@ 15/7/2003 Tokyo Institute of Technology 1 Propagating beliefs in spin- glass models Yoshiyuki Kabashima Dept. of Compt. Intel. & Syst. Sci. Tokyo Institute of Technology
Tokyo Institute of Technology2 Background and Motivation Active research on belief propagation (BP) in information sciences (IS) Similarity to methods in physics TMM & Bethe approx. Difference in interest Physics ⇒ obtained solutions IS ⇒ dynamics of the algorithm may cause unexpected developments in both fields
Tokyo Institute of Technology3 Purpose and Results Purpose : analyze dynamics of BP when employed in spin-glass models Results: Macro. dyn. of BP ⇒ RS solution Micro. stability of BP ⇔ AT condition
Tokyo Institute of Technology4 Outline SG model in Bayesian framework Belief propagation Macro. dyn. and RS solution Micro. stability and AT condition Summary
Tokyo Institute of Technology5 Spin-glass models SG models on a random (Bethe) lattice K-body interaction C-bonds/spin Randomly constructed for other aspects
Tokyo Institute of Technology6 SM and Bayesian Statistics Boltzmann dist. = Bayes formula Magnetization = Posterior average
Tokyo Institute of Technology7 Graph Expression Expression by a bipartite graph … … …
Tokyo Institute of Technology8 Belief Propagation Iterative inference by passing beliefs … … …
Tokyo Institute of Technology9 More Precisely : Posterior average when is left out. : Effective field when comes in. : Estimator
Tokyo Institute of Technology10 Macro. Dyn. vs. RS Solution Distribution (histogram) of beliefs Known result for finite C: Tree approximation (resampling graph/update) Density evolution ⇒ RS solution Vicente, Saad, YK (2000)Richardson & Urbanke (2001)
Tokyo Institute of Technology11 Novel Result for Infinite C Central limit theorem for infinite C Evolution of average and variance Natural iteration of RS SP eqs.
Tokyo Institute of Technology12 Experimental Validation SK model ( N=1000,J=1,T=0.5 ) (AT stable) (AT unstable)
Tokyo Institute of Technology13 Microscopic Instability Possible microscopic instability while BP seems to macroscopically converge Stability analysis of the fixed point
Tokyo Institute of Technology14 Evolution of Perturbation Dist. of Perturb. Perturbation Evolution is attractive ⇔ the fixed point(=RS solution) is stable
Tokyo Institute of Technology15 Pictorial Expression What is performed? … … … … … … … … …
Tokyo Institute of Technology16 Meaning of P. Evolution Link to known results for infinite C Central limit theorem : P. Evolution → Update of average & variance :Gaussian dist.
Tokyo Institute of Technology17 Meaning of P. Evolution Critical conditions for growth of fluctuation For K=2 (SK model) Average → T f : Para-Ferro transition Variance → T AT : AT condition :Average :Variance
Tokyo Institute of Technology18 Analysis for finite C Is P. evolution equivalent to AT analysis even for finite C? AT analysis for finite C is complicated. But, P. evolution is (numerically) possible. K=2 (Wong-Sherrington model) Paramagnetic solution Average → T f Variance → T AT Known result Klein et al (1979) Mezard & Parisi (1987)
Tokyo Institute of Technology19 Analysis for Finite C Ferromagnetic solution Numerical evaluation of T pevol : New result! N=2000, K=2, C=4, 20000MCS/Spin : T pevol in Ferro phase
Tokyo Institute of Technology20 Summary Close relationship between BP and the replica analysis Macro. dyn. ⇒ RS solution Micro. stability ⇔ AT condition This correspondence may be useful for AT analysis for SG models of finite connectivity. Application to CDMA multiuser detection (Kabashima, to appear in J. Phys. A, 2003)