approximating the partition function of the ferromagnetic...
TRANSCRIPT
![Page 1: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/1.jpg)
Approximating the partition function
of the ferromagnetic Ising model
Mark JerrumSchool of Mathematical Sciences
Queen Mary, University of London
Describing joint work with Leslie Ann Goldberg (Liverpool)and Alistair Sinclair (Berkeley)
Counting, Inference and Optimization on GraphsPrinceton, 2nd-5th November 2011
![Page 2: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/2.jpg)
The Ising model
Let G = (V ,E ) be a graph, with indeterminatesγ = (γe : e ∈ E ) associated to the edges.
Definition
The Ising partition function is
ZIsing(G ; γ) =∑
σ:V→{0,1}
∏e∈E
(1 + γeδe(σ)
),
where δe(σ) is 1 if σ assigns the same value to the twoendpoints of e, and 0 otherwise.
We are interested in computing the partition function in theferromagnetic case, which corresponds to evaluating thepolynomial ZIsing(G ; γ) in the positive orthant, γ ≥ 0.
![Page 3: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/3.jpg)
Some remarks
In the q-state Potts model, a configuration is a functionV → {0, 1, . . . , q − 1}. The Ising model is the specialcase q = 2. Leslie Goldberg will consider the generalPotts model later. For the time being, we stick to the2-spin situation.
It is #P-hard to compute ZIsing(G ; γ) exactly, given Gand an assignment to γ.
It is NP-hard even to approximate ZIsing(G ; γ) in thenon-ferromagnetic case, corresponding to γ ≥ −1.
![Page 4: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/4.jpg)
Some remarks
In the q-state Potts model, a configuration is a functionV → {0, 1, . . . , q − 1}. The Ising model is the specialcase q = 2. Leslie Goldberg will consider the generalPotts model later. For the time being, we stick to the2-spin situation.
It is #P-hard to compute ZIsing(G ; γ) exactly, given Gand an assignment to γ.
It is NP-hard even to approximate ZIsing(G ; γ) in thenon-ferromagnetic case, corresponding to γ ≥ −1.
![Page 5: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/5.jpg)
Some remarks
In the q-state Potts model, a configuration is a functionV → {0, 1, . . . , q − 1}. The Ising model is the specialcase q = 2. Leslie Goldberg will consider the generalPotts model later. For the time being, we stick to the2-spin situation.
It is #P-hard to compute ZIsing(G ; γ) exactly, given Gand an assignment to γ.
It is NP-hard even to approximate ZIsing(G ; γ) in thenon-ferromagnetic case, corresponding to γ ≥ −1.
![Page 6: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/6.jpg)
Approximate computation: FPRAS
Definition
An FPRAS is a randomised algorithm that produces a resultthat is correct to within relative error 1± ε with highprobability. It must run in time poly(n, ε−1), where n is theinput size.
Theorem (Jerrum & Sinclair 1990)
There is an FPRAS for ZIsing(G ; γ) in the ferromagnetic region(γ ≥ 0).
(Alternatively: there is an FPRAS for the Tutte polynomialalong the positive branch of the hyperbola defined by q = 2.)
![Page 7: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/7.jpg)
Approximate computation: FPRAS
Definition
An FPRAS is a randomised algorithm that produces a resultthat is correct to within relative error 1± ε with highprobability. It must run in time poly(n, ε−1), where n is theinput size.
Theorem (Jerrum & Sinclair 1990)
There is an FPRAS for ZIsing(G ; γ) in the ferromagnetic region(γ ≥ 0).
(Alternatively: there is an FPRAS for the Tutte polynomialalong the positive branch of the hyperbola defined by q = 2.)
![Page 8: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/8.jpg)
Attempt 1: direct simulation
The FPRAS for the ferromagnetic Ising partition function is anexample of the Markov chain Monte Carlo (MCMC) method,but a direct application, based on single-site updates clearlyfails. The following simulation illustrates the point:
Demo
(Acknowledgements to Bernd Nottelmann and Peter Young.)
We can say that the Markov chain is not “rapidly mixing”.
![Page 9: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/9.jpg)
Attempt 2: expansion in terms of even subgraphs
Call an edge subset A even if every vertex in (V ,A) has evendegree.
Then we have the following alternative “high temperature”expansion of the partition function:
ZIsing(G ; γ) = 2|V |∏e∈E
w ′e∑A⊆E
A even
∏e∈A
we
where w ′e = (γe + 2)/2 and we = γe/(γe + 2).
The Markov chain based on single-edge updates of evensubsets (and defective even subsets) is rapidly mixing. . .yielding an FPRAS.
![Page 10: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/10.jpg)
Extension to “consistent” external field
Now suppose there is a multiplicative weight 1 + µvσ(v) ateach vertex v . There is again a high-temperature expansion:
ZIsing(G ; γ,µ) = 2|V | · · ·∑A⊆E
∏e∈A
we
∏v∈V
degA(v) odd
zv
where zv = µv/(µv + 2) and we is as before. (An easilycomputable product of weights has been omitted.)
A slight modification of the earlier Markov chain withsingle-edge updates works here also, provided µv ≥ 0 (orµv ≤ 0) for all v ∈ V .
But what if the field is inconsistent, i.e., µv takes negativevalues as well as positive?
![Page 11: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/11.jpg)
Interlude: an interesting class of counting problems
Denote by #BIS the problem of counting independent sets ina bipartite graph.
Fact (Dyer, Goldberg, Greenhill & Jerrum, 2000)
#BIS is inter-reducible — in an approximation-preservingsense — with several other counting problems (e.g., downsetsin a partial order, stable matchings, Widom-Rowlinson modelin statistical physics).
A class of sampling problems of intermediate computationalcomplexity or an illusion?
![Page 12: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/12.jpg)
A logically defined complexity class
The complexity class, #RHΠ1, containing “BipartiteIndependent Set” and its peers is characterised by syntacticallyrestricted sentences in first order logic. In fact, #BIS iscomplete for this class with respect toapproximation-preserving reducibility. (C.f. “restricted KromSNP”/“Linear Datalog”.)
E.g., the set of downsets in a partial order (A,≺) may beexpressed as{
D : ∀x , y ∈ A. ¬D(x) ∨ ¬(y ≺ x) ∨ D(y)}.
![Page 13: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/13.jpg)
A logically defined complexity class
The complexity class, #RHΠ1, containing “BipartiteIndependent Set” and its peers is characterised by syntacticallyrestricted sentences in first order logic. In fact, #BIS iscomplete for this class with respect toapproximation-preserving reducibility. (C.f. “restricted KromSNP”/“Linear Datalog”.)
E.g., the set of downsets in a partial order (A,≺) may beexpressed as{
D : ∀x , y ∈ A. ¬D(x) ∨ ¬(y ≺ x) ∨ D(y)}.
(≡ ¬(D(x) ∧ (y ≺ x) ∧ ¬D(y)
))
![Page 14: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/14.jpg)
A logically defined complexity class
The complexity class, #RHΠ1, containing “BipartiteIndependent Set” and its peers is characterised by syntacticallyrestricted sentences in first order logic. In fact, #BIS iscomplete for this class with respect toapproximation-preserving reducibility. (C.f. “restricted KromSNP”/“Linear Datalog”.)
E.g., the set of downsets in a partial order (A,≺) may beexpressed as{
D : ∀x , y ∈ A. ¬D(x) ∨ ¬(y ≺ x) ∨ D(y)}.
First order universal quantification.
![Page 15: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/15.jpg)
A logically defined complexity class
The complexity class, #RHΠ1, containing “BipartiteIndependent Set” and its peers is characterised by syntacticallyrestricted sentences in first order logic. In fact, #BIS iscomplete for this class with respect toapproximation-preserving reducibility. (C.f. “restricted KromSNP”/“Linear Datalog”.)
E.g., the set of downsets in a partial order (A,≺) may beexpressed as{
D : ∀x , y ∈ A. ¬D(x) ∨ ¬(y ≺ x) ∨ D(y)}.
CNF. (Only one clause in this instance!)
![Page 16: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/16.jpg)
A logically defined complexity class
The complexity class, #RHΠ1, containing “BipartiteIndependent Set” and its peers is characterised by syntacticallyrestricted sentences in first order logic. In fact, #BIS iscomplete for this class with respect toapproximation-preserving reducibility. (C.f. “restricted KromSNP”/“Linear Datalog”.)
E.g., the set of downsets in a partial order (A,≺) may beexpressed as{
D : ∀x , y ∈ A. ¬D(x) ∨ ¬(y ≺ x) ∨ D(y)}.
Each clause has at most one unnegated relation symbol and atmost one negated relation symbol.
![Page 17: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/17.jpg)
Some restrictions of the Ising model
Zero external field: µv = 0, for all v ∈ V .
Consistent external field: µv ≥ 0, for all v(or µv ≤ 0, for all v).
Ferromagnetic: γe ≥ 0, for all e ∈ E .
![Page 18: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/18.jpg)
Computational complexity of some variants
Ising partition function Exact Approximate
Zero field, planar FP1 -Ferromagnetic, consistent field #P-hard FPRAS2
Ferromagnetic, general field #P-hard ?Antiferromagnetic/spinglass #P-hard NP-hard3
Notes
![Page 19: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/19.jpg)
Computational complexity of some variants
Ising partition function Exact Approximate
Zero field, planar FP1 -Ferromagnetic, consistent field #P-hard FPRAS2
Ferromagnetic, general field #P-hard ?Antiferromagnetic/spinglass #P-hard NP-hard3
Notes1Reduction to dimer coverings (perfect matchings) [Fisher].
![Page 20: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/20.jpg)
Computational complexity of some variants
Ising partition function Exact Approximate
Zero field, planar FP1 -Ferromagnetic, consistent field #P-hard FPRAS2
Ferromagnetic, general field #P-hard ?Antiferromagnetic/spinglass #P-hard NP-hard3
Notes2We saw this earlier.
![Page 21: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/21.jpg)
Computational complexity of some variants
Ising partition function Exact Approximate
Zero field, planar FP1 -Ferromagnetic, consistent field #P-hard FPRAS2
Ferromagnetic, general field #P-hard ?Antiferromagnetic/spinglass #P-hard NP-hard3
Notes3Essentially Max Cut.
![Page 22: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/22.jpg)
General external field: idea for a reduction
It is possible to encode an instance of #BIS as an instance ofthe ferromagnetic Ising model with general field:
1 means “OUT”;
0 means “IN”;
1 + µu =(1 + γ)deg(u).
u
v
γe = γ > 0
1 means “IN”;
0 means “OUT”;
1 + µv =(1 + γ)−deg(v).
![Page 23: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/23.jpg)
Calculation
σ(u) – σ(v) Contribution Equals
IN – IN 1× 1× (1 + γ)−1 (1 + γ)−1
IN – OUT 1× (1 + γ)× 1 1 + γOUT – IN (1 + γ)× (1 + γ)× (1 + γ)−1 1 + γ
OUT – OUT (1 + γ)× 1× 1 1 + γ
Fleshing out the details (and also doing the reduction in theother direction) yields:
Theorem (Goldberg and Jerrum, 2007)
Computing the partition function of a ferromagnetic Isingmodel with a general fields is equivalent to #BIS underapproximation-preserving reductions.
![Page 24: Approximating the partition function of the ferromagnetic ...people.csail.mit.edu/andyd/CIOG_slides/jerrum_talk_ciog2011.pdfThe Ising model Let G = (V;E) be a graph, with indeterminates](https://reader033.vdocuments.site/reader033/viewer/2022060801/60850d18d8d099705c5a6140/html5/thumbnails/24.jpg)
Complexity of some variants (reprise)
Ising partition function Exact Approximate
Zero field, planar FP -Ferromagnetic, consistent field #P-hard FPRASFerromagnetic, general field #P-hard #BIS-equivalentAntiferromagnetic/spinglass #P-hard NP-hard