An approximation for binary Markov random fields Håkon Tjelmeland and Haakon Austad Haakon.Tjelmeland@stat.ntnu.no Department of Mathematical Sciences, Norwegian University of Science and Technology An approximation for binary Markov random fields – p.1/53 Introduction Binary (hidden) Markov random field ( ) X 1 π(x|θ) = exp − VC (xC |θ) b(θ) C∈C ex: Ising model (with external field) Likelihood evaluation not possible (MLE via MCMCMLE) Direct simulation not possible (simulation possible by MCMC) Our goals: define approximate normalising constant define approximate model from which direct simulation is possible Starting point: forward-backward recursion for MRF An approximation for binary Markov random fields – p.2/53 Plan Notation Exact forward-backward recursion Approximate forward-backward recursion Evaluation criteria for approximation compare to exact results (for small lattices) compare results for different approximation levels use approximation as proposal in independent proposal Metropolis–Hastings algorithm Simulation examples. An approximation for binary Markov random fields – p.3/53 Notation Binary Markov random field on nI × nJ lattice, N = nI · nJ State vector x = (x1 , . . . , xN ), with xi ∈ {0, 1} Set of cliques: C 1 2 3 4 5 6 7 8 9 10 11 12 13 N−1 N Joint distribution ( X a(x|θ) ∝ a(x|θ) = exp − VC (xC , θ) π(x|θ) = b(θ) C∈C ) = Y e−VC (xC ,θ) C∈C Reformulation a(x|θ) = N Y ak (xk:k+p+1 , θ) k=1 p is a function of nJ and the clique size for first-order pairwise interaction model: p = nJ An approximation for binary Markov random fields – p.4/53 Exact forward recursion Following Pettitt and Reeves (2004) and Friel and Rue (2007) For illustration: Small 6 × 5 lattice, Ising model π(x1:30 |θ) = b(θ)−1 a1 (x1:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · An approximation for binary Markov random fields – p.5/53 Exact forward recursion Following Pettitt and Reeves (2004) and Friel and Rue (2007) For illustration: Small 6 × 5 lattice, Ising model π(x1:30 |θ) = b(θ)−1 a1 (x1:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · # " X −1 a1 (x1:6 , θ) a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x2:30 |θ) = b(θ) x1 An approximation for binary Markov random fields – p.6/53 Exact forward recursion Following Pettitt and Reeves (2004) and Friel and Rue (2007) For illustration: Small 6 × 5 lattice, Ising model π(x1:30 |θ) = b(θ)−1 a1 (x1:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · # " X −1 a1 (x1:6 , θ) a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x2:30 |θ) = b(θ) x1 = b(θ)−1 b2 (x2:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · An approximation for binary Markov random fields – p.7/53 Exact forward recursion Following Pettitt and Reeves (2004) and Friel and Rue (2007) For illustration: Small 6 × 5 lattice, Ising model π(x1:30 |θ) = b(θ)−1 a1 (x1:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · # " X −1 a1 (x1:6 , θ) a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x2:30 |θ) = b(θ) x1 = b(θ)−1 b2 (x2:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x3:30 |θ) = b(θ)−1 " X x2 # b2 (x2:6 , θ)a2 (x2:7 , θ) a3 (x3:8 , θ) · · · An approximation for binary Markov random fields – p.8/53 Exact forward recursion Following Pettitt and Reeves (2004) and Friel and Rue (2007) For illustration: Small 6 × 5 lattice, Ising model π(x1:30 |θ) = b(θ)−1 a1 (x1:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · # " X −1 a1 (x1:6 , θ) a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x2:30 |θ) = b(θ) x1 = b(θ)−1 b2 (x2:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x3:30 |θ) = b(θ)−1 " X x2 # b2 (x2:6 , θ)a2 (x2:7 , θ) a3 (x3:8 , θ) · · · π(x3:30 |θ) = b(θ)−1 b3 (x3:6 , θ)a3 (x3:8 , θ) · · · An approximation for binary Markov random fields – p.9/53 Exact forward recursion Following Pettitt and Reeves (2004) and Friel and Rue (2007) For illustration: Small 6 × 5 lattice, Ising model π(x1:30 |θ) = b(θ)−1 a1 (x1:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · # " X −1 a1 (x1:6 , θ) a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x2:30 |θ) = b(θ) x1 = b(θ)−1 b2 (x2:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x3:30 |θ) = b(θ)−1 " X x2 # b2 (x2:6 , θ)a2 (x2:7 , θ) a3 (x3:8 , θ) · · · π(x3:30 |θ) = b(θ)−1 b3 (x3:6 , θ)a3 (x3:8 , θ) · · · .. . π(x30 |θ) = b(θ)−1 b30 (x30 , θ)a30 (x30 , θ) An approximation for binary Markov random fields – p.10/53 Exact backward recursion Following Pettitt and Reeves (2004) and Friel and Rue (2007) For illustration: Small 6 × 5 lattice, Ising model π(x30 |θ) = b(θ)−1 b30 (x30 , θ)a30 (x30 , θ) ∝ b30 (x30 , θ)a30 (x30 , θ) An approximation for binary Markov random fields – p.11/53 Exact backward recursion Following Pettitt and Reeves (2004) and Friel and Rue (2007) For illustration: Small 6 × 5 lattice, Ising model π(x30 |θ) = b(θ)−1 b30 (x30 , θ)a30 (x30 , θ) ∝ b30 (x30 , θ)a30 (x30 , θ) π(x29 |x30 , θ) ∝ π(x29:30 |θ) = b29 (x29:30 , θ)a29 (x29:30 , θ)a30 (x30 , θ) ∝ b29 (x29:30 , θ)a29 (x29:30 , θ) An approximation for binary Markov random fields – p.12/53 Exact backward recursion Following Pettitt and Reeves (2004) and Friel and Rue (2007) For illustration: Small 6 × 5 lattice, Ising model π(x30 |θ) = b(θ)−1 b30 (x30 , θ)a30 (x30 , θ) ∝ b30 (x30 , θ)a30 (x30 , θ) π(x29 |x30 , θ) ∝ π(x29:30 |θ) = b29 (x29:30 , θ)a29 (x29:30 , θ)a30 (x30 , θ) ∝ b29 (x29:30 , θ)a29 (x29:30 , θ) π(x28 |x29:30 , θ) ∝ b28 (x28:30 , θ)a28 (x28:30 , θ) An approximation for binary Markov random fields – p.13/53 Exact backward recursion Following Pettitt and Reeves (2004) and Friel and Rue (2007) For illustration: Small 6 × 5 lattice, Ising model π(x30 |θ) = b(θ)−1 b30 (x30 , θ)a30 (x30 , θ) ∝ b30 (x30 , θ)a30 (x30 , θ) π(x29 |x30 , θ) ∝ π(x29:30 |θ) = b29 (x29:30 , θ)a29 (x29:30 , θ)a30 (x30 , θ) ∝ b29 (x29:30 , θ)a29 (x29:30 , θ) π(x28 |x29:30 , θ) ∝ b28 (x28:30 , θ)a28 (x28:30 , θ) .. . π(x2 |x2:30 , θ) ∝ b2 (x2:6 , θ)a2 (x2:7 , θ) π(x1 |x2:30 , θ) ∝ b1 (x1:5 , θ)a1 (x1:6 , θ) An approximation for binary Markov random fields – p.14/53 Exact backward recursion Following Pettitt and Reeves (2004) and Friel and Rue (2007) For illustration: Small 6 × 5 lattice, Ising model π(x30 |θ) = b(θ)−1 b30 (x30 , θ)a30 (x30 , θ) ∝ b30 (x30 , θ)a30 (x30 , θ) π(x29 |x30 , θ) ∝ π(x29:30 |θ) = b29 (x29:30 , θ)a29 (x29:30 , θ)a30 (x30 , θ) ∝ b29 (x29:30 , θ)a29 (x29:30 , θ) π(x28 |x29:30 , θ) ∝ b28 (x28:30 , θ)a28 (x28:30 , θ) .. . π(x2 |x2:30 , θ) ∝ b2 (x2:6 , θ)a2 (x2:7 , θ) π(x1 |x2:30 , θ) ∝ b1 (x1:5 , θ)a1 (x1:6 , θ) Computational complexity: forward: O(N 2p ), backward: cheap An approximation for binary Markov random fields – p.15/53 Approximate forward recursion For illustration: Small 6 × 5 lattice, Ising model π(x1:30 |θ) = b(θ)−1 a1 (x1:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · # " X −1 a1 (x1:6 , θ) a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x2:30 |θ) = b(θ) x1 = b(θ)−1 b2 (x2:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x3:30 |θ) = b(θ)−1 " X x2 # b2 (x2:6 , θ)a2 (x2:7 , θ) a3 (x3:8 , θ) · · · π(x3:30 |θ) = b(θ)−1 b3 (x3:6 , θ)a3 (x3:8 , θ) · · · .. . π(x30 |θ) = b(θ)−1 b30 (x30 , θ)a30 (x30 , θ) An approximation for binary Markov random fields – p.16/53 Approximate forward recursion For illustration: Small 6 × 5 lattice, Ising model π(x1:30 |θ) = b(θ)−1 a1 (x1:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · # " X −1 a1 (x1:6 , θ) a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x2:30 |θ) = b(θ) x1 ≈ eb(θ)−1eb2 (x2:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x3:30 |θ) = b(θ)−1 " X x2 # b2 (x2:6 , θ)a2 (x2:7 , θ) a3 (x3:8 , θ) · · · π(x3:30 |θ) = b(θ)−1 b3 (x3:6 , θ)a3 (x3:8 , θ) · · · .. . π(x30 |θ) = b(θ)−1 b30 (x30 , θ)a30 (x30 , θ) An approximation for binary Markov random fields – p.17/53 Approximate forward recursion For illustration: Small 6 × 5 lattice, Ising model π(x1:30 |θ) = b(θ)−1 a1 (x1:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · # " X −1 a1 (x1:6 , θ) a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x2:30 |θ) = b(θ) x1 ≈ eb(θ)−1eb2 (x2:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x3:30 |θ) ≈ eb(θ)−1 " X x2 # eb2 (x2:6 , θ)a2 (x2:7 , θ) a3 (x3:8 , θ) · · · π(x3:30 |θ) = b(θ)−1 b3 (x3:6 , θ)a3 (x3:8 , θ) · · · .. . π(x30 |θ) = b(θ)−1 b30 (x30 , θ)a30 (x30 , θ) An approximation for binary Markov random fields – p.18/53 Approximate forward recursion For illustration: Small 6 × 5 lattice, Ising model π(x1:30 |θ) = b(θ)−1 a1 (x1:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · # " X −1 a1 (x1:6 , θ) a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x2:30 |θ) = b(θ) x1 ≈ eb(θ)−1eb2 (x2:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x3:30 |θ) ≈ eb(θ)−1 " X x2 # eb2 (x2:6 , θ)a2 (x2:7 , θ) a3 (x3:8 , θ) · · · π(x3:30 |θ) ≈ eb(θ)−1eb3 (x3:6 , θ)a3 (x3:8 , θ) · · · .. . π(x30 |θ) = b(θ)−1 b30 (x30 , θ)a30 (x30 , θ) An approximation for binary Markov random fields – p.19/53 Approximate forward recursion For illustration: Small 6 × 5 lattice, Ising model π(x1:30 |θ) = b(θ)−1 a1 (x1:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · # " X −1 a1 (x1:6 , θ) a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x2:30 |θ) = b(θ) x1 ≈ eb(θ)−1eb2 (x2:6 , θ)a2 (x2:7 , θ)a3 (x3:8 , θ) · · · π(x3:30 |θ) ≈ eb(θ)−1 " X x2 # eb2 (x2:6 , θ)a2 (x2:7 , θ) a3 (x3:8 , θ) · · · π(x3:30 |θ) ≈ eb(θ)−1eb3 (x3:6 , θ)a3 (x3:8 , θ) · · · .. . π(x30 |θ) ≈ eb(θ)−1eb30 (x30 , θ)a30 (x30 , θ) An approximation for binary Markov random fields – p.20/53 Approximate backward recursion For illustration: Small 6 × 5 lattice, Ising model π(x30 |θ) ∝ eb30 (x30 , θ)a30 (x30 , θ) π(x29 |x30 , θ) ∝ eb29 (x29:30 , θ)a29 (x29:30 , θ) π(x28 |x29:30 , θ) ∝ eb28 (x28:30 , θ)a28 (x28:30 , θ) .. . π(x2 |x2:30 , θ) ∝ eb2 (x2:6 , θ)a2 (x2:7 , θ) π(x1 |x2:30 , θ) ∝ eb1 (x1:5 , θ)a1 (x1:6 , θ) An approximation for binary Markov random fields – p.21/53 Representation of bk (xk:k+p , θ) Exact recursive formula bk+1 (xk+1:k+p+1 , θ) = X bk (xk:k+p , θ)ak (xk:k+p+1 , θ) xk Number of values to store for each k: 2p Represent bk (xk:k+p , θ) as ln {bk (xk:k+p , θ)} = βk∅ + k+p X {i} βk xi + i=k i=k + k+p−2 X k+p−1 X i=k k+p X {i,j,t} k+p−1 X k+p X xi xj xt + . . . + βk = X xi xj j=i+1 {k,...,k+p} βk {i,j} βk xk · . . . · xk+p j=i+1 t=j+1 Λ⊆{k,...,k+p} βkΛ Y xi i∈Λ An approximation for binary Markov random fields – p.22/53 Ising example Ising model — small 20 × 20 lattice, i.e. p = 20 red: β = 0.6 black: β = 0.8 3. and 4. 5. and 6. 50 100 150 200 −0.005 0 2000 4000 6000 −0.015 0 −0.04 −0.5 0.00 0.0 0.04 0.5 0.005 1. and 2. 0 20000 40000 An approximation for binary Markov random fields – p.23/53 Ising example Ising model — small 20 × 20 lattice, i.e. p = 20 red: β = 0.6 black: β = 0.8 3. and 4. 5. and 6. 50 100 150 200 −0.005 0 2000 4000 6000 −0.015 0 −0.04 −0.5 0.00 0.0 0.04 0.5 0.005 1. and 2. 0 20000 40000 β = 0.8: fraction of interactions larger than ε ε 10−1 10−2 10−3 10−4 10−5 fraction 0.000039 0.000169 0.000906 0.00556 0.0314 An approximation for binary Markov random fields – p.24/53 Forward recursion for new representation Correspondingly we represent X ln {ak (xk:k+p+1 , θ)} = αkΛ Y xi i∈Λ Λ⊆{k,...,k+p+1} Old recursion formula bk+1 (xk+1:k+p+1 , θ) = X bk (xk:k+p , θ)ak (xk:k+p+1 , θ) xk Recursion formula for new representation X Y Λ exp βk+1 xi i∈Λ Λ⊆{k+1,...,k+p+1} X X Y Λ exp βk = xi + x k Λ⊆{k,...,k+p} i∈Λ X Λ⊆{k,...,k+p+1} αkΛ Y i∈Λ xi An approximation for binary Markov random fields – p.25/53 Λ Recursive computation of βk+1 Setting xk+1 = . . . = xk+p+1 = 0 gives n ∅ exp βk+1 o o n o n {k} {k} = exp βk∅ + αk∅ + exp βk∅ + βk + αk∅ + αk An approximation for binary Markov random fields – p.26/53 Λ Recursive computation of βk+1 Setting xk+1 = . . . = xk+p+1 = 0 gives n ∅ exp βk+1 o o n o n {k} {k} = exp βk∅ + αk∅ + exp βk∅ + βk + αk∅ + αk Setting xi = 1 and xj = 0 for all j 6= i gives exp n ∅ βk+1 + {i} βk+1 o o n {i} {i} ∅ ∅ = exp βk + βk + αk + αk n {i} + exp βk∅ + βk {k} + βk {k,i} + βk {i} {k} + αk∅ + αk + αk {k,i} + αk An approximation for binary Markov random fields – p.27/53 Λ Recursive computation of βk+1 Setting xk+1 = . . . = xk+p+1 = 0 gives n ∅ exp βk+1 o o n o n {k} {k} = exp βk∅ + αk∅ + exp βk∅ + βk + αk∅ + αk Setting xi = 1 and xj = 0 for all j 6= i gives exp n ∅ βk+1 + {i} βk+1 o o n {i} {i} ∅ ∅ = exp βk + βk + αk + αk n {i} + exp βk∅ + βk {k} + βk {k,i} + βk {i} {k} + αk∅ + αk + αk {k,i} + αk Setting xi = xj = 1 and xt = 0 for t 6∈ {i, j} gives n ∅ exp βk+1 + {i} βk+1 + {j} βk+1 + {i,j} βk+1 o = ... An approximation for binary Markov random fields – p.28/53 Recursive scheme ln {bk (xk:k+p , θ)} = X Λ⊆{k,...,k+p} βkΛ Y xi i∈Λ 0123 Q A Q A QQ Q A AU s Q + 012 013 023 123 HH H H H H HH H @ H H H HH HH@ HH ? ? HH ? H@ R @ HH H j H j H j H 13 23 03 12 01 02 HH HH HH @ H H H HH H @ HH H HH HH HH @ ? ? ? R H @ H H j H j H j H 0 1 2 3 Q Q A Q Q A Q A sAU Q + ∅ An approximation for binary Markov random fields – p.29/53 Approximation strategy If a computed βkΛ has |βkΛ | ≤ ε, approximate βkΛ to zero If for some Λ, βkλ is (approximated to) zero for all λ ⊂ Λ, |λ| = |Λ| − 1, approximate βkΛ to zero (without computing the exact value) 0123 Q A Q A QQ Q A AU + s Q 012 013 023 123 HH H H H H HH H @ H H H HH HH@ HH ? ? HH ? H@ R @ HH H j H j H j H 01 02 13 23 03 12 HH HH HH @ H HH HH H @ HH HH HH @ H ? H ? R ? HH @ H j H j H H j H 0 1 2 3 Q Q A Q Q A Q A sAU Q + ∅ An approximation for binary Markov random fields – p.30/53 Approximation strategy If a computed βkΛ has |βkΛ | ≤ ε, approximate βkΛ to zero If for some Λ, βkλ is (approximated to) zero for all λ ⊂ Λ, |λ| = |Λ| − 1, approximate βkΛ to zero (without computing the exact value) 0123 Q A Q A QQ Q A AU + s Q 012 013 023 123 H HH H HH H H H @ H H H HH@ HH HH ? ? HH ? H@ HH R @ H j H j H j H 02 03 12 13 23 01 HH HH HH @ H HH HH H @ HH HH HH @ H ? H ? R ? HH @ H j H H j H j H 0 1 2 3 Q Q A Q Q A Q A sAU Q + ∅ An approximation for binary Markov random fields – p.31/53 Evaluation criteria for approximation Compare approximation with exact result — for small lattices Compare results for different cut off values ε Use approximation as proposal in an independent proposal Metropolis–Hastings algorithm — look at acceptance rate An approximation for binary Markov random fields – p.32/53 Comparison to exact results in a small lattice Ising model, small 20 × 20 lattice Simulate first a realisation from the Ising model Compute the likelihood function l(β)for this realisation Results for ε = 0, 10−5 , 10−4 , 10−3 , 10−2 and 10−1 β = 0.8 β = 0.9 230 200 160 150 200 250 240 250 270 280 β = 0.6 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 An approximation for binary Markov random fields – p.33/53 Comparison to exact results in a small lattice Ising model, small 20 × 20 lattice Simulate first a realisation from the Ising model Compute the likelihood function l(β)for this realisation Results for ε = 0, 10−5 , 10−4 , 10−3 , 10−2 and 10−1 β = 0.8 β = 0.9 230 200 160 150 200 250 240 250 270 280 β = 0.6 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 An approximation for binary Markov random fields – p.34/53 Comparison to exact results in a small lattice Ising model, small 20 × 20 lattice Simulate first a realisation from the Ising model Compute the likelihood function l(β)for this realisation Results for ε = 0, 10−5 , 10−4 , 10−3 , 10−2 and 10−1 β = 0.8 β = 0.9 230 200 160 150 200 250 240 250 270 280 β = 0.6 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 An approximation for binary Markov random fields – p.35/53 Comparison to exact results in a small lattice Ising model, small 20 × 20 lattice Simulate first a realisation from the Ising model Compute the likelihood function l(β)for this realisation Results for ε = 0, 10−5 , 10−4 , 10−3 , 10−2 and 10−1 β = 0.8 β = 0.9 230 200 160 150 200 250 240 250 270 280 β = 0.6 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 An approximation for binary Markov random fields – p.36/53 Comparison to exact results in a small lattice Ising model, small 20 × 20 lattice Simulate first a realisation from the Ising model Compute the likelihood function l(β)for this realisation Results for ε = 0, 10−5 , 10−4 , 10−3 , 10−2 and 10−1 β = 0.8 β = 0.9 230 200 160 150 200 250 240 250 270 280 β = 0.6 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 An approximation for binary Markov random fields – p.37/53 Comparison to exact results in a small lattice Ising model, small 20 × 20 lattice Simulate first a realisation from the Ising model Compute the likelihood function l(β)for this realisation Results for ε = 0, 10−5 , 10−4 , 10−3 , 10−2 and 10−1 β = 0.8 β = 0.9 230 200 160 150 200 250 240 250 270 280 β = 0.6 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 An approximation for binary Markov random fields – p.38/53 Compare results for different cut off value ε Ising model, 50 × 50 lattice Simulate first a realisation from the Ising model Compute the likelihood function l(β) for this realisation Results for ε = 10−5 , 10−4 , 10−3 , 10−2 and 10−1 β = 0.8 1450 1100 1300 1550 1500 1650 1700 β = 0.6 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 An approximation for binary Markov random fields – p.39/53 Compare results for different cut off value ε Ising model, 50 × 50 lattice Simulate first a realisation from the Ising model Compute the likelihood function l(β) for this realisation Results for ε = 10−5 , 10−4 , 10−3 , 10−2 and 10−1 β = 0.8 1450 1100 1300 1550 1500 1650 1700 β = 0.6 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 An approximation for binary Markov random fields – p.40/53 Compare results for different cut off value ε Ising model, 50 × 50 lattice Simulate first a realisation from the Ising model Compute the likelihood function l(β) for this realisation Results for ε = 10−5 , 10−4 , 10−3 , 10−2 and 10−1 β = 0.8 1450 1100 1300 1550 1500 1650 1700 β = 0.6 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 An approximation for binary Markov random fields – p.41/53 Compare results for different cut off value ε Ising model, 50 × 50 lattice Simulate first a realisation from the Ising model Compute the likelihood function l(β) for this realisation Results for ε = 10−5 , 10−4 , 10−3 , 10−2 and 10−1 β = 0.8 1450 1100 1300 1550 1500 1650 1700 β = 0.6 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 An approximation for binary Markov random fields – p.42/53 Compare results for different cut off value ε Ising model, 50 × 50 lattice Simulate first a realisation from the Ising model Compute the likelihood function l(β) for this realisation Results for ε = 10−5 , 10−4 , 10−3 , 10−2 and 10−1 β = 0.8 1450 1100 1300 1550 1500 1650 1700 β = 0.6 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 An approximation for binary Markov random fields – p.43/53 Compare results for different cut off value ε Ising model, 50 × 50 lattice Simulate first a realisation from the Ising model Compute the likelihood function l(β) for this realisation Results for ε = 10−5 , 10−4 , 10−3 , 10−2 and 10−1 β = 0.8 1450 1100 1300 1550 1500 1650 1700 β = 0.6 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 Fraction of interactions computed for ε = 10−3 β 0.1 0.3 0.5 0.7 0.9 fraction 2.1 · 10−13 5.0 · 10−12 7.0 · 10−12 3.3 · 10−11 1.5 · 10−10 An approximation for binary Markov random fields – p.44/53 Used as proposal in a Metropolis–Hastings algorithm Ising model, 50 × 50 lattice Initiate Metropolis–Hastings algorithm with exact sample Run for ε = 10−1 , 10−2 , 10−3 , 10−4 and 10−5 0.0 0.2 0.4 0.6 0.8 1.0 Monitor the Metropolis–Hastings acceptance rate 0.0 0.2 0.4 0.6 0.8 An approximation for binary Markov random fields – p.45/53 Image analysis example Ising prior, 50 × 50 lattice Conditionally independent Gaussian observations, yi |xi ∼ N(xi , σ 2 ) Simulated data (true parameter values β = 0.8, σ = 0.5) Compute marginal likelihood function l(θ) = π(y|θ) = π(x|θ)π(y|x, θ) π(x|y, θ) An approximation for binary Markov random fields – p.46/53 Image analysis example (cont.) ε = 10−2 ε = 10−3 ratio 5000 0.8 5000 0.8 4500 0.8 1.006 1.004 4500 0.6 4000 0.6 4000 0.6 0.4 3500 0.4 3500 0.4 3000 0.2 0.2 0.4 0.6 0.8 1.0 1.000 3000 0.2 2500 0.0 0.2 1.002 0.998 2500 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 An approximation for binary Markov random fields – p.47/53 Image analysis example (cont.) 0.90 ε = 10−3 2400 0.85 2380 2360 0.80 2340 0.75 0.70 0.40 2320 2300 0.45 0.50 0.55 0.60 An approximation for binary Markov random fields – p.48/53 Example with 3 × 3 clique 3 × 3 clique model, pairwise interaction only, 100 × 100 lattice Realisations from specified model, by MCMC 0.0 0.2 0.4 0.6 0.8 1.0 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.8 0.6 0.4 0.2 0.0 0.0 0.2 0.4 0.6 0.8 1.0 Realisations from approximate model, ε = 10−3 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 An approximation for binary Markov random fields – p.49/53 Problematic example 4 × 4 clique, strong higher order interactions Realisations from the model, by MCMC Too few interactions are approximated to zero — not able to run the approximation An approximation for binary Markov random fields – p.50/53 A closer look at the approximate model Approximate model is π e(x1:n |θ) = π e(xn |θ)e π (xn−1 |xn , θ)e π (xn−2 |xn−1:n , θ) · . . . · π e(x1 |x2:n , θ) Conditionally independence because: original MRF approximation An approximation for binary Markov random fields – p.51/53 A closer look at the approximate model (cont.) ε = 10−2 ε = 10−4 β = 0.4 β = 0.8 An approximation for binary Markov random fields – p.52/53 Closing remarks For MRFs we have defined approximation to the normalising constant — that (often) can be computed the joint distribution — that (often) can be simulated directly The approximation is based on representation of bk (xk:k+p , θ) by interaction parameters most interaction parameters must be small have assumed higher order interactions to be small whenever lower order interactions are small We have demonstrated the quality of the approximation in a number of examples The approximative model is a Markov mesh model (and POMM) Can sum out the variables in a different order An approximation for binary Markov random fields – p.53/53