1. Jesper H. Sørensen, Toshiaki Koike-Akino, and Philip Orlik 2012 IEEE International Symposium on Information Theory Proceedings Rateless Feedback Codes

2. Outline Introduction Background of rateless codes Rateless feedback codes LT feedback codes Raptor feedback codes Numerical results

3. Introduction On the internet When a strong feedback channel is available Using automatic repeat-request (ARQ) schemes. Difficult to employ this scheme in many practical systems. With no frequent feedback A rateless coding can achieve reliable communications. LT codes Raptor codes Complexity : O(klogk)

4. Introduction If the receiver has m feedback opportunities (0 < m < k), What scheme should we applied ? Doped fountain coding [3] The receiver feeds back information on undecoded symbols. Making the transmitter able to transmit input symbols Accelerate the decoding process. Real-time oblivious erasure correcting [4] Telling how many of the k input symbols have been decoded The transmitter chooses a fixed degree Maximizes the probability of decoding new symbols.

5. In this paper We propose LT feedback codes and Raptor feedback codes. Can use any amount of feedback opportunities. The feedback tells the encoder which source symbols have been recovered. Helping the encoder to modify the degree distribution Our goal To decrease the coding overhead especially for short k. We will show that the proposed feedback scheme can decrease both the coding overhead and the complexity

6. Background of rateless codes LT codes encoding process : 1. Randomly choose a degree d by sampling Ω (d). 2. Choose uniformly at random d of the k input symbols. 3. Perform XOR of the d chosen input symbols. LT decoding process : 1. Find all degree-1 output symbols and put the neighboring input symbol into ripple. 2. Symbols in the ripple are processed one by one We potentially reduce a buffered symbols to degree one, we call this a symbol release

7. Rateless feedback codes The ripple size is an important parameter in the design of LT codes. This design contains two steps: 1. Find a suitable ripple evolution to aim for 2. Find a degree distribution which achieves that ripple evolution. We first focus on the case of a single feedback located when f 1 k symbols have been decoded.

8. LT feedback codes The feedback informs the transmitter which input symbols have been decoded. The first f 1 k symbols has no influence on the release of the symbols received after the feedback. The probability that a symbol of degree d is released when L input symbols remain unprocessed :

9. LT feedback codes Symbols received after the feedback Based on the reduced set of input symbols of size (1- f 1 )k. Their releases are independent of the processing of the first f 1 k input symbols. Their release probabilities is We can give the ripple a boost at an intermediate point. k = 100 f1 = 0.5

10. LT feedback codes The ripple size should be kept larger than, for some positive constant c. Every time a symbol is processed, the ripple size is either increased or decreased by one with equal probabilities. Viewing the ripple evolution as a simple random walk. The expected distance from the origin after z steps is

11. LT feedback codes Due to the feedback, the ripple evolution can be viewed as two random walk. Ripple evolution :

12. LT feedback codes The next step in the design is to find a degree distribution which achieves the proposed ripple evolution. We let Q(L) denotes the expected number of releases in the (k L) –th decoding step. We want to let the vector R(L) map into Q(L).

13. The achieved Q(L) can be expressed as a function of the applied degree distribution,, through (1). This is done for the general case without feedback as For the case with feedback, we have contributions from two different degree distributions, LT feedback codes

14. The solution tells the degrees amount before and after the feedback in order to achieve the desired ripple evolution. n 1 and n 2 are the free parameters. Normalizing the solution vectors with n 1 and n 2 provides the degree distributions. We propose to use the least-squares nonnegative solution for (7) to achieve close to the desired ripple evolution. LT feedback codes

16. Consider m feedback opportunities whose locations are at f i, for i =1, 2,...,m. The decoding can be viewed as m+1 random walks. The proposed ripple evolution : LT feedback codes

18. An important performance metric is the average degree, Showing that LT feedback codes decrease the average degree, and thereby computational complexity. LT feedback codes

19. Raptor coding is a concatenation of an LT code and a highrate block code. A weight w : To make it more strict. To avoid the case that the ripple lacks robustness around the feedback point Raptor feedback codes

20. Raptor feedback codes k = 128, f 1 = 0.75, c 1 = c 2 = 1 and d = 2.7 A slight lack of robustness near the feedback point and near the end of decoding Fig. 5. The ripple evolution achieved in the Raptor feedback code compared to the one proposed for LT feedback codes.

21. Numerical results 5000 runs for c 1 = c 2 = 1 The best performance : for k = 32, 64 and 96 for k = 32, 64,96 and 128.

22. Numerical results

23. Raptor feedback codes,for zero, one and two feedback opportunities, we choose d of 5, 3 and 2.5 respectively.

24. Numerical results The average overhead and complexity of LT feedback codes for k = 128 as a function of the number of feedback opportunities.

25. Reference