Universal Rigidity of Generic Frameworks. Steven Gortler joint with Dylan Thurston

Transcription

1 Universal Rigidity of Generic Frameworks Steven Gortler joint with Dylan Thurston

2 Degrees of Graph Rigidity

3 Framework of G Given a graph: G, v vertices, e edges Framework p: drawing in E d (Don t care about edge-edge crossing)

4 Generic We will restrict ourselves to frameworks that are generic in E d Think randomly perturbed in No algebraic coincidences Coincidences ruin the thms. Quite typical scenario for rigidity E d

5 Framework of G Given a graph: G, v vertices, e edges Framework p: drawing in E d (Don t care about edge-edge crossing)

6 Locally flexible Can find continuum of frameworks in d- dims with same edge lengths Graph property (for generic frameworks)

7 Locally rigid No continuum Well understood Rank of appropriate matrix Graph property

8 Locally rigid No continuum But can find other discrete framework in d-dims

9 Globally Rigid No other framework in d-dims Well understood [C 82, GHT 07] Rank of appropriate matrix Graph property

10 Globally Rigid No other framework in d-dims But can be other frameworks in higher dims E 2 E 3

11 Universally Rigid (UR) No other framework in ANY-dims

12 UR: can depend on the generic embedding: 1D Unlike local and global rigidity

13 UR: can depend on the generic embedding: 2D

14 Genericaly UR Graphs Graphs such that ALL d-dim generic frameworks are UR

15 Degrees of rigidity Locally Flexible graph Locally Rigid graph Globally Rigid graph Universally Rigid Framework Universally Rigid Graph

16 Degrees of rigidity 2007 Locally Flexible graph Locally Rigid graph Globally Rigid graph Universally Rigid Framework Universally Rigid Graph

17 Degrees of rigidity Locally Flexible graph Locally Rigid graph Globally Rigid graph 2009 Universally Rigid Framework Universally Rigid Graph

18 Motivation for UR

19 Motivation for UR General embeddings from distances is NP-hard But it is useful in application Molecular structure from NMR data Sensor networks

20 Motivation for UR Try semidefinite programming (SDP) [LLR 95] Unknown Gram matrix Positive semidefinite (PSD) constraint Length squared constraints are linear Will get solution in very high dimension Rank constraints are non-convex UR: No need for explicit rank constraint [So Ye 07] Wish to characterize this class

21 Define terms

22 Equilibrium Stress Vector (ESV) of p A real number w uv on each edge e uv u v ω uv[ρ(v) ρ(u)]= 0 a g b R 2 d c e h f p(u)

23 ESV of p A real number w uv on each edge e uv u v ω uv[ρ(v) ρ(u)]= 0 g R 2 c e f p(u)

24 ESV of p A real number w uv on each edge e uv u v ω uv[ρ(v) ρ(u)]= 0 b R 2 h f p(u)

25 Equilibrium Stress Matrix (ESM) v by v matrix of p Ω uv :=Ω vu :=ω uv Ω uv :=0fornon-edges Ω uu := v ω uv fordiagonals Matricize the ESV

26 Rank

27 Equilibrium Stress Matrix (ESM) of p By construction, kernel must contain: Each of the d-coordinates of p Ωρ j

28 Equilibrium Stress Matrix (ESM) of p By construction, kernel must contain: Each of the d-coordinates of p Ωρ j = v ω uv[ρ j (v) ρ j (u)]

29 Equilibrium Stress Matrix (ESM) of p By construction, kernel must contain: Each of the d-coordinates of p Ωρ j = v ω uv[ρ j (v) ρ j (u)]=0

30 Equilibrium Stress Matrix (ESM) of p By construction, kernel must contain: Each of the d-coordinates of p The all-ones vector Ωρ j = v ω uv[ρ j (v) ρ j (u)]=0 Ω 1= v ω uv[1 1]=0

31 Equilibrium Stress Matrix (ESM) of p By construction, kernel must contain: Each of the d-coordinates of p The all-ones vector Maximal possible rank is v-d-1 Ωρ j = v ω uv[ρ j (v) ρ j (u)]=0 Ω 1= v ω uv[1 1]=0

32 Maximal possible rank is v-d-1 Any affine transform T(p) will satisfy any stress w of p Ω If has maximal rank, then these are the only satisfying drawings p T(p)

33 Springs and Struts

34 Stress Energy Stress vector -> a quadratic spring/strut energy E(ρ):= uv ω uv ρ(v) ρ(u) 2 ESV condition == p is an equilibrium point of this energy Max/Min/Saddle R 2

35 Stress Energy Stress vector -> a quadratic spring/strut energy E(ρ):= uv ω uv ρ(v) ρ(u) 2 = j (ρj ) t Ωρ j ESV condition == p is an equilibrium point of this energy Max/Min/Saddle R 2

36 Positive semidefinite E(ρ):= j (ρj ) t Ωρ j If Ω is positive semidefinite, then p is in fact a min of the spring/strut energy R 2

37 Theorems

38 Sufficiency [Con 82] If p is generic in E. p has a ESM: W of rank v-d-1. W is PSD Ω Then p is UR E d Ω

39 Main result today [GT 09] If p is generic in E. p is UR E d Then p has a ESM: W of rank v-d-1. W is PSD Ω Ω

40 If p is generic in E. p is UR Characterization E d == p has a ESM: W of rank v-d-1. W is PSD Ω Ω

41 Compare to GGR [Con 89+ GHT 07] If p is generic in E. p is UR GR in E d E d == p has a ESM: W of rank v-d-1. W is PSD Ω Ω

42 Proof

43 Convex preliminaries Stratify by dimension Obvious for polytopes Two notions in general

44 Convex extremity k-extreme: not interior to k+1 simplex in the body d b a c

45 Convex extremity k-extreme: not interior to k+1 simplex in the body d b a c 0-extreme

46 Convex extremity k-extreme: not interior to k+1 simplex in the body d b a c 1-extreme

47 Convex extremity k-extreme: not interior to k+1 simplex in the body Face(x): largest convex subset with x in interior (DimFace(x)=smallestkstxisk-extreme) d b a c 1-extreme

48 Convex extremity k-extreme: not interior to k+1 simplex in the body Face(x): largest convex subset with x in interior (DimFace(x)=smallestkstxisk-extreme) d b a c 2-extreme

49 Convex extremity k-extreme: not interior to k+1 simplex in the body Face(x): largest convex subset with x in interior (DimFace(x)=smallestkstxisk-extreme) d b a c 3-extreme

50 Convex exposedness k-exposed: exists halfspace with k dim intersection d b a c 0-exposed

51 Convex exposedness Point d is 0-extreme (with 0-dim face) But not 0-exposed! Every supporting hyperplane includes more stuff d b a c

52 Convex exposedness k-exposed: exists halfspace with k dim intersection d b a c 1-exposed

53 Convex exposedness k-exposed: exists halfspace with k dim intersection d b a c 1-exposed

54 Convex exposedness k-exposed: exists halfspace with k dim intersection d b a c 2-exposed

55 Convex exposedness k-exposed: exists halfspace with k dim intersection d b a c 3-exposed

56 Convex exposedness k-exposed implies k-extreme But not vice-versa d b a c

57 Straszewicz and Asplund [Stra 35]: The 0-exposed are dense in the 0-extreme [Asp 63]: The k-exposed are dense in the k-extreme d b a c

58 Onto our convex body We will be interested in our graph G But we will start with the complete graph on v-verts

59 Measuring the simplex v : thesimplexonvvertices l(ρ) R (v 2) : squarededgelengths ρ l R (v 2)

60 Measuring the simplex v : thesimplexonvvertices l(ρ) R (v 2) : squarededgelengths M( v ): imageofl(anydimspan) ρ M( v ) l R (v 2)

61 Face structure of M(. v ) M( v )isisometrictoconeofpsd(v-1)by(v-1)matrices l(q) Face(l(p))==qisaffineimageofp ρ l

62 Face structure of M(. v ) M( v )isisometrictoconeofpsd(v-1)by(v-1)matrices l(q) Face(l(p))==qisaffineimageofp ρ l l

63 Face structure of M(. v ) M( v )isisometrictoconeofpsd(v-1)by(v-1)matrices l(q) Face(l(p))==qisaffineimageofp ρ l q l

64 Hyperplane distance x ψ x ψ

65 Pull Back x ψ=l(ρ) ψ ρ x=l(ρ) l ψ

66 Matricize x ψ=l(ρ) ψ= j (ρj ) t Ψρ j ρ x=l(ρ) l ψ Ψ

67 Bounding Hyperplane ρ 0 j (ρj ) t Ψρ j ΨisPSD

68 Supporting Hyperplane ΨisPSD. j (ρj ) t Ψρ j =0 Ψρ j =0 ρ j isinthekernel PSD contact == kernel == stress

69 Supporting Hyperplane PSDcontactatρ Contactatthefaceofl(ρ)(atleast)

70 Starting the argument We assume that p is generic framework of G and is UR We want to find a PSD ESM with rank v-d-1

71 What do we need PSD ESM matrix Need a supporting hyperplane of With intersection at x Rank v-d-1 Only affines of p are in kernel Hyperplane intersection exactly Face l(ρ) M( v ) x=l(ρ)

72 What else do we need Recall that we are interested in the graph G, not the simplex

73 What else do we need We need zeros at non edges of G

74 Lets model this We need zeros at non edges of G ω non-edges R (v 2) edges

75 Lets model this We need zeros at non edges of G Projects down to a hyperplane ω π b non-edges d c R (v 2) edges a

76 Faces correspond Lemma: Universalrigidity: π 1 (Face(π(x)))=Face(x) π b non-edges d c R (v 2) edges a

77 So we really just need Any hyperplane downstairs with support Face(π(l(ρ))) b non-edges d c R (v 2) edges a

78 So we really just need Any hyperplane downstairs with support Face(π(l(ρ))) upstairs has support Face(l(ρ)) ω b non-edges d c R (v 2) edges a

79 What Could go Wrong Any hyperplane downstairs with support Face(π(l(ρ))) b non-edges d c R (v 2) edges a

80 Straszewicz and Asplund [Stra 35]: The 0-exposed are dense in the 0-extreme [Asp 63]: The k-exposed are dense in the k-extreme d b a c Use genericity of p (+ a few technicalities)

81 Now we are done Generic point upstairs, maps to generic downstairs Downstairs point has supporting plane with right contact (Asplund) Corresponds to PSD stress of G Contact dim upstairs gives right rank b non-edges d c edges a

82 Computation and SDP

83 Gram UR testing Suppose generic d-dimensional framework Given distances, set up SDP feasibility to look for embeddings Complexity of (exact) SDP feasibility unknown UR => only rank(x)==d+1 Lin span = aff span + 1

84 Gram UR testing Suppose generic d-dimensional framework Given distances, set up SDP feasibility to look for embeddings Complexity of (exact) SDP feasibility unknown UR => only rank(x)=d+1.details.. Not UR => there is x with rank > d+1 UR: highest rank x is d+1

85 Stress UR testing Suppose generic d-dimensional framework Given distances, can set up SDP feasibility to look for PSD stresses UR: highest rank ( ) is v-d-1 (thm) Ω

86 Stress UR testing Suppose generic d-dimensional framework Given distances, can set up SDP feasibility to look for PSD stresses UR: highest rank ( Ω) is v-d-1 (thm) So is it any easier??

87 Dual Computation Suppose generic d-dimensional framework Given distances, can set up SDP feasibility to look for PSD stresses This is simply the dual SDP

88 Dual Computation Suppose generic d-dimensional framework Given distances, can set up SDP feasibility to look for PSD stresses This is simply the dual SDP UR => highest rank x is d+1 UR => highest rank Ω is v-d-1 (thm)

89 Dual Computation Suppose generic d-dimensional framework Given distances, can set up SDP feasibility to look for PSD stresses This is simply the dual SDP UR => highest rank x is d+1 Ω UR => highest rank is v-d-1 (thm) Ω UR => highest [rank(x)+rank( )] is v (thm) (x, Ω ) satisfy strict complementarity if UR (thm)

90 Strict complementarity Min x,β x L+b x S+ n primal Min b,ω Ω L +β Ω S+ n dual

91 Strict complementarity Min x,β x L+b x S+ n primal Min b,ω Ω L +β Ω S+ n dual WC:rank(x )+rank(ω ) n Holds for all solution pairs SC:(x,Ω ) rank(x )+rank(ω )=n Strong form of duality Does not always hold for SDP Needed for fast convergence

92 b What we really have proven We did not rely on any properties of the specific projection. ω π non-edges d c edges a

93 Strict complementarity Min x,β x L+b x S n + Min b,ω Ω L +β Ω S n + [GT09] (L,b,β) x isuniqueandgenericinitsrankoverq(l,β), Ω (x,ω )SC UR Generic d-dim framework

94 Compare to [AHO 97] Min x,β x L+b x S n + Min b,ω Ω L +β Ω S n + [GT09] (L,b,β) x isuniqueandgenericinitsrankoverq(l,β), Ω (x,ω )SC [AHO97] LoverQ, generic(b,β), (x,ω )SC Random lengths= not low rank

95 Remaining Questions Complexity of UR testing Is there a single natural unifying statement for strict complementarity Graphs that are always UR Graphs that are sometimes UR

96 Remaining Questions Complexity of UR testing Is there a single natural unifying statement for strict complementarity Graphs that are always UR Graphs that are sometimes UR [GT09] (L,b,β) x isuniqueandgenericinitsrankoverq(l,β), Ω (x,ω )SC [AHO97] LoverQ, generic(b,β), (x,ω )SC

97 Remaining Questions Complexity of UR testing Is there a single natural unifying statement for strict complementarity Graphs that are always UR Graphs that are sometimes UR

98 Remaining Questions Complexity of UR testing Is there a single natural unifying statement for strict complementarity Graphs that are always UR Graphs that are sometimes UR

99 Thank You