Planar and surface graphical models which are easy Vladimir Chernyak and Michael Chertkov Department of Chemistry, Wayne State University Department of Mathematics, Wayne State University Theory Division, Los Alamos National Laboratory Acknowledgements John Klein (Wayne State University) Martin Loebl (Charles University) Outline • Forney-style graphical models: generically hard • A family of graphical models that turn out to be easy • Gauge invariance and linear algebra: traces and graphical traces • Calculus of Grassman (anticommuting) variables: determinants and Pfaffians • Topological invariants of immersions: equivalence between certain binary and Gaussain Grassman (fermion) models • Planar versus surface easy Forney-style graphical model formulation q-ary variables reside on edges Forney ’01; Loeliger ‘01 Generalization: continuous/Grassman variables Probability of a configuration Marginal probabilities can be expressed in terms of the derivatives of the free energy with respect to factor-functions Partition function Reduced variables Equivalent models: gauge fixing and transformations Replace the model with an equivalent more convenient model Invariant approach (i) Introduce an invariant object that describes partition function Z (ii) Different equivalent models correspond to different coordinate choices (gauge fixing) (iii) Gauge transformations are changing the basis sets Coordinate approach (i) Introduce a set of gauge transformations that do not change Z (ii) Gauge transformations build new equivalent models General strategy (based on linear algebra) (i) Replace q-ary alphabet with a q-dimensional /functional vector space (ii) (letters are basis vectors) (ii) Represent Z by an invariant object graphical trace (iii) Gauge fixing is a basis set choice (iv) Gauge transformations are linear transformation of basis sets Gauge invariance: matrix formulation Gauge transformations of factor-functions with orthogonality conditions do not change the partition function Graphical representation of trace and cyclic trace Trace f ij g ji scalar product Tr( f ) f j j fij g ji Summation over repeating subscripts/superscripts f in jn Cyclic trace g jni1 f i1 j1 g j1i2 g j2i3 f i2 j2 Tr( f n ) f j1j2 f j2j3 ... f jnj1 fi1 j1 g j1i2 fi2 j2 g j2i3 ... fin jn g jni1 Graphical trace and partition function Collection of tensors (poly-vectors) f a a1a2 a3 Wab g f bb1b2 a 3b1 ab f { f a }a 1,..., N { f aa1 ... an }a 1,... N a ia ib g {g ab}ab {g ab }ab Wba Graphic trace g ak b j ab Tr( f ) Trg ( f a ) (... f aa1 ... ak ... ana ... f bb1 ... b j ... bnb ...) a Scalar products u w gab (u w) u Wab w Wba Orthogonality condition ij i i g ab g ab (eab ebaj ) eab ebaj ij Tensors and factor-functions fa n 1 f ( ,..., ) e ... e a 1 n ab1 abn 1 ... n Z Tr ( f ) Partition function and graphical trace: gauge invariance Dual basis set of co-vectors (elements of the dual space) ab,i (eabj ) i j ab Wab* Orthogonality condition (two equivalent forms) ab,i ba, j ij g ab ab, j ba , j j Graphic trace: Evaluate scalar products (reside on edges) on tensors (reside vertices) sb esb Tr( f ) f a ( ab , ac , ad ) f b ( ba , bs )... ab, ba, ac ac e ab ad ad e ab eab ab bs ebs ba eba Gauge invariance: graphic trace is an invariant object, factor-functions are basis-set dependent f a (1,..., n ) ab1 ,1 ... abn , n ( f a ) “Gauge fixing” is a choice of an orthogonal basis set Supersymmetric sigma-models: calculus of Grassman variables and supermanifolds dimension substrate (usual) manifold additional Grassman (anticommuting variables) Functions on a supermanifold Berezin integral (measure in a supermanifold) Any function on a supermanifold can be represented as a sum of its even and odd components Gaussian Grassman models: determinants and Pfaffians Two key formulas of Gaussain calculus: det( A) j d j d j exp ik Aikik Measure is invariant Pfaff ( A) j d j exp ik Aikik Requires an ordering choice in the measure Topological invariants of immersions and SmaleHirsch-Gromov (SHG) Theorem: planar case Loops in phase space Immersions Topologically (homotopy) equivalent Number of self-intersections Number of rotations over 2 (spinor structure) Topological invariants of immersions (II) : planar case Generalization to a set of immersions (equality modulo 2) Number of (self) intersections Effects of orientations (sign factor) Spinor structures are the ways to execute a “square root”. In the planar case unique Topological invariants of immersions: Riemann surface case 22 g spinor structures (ways to square root) for a Riemann surface of genus g Value of the invariant (quadratic form) Effects of orientation Total number of intersections Edge Binary Wick (EBW) models [VC, Chertkov 09] Main Theorem [VC, Chertkov 09/surface] Main “take home” message Summary • We have identified a family of graphical models that are planar/surface easy • We have established two ingredients that make these models easy • Ingredient #1: Invariant nature of linear algebra(traces and graphical traces) • Ingredient #2: Topological invariants of immersions (intersections, self-intersections and spinor structures) • Question (and maybe path forward): Can the easy family be extended? Belief propagation gauge and BP equations Continuous and supersymmetric case: graphical sigma-models M M M M M ab M ba Scalar product: the space of states and its dual are equivalent No-loose-end requirement Continuous version of BP equations Supersymmetric sigma-models: supermanifolds dimension substrate (usual) manifold additional Grassman (anticommuting variables) Functions on a supermanifold Berezin integral (measure in a supermanifold) Any function on a supermanifold can be represented as a sum of its even and odd components Supersymmetric sigma models: graphic supertrace I Natural assumption: factor-functions are even functions on Introduce parities of the beliefs pba pab BP equations for parities pab 0,1 Follows from the first two Edge parity is well-defined elements Euler characteristic (number of connected components) B1 B0 card ( E ) card (V ) Supersymmetric sigma models: graphic supertrace II Decompose the vector spaces into reduced vector spaces Graphic supertrace decomposition (generalizes the supertrace) results in a multi-reference loop expansion is the graphic trace (partition function) of a reduced model