1. The Factor Graph Approach to Model-Based Signal Processing Hans-Andrea Loeliger

2. 2 Outline Introduction Factor graphs Gaussian message passing in linear models Beyond Gaussians Conclusion

3. 3 Outline Introduction Factor graphs Gaussian message passing in linear models Beyond Gaussians Conclusion

4. 4 Introduction Engineers like graphical notation It allow to compose a wealth of nontrivial algorithms from tabulated “local” computational primitive

5. 5 Outline Introduction Factor graphs Gaussian message passing in linear models Beyond Gaussians Conclusion

6. 6 Factor Graphs A factor graph represents the factorization of a function of several variables Using Forney-style factor graphs

7. 7 Factor Graphs cont’d Example:

8. 8 Factor Graphs cont’d (a)Forney-style factor graph (FFG); (b) factor graph as in [3]; (c) Bayesian network; (d) Markov random field (MRF)

9. 9 Factor Graphs cont’d Advantages of FFGs: suited for hierarchical modeling compatible with standard block diagram simplest formulation of the summary- product message update rule natural setting for Forney’s result on FT and duality

10. 10 Auxiliary Variables Let Y 1 and Y 2 be two independent observations of X:

11. 11 Modularity and Special Symbols Let and with Z 1, Z 2 and X independent The “+”-nodes represent the factors and

12. 12 Outline Introduction Factor graphs Gaussian message passing in linear models Beyond Gaussians Conclusion

13. 13 Computing Marginals Assume we wish to compute For example, assume that can be written as

14. 14 Computing Marginals cont’d

15. 15 Message Passing View cont’d

16. 16 Sum-Product Rule The message out of some node/factor along some edge is formed as the product of and all incoming messages along all edges except, summed over all involved variables except

17. 17 denotes the message in the direction of the arrow denotes the message in the opposite direction Arrows and Notation for Messages

18. 18 Marginals and Output Edges

19. 19 Max-Product Rule The message out of some node/factor along some edge is formed as the product of and all incoming messages along all edges except, maximized over all involved variables except

20. 20 Message of the form: Arrow notation: / is parameterized by mean / and variance / Scalar Gaussian Message

21. 21 Scalar Gaussian Computation Rules

22. 22 Vector Gaussian Messages Message of the form: Message is parameterized either by mean vector m and covariance matrix V=W -1 or by W and Wm

23. 23 Vector Gaussian Messages cont’d Arrow notation: is parameterized by and or by and Marginal: is the Gaussian with mean and covariance matrix

24. 24 Single Edge Quantities

25. 25 Elementary Nodes

26. 26 Matrix Multiplication Node

27. 27 Composite Blocks

28. 28 Reversing a Matrix Multiplication

29. 29 Combinations

30. 30 General Linear State Space Model

31. 31 General Linear State Space Model If is nonsingular and － forward and － backward If is singular and － forward and － backward Cont’d

32. 32 General Linear State Space Model By combining the forward version with backward version, we can get Cont’d

33. 33 Gaussian to Binary

34. 34 Outline Introduction Factor graphs Gaussian message passing in linear models Beyond Gaussians Conclusion

35. 35 Message Types A key issue with all message passing algorithms is the representation of messages for continuous variables The following message types are widely applicable Quantization of continuous variables Function value and gradient List of samples

36. 36 Message Types cont’d All these message types, and many different message computation rules, can coexist in large system models SD and EM are two example of message computation rules beyond the sum-product and max-product rules

37. 37 LSSM with Unknown Vector C

38. 38 Steep Descent as Message Passing Suppose we wish to find

39. 39 Steep Descent as Message Passing Steepest descent: where s is a positive step-size parameter Cont’d

40. 40 Steep Descent as Message Passing Gradient messages: Cont’d

41. 41 Steep Descent as Message Passing Cont’d

42. 42 Outline Introduction Factor graphs Gaussian message passing in linear models Beyond Gaussians Conclusion

43. 43 Conclusion The factor graph approach to signal processing involves the following steps: 1)Choose a factor graph to represent the system model 2)Choose the message types and suitable message computation rules 3)Choose a message update schedules

44. 44 Reference [1] H.-A. Loeliger, et al., “The factor graph approach to model- based signal processing” [2] H.-A. Loeliger, “An introduction to factor graphs,” IEEE Signal Proc. Mag., Jan. 2004, pp.28-41 [3] F.R. Kschischang, B.J. Fery, and H.-A. Loeliger, “Factor graphs and the sum-product algorithm,” IEEE Trans. Inform. Theory, vol. 47, pp.498-519