Vector representations of graphs László Lovász Microsoft Research One Microsoft Way, Redmond, WA 98052 lovasz@microsoft.com Cim
Every planar graph can be drawn in the plane with straight edges Fáry-Wagner
3-connected planar graph Steinitz Every 3-connected planar graph is the skeleton of a polytope.
Rubber bands and planarity Tutte (1963) G: 3-connected planar graph outer face fixed to convex polygon edges replaced by rubber bands Energy: Equilibrium:
rubber band embedding is planar G 3-connected planar rubber band embedding is planar (Easily) polynomial time computable Lifts to polyhedral representation Maxwell Tutte
Largest has multiplicity 1. The Colin de Verdière number G: connected graph Roughly: multiplicity of second largest eigenvalue of adjacency matrix Largest has multiplicity 1. But: maximize over weighting the edges and diagonal entries But: non-degeneracy condition on weightings
Strong Arnold Property M=(Mij): symmetric VxV matrix Mij <0, if ijE 0, if Mii arbitrary normalization M has =1 negative eigenvalue symmetric, X=0 Strong Arnold Property
Basic Properties μk is polynomial time decidable for fixed k μ(G) is minor monotone deleting and contracting edges for μ>2, μ(G) is invariant under subdivision for μ>3, μ(G) is invariant under Δ-Y transformation
Special values rank=1
Special values Violates the Strong Arnold property! rank=2
Special values μ(G)1 G is a path non-singular
Special values μ(G)1 G is a path μ(G)2 G is outerplanar μ(G)3 G is a planar Colin de Verdière, using pde’s Van der Holst, elementary proof μ(G)4 G is linklessly embeddable in 3-space … μ(G)n-4 complement G is planar _ ~ Kotlov-L-Vempala
homological, homotopical,… Linklessly embeddable graphs embeddable in R3 without linked cycles homological, homotopical,… equivalent Apex graph
Basic facts about linklessly embeddable graphs Closed under: - subdivision - minor - Δ-Y and Y- Δ transformations G linklessly embeddable G has no minor in the “Petersen family” Robertson – Seymour - Thomas
The Petersen family (graphs arising from K6 by Δ-Y and Y- Δ)
G is linklessly embeddable : follows from Robertson-Seymour-Thomas L-Schrijver
Nullspace representation basis of nullspace of M Representation of G in Rd
Van der Holst’s Lemma connected or… like convex polytopes?
Linked Borsuk Theorem P R5 convex polytope A,B: faces of P A, B opposite: parallel supporting hyperplanes H, H’ such that A H, B H’.
Linked Borsuk Theorem embedding φ: P1R3 opposite 2-dimensional faces A,B, such that φ(A) and φ(B) are linked L-Schrijver Special case: K6 is not linklessly embeddable
? G path nullspace representation gives embedding in R1 properly normalized G 2-connected outerplanar nullspace representation gives outerplanar embedding in R2 G 3-connected planar nullspace representation gives planar embedding in S2 L-Schrijver G 4-connected linkless embed. nullspace representation gives linkless embedding in R3 ?
planar embedding nullspace representation
G 3-connected nullspace representation gives planar planar embedding in S2 The vectors can be rescaled so that we get a convex polytope.
Colin de Verdière matrix M Steinitz representation P q p u v
Every planar graph can be represented by touching circles Coin representation Koebe (1936) Every planar graph can be represented by touching circles
Polyhedral version Every 3-connected planar graph is the skeleton of a convex polytope such that every edge touches the unit sphere Andre’ev
From polyhedra to circles horizon
From polyhedra to representation of the dual
The Gram representation Kotlov – L - Vempala pos semidefinite Gram representation
Properties of the Gram representation Assume: G has no twin nodes, and exceptional ui is a vertex of P is an edge of P 0 int P If G has no twin nodes, and μ(G)n-4, then is planar.
Vectors to spheres ui Ci ui Ci Cj uj representation of by orth circles planar
of hyperbolic geometry Projective distance Hilbert d distance: a b c Cayley-Klein model of hyperbolic geometry
Projective “distance”: b a b a “distance”: 1 “distance”: p q r s d b c “distance”: a C D
G has euclidean distance 1 representation in R1 G is a path G: connected graph G has euclidean distance 1 representation in R1 G is a path G has projective “distance” >1 representation in R2 G is outerplanar G has projective “distance” 1 representation in R3 G is planar Koebe
Rubber bands and connectivity G: arbitrary graph A,B V, |A|=|B|=k A: affine indep in Rd edges: rubber bands
For almost all choices of edge strengths: B affine indep Linial-L- k disjoint (A,B)-paths Linial-L- Wigderson () cutset
For almost all choices of edge strengths: B affine indep Linial-L- k disjoint (A,B)-paths Linial-L- Wigderson () strengthen
edges strength s.t. B is independent no algebraic relation for a.a. choices of edge strengths, B is independent no algebraic relation between edge strength G is k-connected nodes in the generic rubber band embedding, with A fixed, are in general position
Rubber bands and maximum cuts maximize
Polynomial with 12% error Max Cut: NP-hard Approximations? Easy with 50% error Erdős NP-hard with 6% error Hastad Polynomial with 12% error Goemans-Williamson
spring (repulsive) Energy: How to find minimum energy position? dim=1: Max Cut Min energy 4 Max Cut dim=2: probably hard dim=n: Poly time solvable! semidefinite optimization
Solvable in polynomial time Introduce new variables: These satisfy: linear! convex! The objective function is: Solvable in polynomial time
minimum energy in n dimension random hyperplane Probability of edge ij cut: Expected number of edges cut: