achieving codes for erasure decoding, achieve the minimum rate for erasure quantization.

53 In theory, there is no difference between theory and practice. But, in practice, there is. Jan L.A. van de Snepscheut

54 In theory, there is no difference between theory and practice. But, in practice, there is. Jan L.A. van de Snepscheut

55 Concluding Remarks: LDPC codes = bad quantizers; dual codes = good quantizers. Duality between optimal/iterative decoding/quantization. Future work: extend analysis/algorithms/duality to all source models and distortion measures.