Finite-Length Scaling and Error Floors

53
1 Finite-Length Scaling and Error Floors Abdelaziz Amraoui Andrea Montanari Ruediger Urbanke Tom Richardson

description

Finite-Length Scaling and Error Floors. Tom Richardson. Abdelaziz Amraoui Andrea Montanari Ruediger Urbanke. Approach to Asymptotic. Finite Length Scaling. Finite Length Scaling. Finite Length Scaling. Finite Length Scaling. Analysis (BEC): Covariance evolution. - PowerPoint PPT Presentation

Transcript of Finite-Length Scaling and Error Floors

Page 1: Finite-Length Scaling and Error Floors

1

Finite-Length Scalingand Error Floors

Abdelaziz AmraouiAndrea MontanariRuediger Urbanke

Tom Richardson

Page 2: Finite-Length Scaling and Error Floors

2

Approach to Asymptotic PDF File

Page 3: Finite-Length Scaling and Error Floors

3

Finite Length Scaling

Page 4: Finite-Length Scaling and Error Floors

4

Finite Length Scaling

Page 5: Finite-Length Scaling and Error Floors

5

Finite Length Scaling

Page 6: Finite-Length Scaling and Error Floors

6

Finite Length Scaling

Page 7: Finite-Length Scaling and Error Floors

7

Analysis (BEC): Covariance evolution

Fraction of check nodes of degreegreater than one and equal to one.

Covariance terms.

As a function of residual graph fractional size.

Page 8: Finite-Length Scaling and Error Floors

8

Covariance evolution

Page 9: Finite-Length Scaling and Error Floors

9

Finite Length Curves

Page 10: Finite-Length Scaling and Error Floors

10

Analysis (BEC)

• Follow Luby et al: single variable at a time with the trajectory converging to a differential equation.

• Covariance of state space variables also follows a d.e.

• Increments have Markov property and regularity.

Page 11: Finite-Length Scaling and Error Floors

11

Results

Page 12: Finite-Length Scaling and Error Floors

12

Finite Threshold Shift

Page 13: Finite-Length Scaling and Error Floors

13

Generalizing from the BEC?

•No obvious incremental form (diff. Eq.)•No state space characterization of failure.•No clear finite dimensional state space.•Not clear what the right coordinates are for the general case (Capacity?).

Nevertheless, it is useful in practice to havethis interpretation of iterative failure and tohave the basic form of the scaling law.

Page 14: Finite-Length Scaling and Error Floors

14

Empirical Evidence

Page 15: Finite-Length Scaling and Error Floors

15

Error Floors

Page 16: Finite-Length Scaling and Error Floors

16

Page 17: Finite-Length Scaling and Error Floors

17

Page 18: Finite-Length Scaling and Error Floors

18

Error Floors:Review of the BEC

Page 19: Finite-Length Scaling and Error Floors

19

Error floors on the erasure channel: Stopping sets.

Page 20: Finite-Length Scaling and Error Floors

20

Error floors on the erasure channel: average performance.

Page 21: Finite-Length Scaling and Error Floors

21

Error floors on the erasure channel: decomposition

Page 22: Finite-Length Scaling and Error Floors

22

Error floors on the erasure channel: average and typical performance.

Page 23: Finite-Length Scaling and Error Floors

23

Error floors for general channels: Expurgated Ensemble Experiments.

AWGN channel rate 51/64 block lengths 4k

Random

Girth (8) optimized

Neighborhood optimized

Page 24: Finite-Length Scaling and Error Floors

24

Error floors for general channels: Trapping set distribution.

AWGN channel rate 51/64 block lengths 4k

(3,1)

(5,1)

(7,1)

Page 25: Finite-Length Scaling and Error Floors

25

Observations.

•Error floor region dominated by small weight errors.

•Subset on which error occur usually induces a subgraph with only degree 2 and degree 1 check nodes where the number of degree 1 check nodes is relatively small.

•Optimized graphs exhibit concentration of types of errors.

Page 26: Finite-Length Scaling and Error Floors

26

IntuitionIn the error floor event, nodes in the trapping sets receive 1’s with some reliability. Other nodes receive typical inputs.

(Reliable 1)

(Definite 0)

After a few iterations ‘exterior’ nodes and messages converge to high reliability 0s. Internally messages are 1s.

(Definite 0)

1

1

11

1

Nevertheless, if internal received are 1s, internal messaging reaches highly reliable 1 and message state gets trapped.

(Definite 0)

(9,3)

Page 27: Finite-Length Scaling and Error Floors

27

A decoder on an input ℇ Y is a sequence of maps:

Dl : ℇ {0,1}n

Defining Failure: Trapping Sets

(Assume the all-0 codeword is the desired decoding.

For the BEC let 1 denote an erasure.)

We say that bit j is eventually correct if there exists L so that l > L implies Dl(ℇ ) = 0.

Assuming failure, the trapping set T is the set of all bits that are not eventually correct.

Page 28: Finite-Length Scaling and Error Floors

28

Defining Failure for BP: Practice

Decode for 200 iterations. If the decoding is not successful decode an additional 20 iterations and take the union of all bits that do not decode to 0 during this time.

Page 29: Finite-Length Scaling and Error Floors

29

Trapping Sets: Examples

1. Let the decoder be the maximum likelihood decoder in one step. Then the trapping sets are the non-zero codewords.

2. Let the decoder be belief propagation over the BEC. Then the trapping sets are the stopping sets.

3. Let the decoder be serial (strict) flipping over the BSC. T is a trapping set if and only if the in the subgraph induced by T each node has more even then odd degree neighbors, and the same holds for the complement of T.

Page 30: Finite-Length Scaling and Error Floors

30

Analysis with Trapping Sets:Decomposition of failure

FER() = T P(ℇT, )

ℇT: The set of all inputs giving rise to failure on trapping set T.

Error Floors dominated by “small” events.

Page 31: Finite-Length Scaling and Error Floors

31

1. Find (cover) all trapping sets likely to have significant contribution to the error floor region.

T1,T2,T3,….,Tm

2. Evaluate contribution of each set to the error floor.

P(ℇT1, ), P(ℇT2, ),…

Predicting Error Floors: A two pronged attack.

Strictly speaking, we get a (tight) lower bound

FER() > i P(ℇTi, )

Page 32: Finite-Length Scaling and Error Floors

32

Finding Trapping Sets

Simulation of decoding can be viewed as stochastic process for finding trapping sets.

It is very inefficient, however.

We could use (aided) flipping to get some speed up.

It is still too inefficient.

Page 33: Finite-Length Scaling and Error Floors

33

Finding Trapping Sets (Flipping)

•Trapping sets can be viewed as “local” extrema of certain functions. E.g., number of odd degree induced checks.

•“Local” means, e.g., under single element removal, addition, or swap.

Therefore, we can look for subsets that are “local” extrema.

Page 34: Finite-Length Scaling and Error Floors

34

Finding Trapping Sets (Flipping)Basic idea:

•Build up a connected subset with bias towards minimizing induced odd degree checks.

•Check occasionally for containment of a in-flipping stable set by applying flipping decoding. Eventually such a set is contained.

•Check now for other types of variation:

•Out-flipping stability.

•Single aided flip stability (chains).

•……

Page 35: Finite-Length Scaling and Error Floors

35

Differences: BP and Flipping

Page 36: Finite-Length Scaling and Error Floors

36

Differences: BP and Flipping

r1

r1+r2+r3

r3

r2

Page 37: Finite-Length Scaling and Error Floors

37

Differences: BP and Flipping

Page 38: Finite-Length Scaling and Error Floors

38

Differences: BP and FlippingPDF File

Page 39: Finite-Length Scaling and Error Floors

39

•Find random variable x on which to condition the decoder input Y that “mostly” determines membership in ℇT . I.e.,

Pr{ℇT | x} is nearly a step function in x.

•Perform in situ simulation of trapping set while varying x to measure Pr{ℇT | x}.

•Combine with density of x to get Pr{ℇT }.

Evaluating Trapping SetsBasic idea:

Page 40: Finite-Length Scaling and Error Floors

40

Evaluating Trapping Sets

Condition input to trapping set

Otherwise simulate channel

Page 41: Finite-Length Scaling and Error Floors

41

Evaluating Trapping Sets: BECX is the number of erasures in T (=S).

Page 42: Finite-Length Scaling and Error Floors

42

Evaluating Trapping Sets: AWGNX is the mean noise input in T.

Page 43: Finite-Length Scaling and Error Floors

43

Evaluating Trapping Sets: AWGN

Page 44: Finite-Length Scaling and Error Floors

44

Evaluating Trapping Sets: Margulis 12,4

Page 45: Finite-Length Scaling and Error Floors

45

A tougher test case:G

Page 46: Finite-Length Scaling and Error Floors

46

Evaluating Trapping Sets: G

Page 47: Finite-Length Scaling and Error Floors

47

A Curve from a single point

Page 48: Finite-Length Scaling and Error Floors

48

Extrapolating a curve

Page 49: Finite-Length Scaling and Error Floors

49

Variation in Trapping Sets: (10,4) (10,2) (10,0)

Page 50: Finite-Length Scaling and Error Floors

50

Variation in Trapping Sets: (12,4) (12,2) (12,0)

Page 51: Finite-Length Scaling and Error Floors

51

Variation in Trapping Sets

Page 52: Finite-Length Scaling and Error Floors

52

Conclusions

•Error floor performance is predictable with considerable computational effort. (Would be nice to have scaling law for “best” codes.)

•Trade off between error floor and threshold (waterfall) can be optimized even for very deep error floors.

Page 53: Finite-Length Scaling and Error Floors

53

Conclusions

“It is interesting to observe that the search for theoretical understanding of turbo codes has transformed coding theorists into experimental scientists.”physicists.”