Characterizing Generic Global Rigidity Steven J. Gortler (with A. Healy and D. Thurston)

Post on 22-Dec-2015

219 views 1 download

Transcript of Characterizing Generic Global Rigidity Steven J. Gortler (with A. Healy and D. Thurston)

Characterizing Generic Global Rigidity

Steven J. Gortler

(with A. Healy and D. Thurston)

Global Rigidity

• Given a graph: “G”,

• Given a “framework”: p, in Rd

– For some fixed d

R2

Global Rigidity

• p is GR in Rd : there is no second framework in Rd with the same edge lengths

R2 globally rigid

Global Rigidity

• p is GR in Rd : there is no second framework in Rd with the same edge lengths

R2

Global Rigidity

• p is GR in Rd : there is no second framework in Rd with the same edge lengths

R2

Global Rigidity

• p is GR in Rd : there is no second framework in Rd with the same edge lengths

• Else: p is GF in Rd

R2

Global Rigidity (notes)

• Edge crossing is allowed

• Euclidean: rot, trans, reflect is not considered different

R2

Global Rigidity (notes)

• Locally flexible => globally flexible

• Easier to characterize

R2

Motivation: Distance Geometry

• Input: some pairwise distances

• Output: geometric framework/Eucl– Chemistry, sensor networks

Motivation: Distance Geometry

• Input: some pairwise distances

• Output: geometric framework/Eucl– GR: Well posed-ness of problem– GR: Divide and conquer

Motivation: Distance Geometry

• Molecule problem

• MDS with partial information

• Rank constrained distance matrix completion

Generic

• Given a graph G, and framework p, the GR problem is NP-Hard [Saxe ‘79]

Generic

• Given a graph G, and framework p, the GR problem is NP-Hard [Saxe ‘79]

• The reductions all involve special “coincidences” in the framework

R1

Generic

• Given a graph G, and framework p, the GR problem is NP-Hard [Saxe ‘79]

• The reductions all involve special “coincidences” in the framework

R1

Generic

• Given a graph G, and framework p, the GR problem is NP-Hard [Saxe ‘79]

• Problem seems simpler if we assume no coincidences

R1 globally rigid

Generic

• Given a graph G, and framework p, the GR problem is NP-Hard [Saxe ‘79]

• Problem seems simpler if we assume no coincidences

R1 globally rigid globally flexible

Generic

• Perhaps the problem is easier if we assume that the input framework is “generic”– Think: randomly perturbed

R1 globally rigid globally flexible

Generic

• Perhaps the problem is easier if we assume that the input framework is “generic”

• In 1D, a generic framework is GR iff the graph is 2-connected

R1 globally rigid globally flexible

Generic

• Perhaps the problem is easier if we assume that the input framework is “generic”

• In 1D, a generic framework is GR iff the graph is 2-connected– So GGR in R1 is a property of the graph alone

R1 globally rigid globally flexible

History of GGR

CC: (Connelly condition) Sufficient for all d [’89, H‘95, ‘05]

HC: (Hendrickson condition) Necessary for all d [’88, ’92]

History of GGR

CC: (Connelly condition) Sufficient for all d [’89, H‘95, ‘05]

HC: (Hendrickson condition) Necessary for all d [’88, ’92]

HC = CC (nec & suff) for d=2 [JJ’05]

History of GGR

CC: (Connelly condition) Sufficient for all d [’89, H‘95, ‘05]

HC: (Hendrickson condition) Necessary for all d [’88, ’92]

HC not sufficient for d >= 3 [C ’91]

HC = CC (nec & suff) for d=2 [JJ’05]

K5,5

History of GGR

CC: (Connelly condition) Sufficient for all d [’89, H‘95, ‘05]

HC: (Hendrickson condition) Necessary for all d [’88, ’92]

HC not sufficient for d >= 3 [C ’91]

HC = CC (nec & suff) for d=2 [JJ’05]

K5,5

History of GGR

CC: (Connelly condition) Sufficient for all d [’89, H‘95, ‘05]

HC: (Hendrickson condition) Necessary for all d [’88, ’92]

HC not sufficient for d >= 3 [C ’91]

HC = CC (nec & suff) for d=2 [JJ’05]

CC is necessary for all d [this work]

Main result

• [Connelly] If CC is satisfied by a generic framework in Rd, it is globally rigid in Rd .

• [Thm] If CC is not satisfied by a generic framework in Rd then it is globally flexible in Rd

Main result

• [Connelly] If CC is satisfied by a generic framework in Rd, it is globally rigid in Rd .

• [Thm] If CC is not satisfied by a generic framework in Rd then it is globally flexible in Rd

• Note: CC can be tested with an efficient randomized algorithm.

Main result

• [Connelly] If CC is satisfied by a generic framework in Rd, it is globally rigid in Rd .

• [Thm] If CC is not satisfied by a generic framework in Rd then it is globally flexible in Rd

• Note: CC test gives same answer for all generic frameworks of G in Rd

Main result

• [Connelly] If CC is satisfied by a generic framework, it is globally rigid.

• [Thm] If CC is not satisfied by a generic framework in Rd then it is globally flexible in Rd

• [Cor] A graph is either GR for all generic fmwks in Rd, or is not GR for any generic fmwk in Rd.– So GGR in Rd is a property of the graph alone

On to the condition….

• A real number wuv on each edge euv

Σv wuv [p(v) – p(u)] = 0

Stress Vector satisfied by a framework

R2

f

a

c

b

e

dh

g

p(u)

u

• A real number wuv on each edge euv

Σv wuv [p(v) – p(u)] = 0

Stress Vector satisfied by a framework

R2p(u)e

f

g

c

u

• A real number wuv on each edge euv

Σv wuv [p(v) – p(u)] = 0

Stress Vector satisfied by a framework

u

R2

f

b

h

p(u)

Stress Vector satisfied by a framework

• Equivalent to (symmetrically) writing each vertex as an affine comb of its nbrs

[1/ Σv wuv] Σv wuv p(v) = p(u)

R2

u

fc

e

g

p(u)

Stress Vector satisfied by a framework

Equivalent to equilibrium point of quadratic spring/strut energy (no pins)

E(p)= Σu Σv wuv |p(v) – p(u)|2

R2

Stress vectors: easy facts

• Stress vectors satisfied by a fixed p form a linear space W(p)

p

Stress vectors: easy facts

• Stress vectors satisfied by a fixed p form a linear space W(p)

• Any affine transform ‘T(p)’ will satisfy all stresses in W(p)

p T(p)

Stress vectors: easy facts

• For some graphs, there are even more fmwks than the affine transforms of p,

• that still satisfy all stresses in W(p)

p not an affine tform of p

Connelly’s Condition

• CC := The only fmwks that satisfy all of W(p) are the affine transforms of p

• This will somehow describe GGR!

Stresses and lengths

• What is the relationship between stress vectors– Affine invariant

• ….. and lengths– Euclidean invariant

The mapping “L”

• “L” is the mapping from d-dim fmwks to Re

• Describing each edge’s squared length

Re

R2

L

The mapping “L”

• “L” is the mapping from d-dim fmwks to Re

• Describing each edge’s squared length

Re

R2

L

The set “M”

• “M” is the image of “L” – All possible measurements

Re

R2

L

The set “M”

• “M” is the image of “L” – All possible measurements– A semi-algebraic set

• A smooth manifold a.e.

Re

R2

L

Re

Lengths and stresses: the connection

• At a generic fmwk p,

span L* = the tangent space of M at L(p)

R2

L

Re

Lengths and stresses: the connection

• At a generic fmwk p,

span L* = the tangent space of M at L(p)

• At a generic fmwk p,

W(p) spans the normal space of M at L(p)– [Maxwell]

R2

L

Re

Lengths and stresses: the connection

• So a generic fmwk that satisfies all the same stresses as p must have the same normal space

R2

L

Re

Lengths and stresses: the connection

• So a generic fmwk that satisfies all the same stresses as p must have the same normal space

R2

L

Re

Lengths and stresses: the connection

• So a generic fmwk that satisfies all the same stresses as p must have the same tangent space

R2

L

Re

Lengths and stresses: the connection

• So a generic fmwk that satisfies all the same stresses as p must have the same tangent space– M is a cone: λM = M– Same tangent space, not just parallel

R2

L

Re

Lengths and stresses: the connection

• So a generic fmwk that satisfies all the same stresses as p must have the same tangent space

• This is the key connection

• We will return to this in the proof

R2

L

On to the proof…

Degree mod two thm

Degree mod two thm

2

0

Degree mod two thm

2

0

4

Degree mod two thm

• Typical version:

• Given smooth map from a compact manifold to a connected manifold of same dimension

• Every generic point in the range has the same number of pre-images mod 2– The creases are singular

• This number {0,1} is called the degree

Degree mod two thm

• General version:

• Given smooth map from a compact manifold to a connected manifold of same dimension

• Every generic point in the range has the same number of pre-images mod 2

proper

Our plan

2

0

4

Our plan

2

0

4

Our plan

• Assume (!CC)• Start with the map L• Define a domain

– Equate Euclidean transforms in the domain

• Define range– Connected smooth manifold

• Will need to remove some singularities while maintaining connectivity of range and properness of map

• Show the map has degree 0

• Each point in image has multiple pre-images– Framework is globally flexible

Domain

• Start with stress satisfiers: A(p)– Frameworks that satisfy all of the stresses

that are satisfied by p– Affine tforms of p plus maybe more

Domain

• Start with stress satisfiers: A(p)

• Mod out Euclideans: A(p)/Eucl

Domain

• Start with stress satisfiers: A(p)

• Mod out Euclideans: A(p)/Eucl

• Result is smooth manifold + singularities

Domain

• Start with stress satisfiers: A(p)

• Mod out Euclideans: A(p)/Eucl

• Result is smooth manifold + singularities– Singularities: fmwks stabilized by a n.t. euclidean.

Domain

• Start with stress satisfiers: A(p)

• Mod out Euclideans: A(p)/Eucl

• Result is smooth manifold + singularities– Singularities: deficient affine span.

Lemma 1

• Lemma: If (!CC)– A(p) is “big”

• then the singularities of A(p)/Eucl are of co-dimension >= 2.– Proof: counting

Codimension and cutting

• The singular co-dim will carry over to the range

• Removal a co-dimension 2 set does not disconnect– Needed for degree thm

co-dim 2

Codimension and cutting

• The singular co-dim will carry over to the range

• Removal a co-dimension 2 set does not disconnect– Needed for degree thm

co-dim 2 co-dim 1

The range

• Let B(p) := L(A(p))– Achievable measurements of stress satisfiers– Some subset of M

Re

L

The range

• For the degree to be well defined– Need to include B(p) as a full dimensional

subset of a connected range manifold

Re

L

The range

• For the degree to be well defined– Need to include B(p) as a full dimensional

subset of a connected range manifold

• To show the degree is 0– Sufficient for range to include pts not in B(p)

Re

L

0

The range

• For the degree to be well defined– Need to include B(p) as a full dimensional

subset of a connected range manifold

• To show the degree is 0– Sufficient for range to include pts not in B(p)

Re

L

The range

• For the degree to be well defined– Need to include B(p) as a full dimensional

subset of a connected range manifold

• To show the degree is 0– Sufficient for range to include pts not in B(p)

Re

L

The range

• So we need to understand the shape of B(p)

Re

L

0

Gauss fiber

Re

Digression

• Gauss fiber: points with same (not just parallel) tangent as chosen point

Gauss fiber

• Gauss fiber thm: The Gauss fiber at a generic point of an irreducible algebraic variety is an affine space

Re1d fiber

Digression

Gauss fiber

• Gauss fiber thm: The Gauss fiber at a generic point of an irreducible algebraic variety is a affine space

Re

1d non-affine fiber

Digression

Gauss fiber

• Gauss fiber thm: The Gauss fiber at a generic point of an irreducible algebraic variety is a affine space

Re

1d non-affine fiber, exceptional

Digression

Gauss fiber

• Gauss fiber thm: The Gauss fiber at a generic point of an irreducible algebraic variety is a affine space

Re

0-d fiber

Digression

The range

• Recall “the connection” – Same stresses = same tangent in M

• B(p) is a gauss fiber in M

L

Re

The range

• Recall “the connection” – Same stresses = same tangent in M

• B(p) is a gauss fiber in M

• M is not an irreducible algebraic variety • But it is a full dimensional semi-algebraic subset of one

L

Re

Lemma 2

• Lemma: B(p) is a flat space– Full dimensional subset of an affine space

L

Re

Lemma 2

• Lemma: B(p) is a flat space– Full dimensional subset of an affine space

• So define the range to be this whole affine space

L

Re

Lemma 2

• Lemma: B(p) is a flat space– Full dimensional subset of an affine space

• So define the range to be this whole affine space

L

Re

Lemma 2

• Lemma: B(p) is a flat space– Full dimensional subset of an affine space

• M is contained in first octant, the affine space is not– The degree will be zero

L

Re

0

Lemma 2

• Lemma: B(p) is a flat space– Full dimensional subset of an affine space

• Note: domain and range have same dimension

L

Re

0

Last step

L

• Remove the images of the singularities of A(p)/Eucl from the range and their pre-images– Range remains connected if (!CC)– Domain and range are smooth manifolds

0

Re

Last step

L

• Remove the images of the singularities of A(p)/Eucl from the range and their pre-images– Range remains connected if (!CC)– Domain and range are smooth manifolds

0

Re

Last step

L

• Remove the images of the singularities of A(p)/Eucl from the range and their pre-images

• Can now apply degree thm

0

Re

24

Main Theorem

L

0

24

Re

Main Theorem

L

0

24

Re

Main Theorem

L

0

24

•[Thm] If CC is not satisfied by a generic framework in Rd then it is globally flexible in Rd

Re

Review

• Defined a domain– A(p)/Eucl– Created some singularities

• Defined same dimensional connected smooth range – Affine space containing B(p) (due to flatness)

• Removed singularities– To get smooth domain manifold– Maintained range connectedness (due to high co-dim if !CC)

• Now we there is a well defined degree• Need to know degree is 0 (due to flatness)

Deep Breath…..

Algorithm

• Input: Graph, d

Algorithm

• Input: Graph, d• Pick random framework p in Rd

• Compute stress vector space W(p)– Linear algebra

• Pick a random stress vector w from W(p)• Compute m: dimensionality of fmwks that satisfy w

– Linear algebra

• If m = d(d+1) output “GR in Rd”

• If m > d(d+1) output “GGR in Rd”

Algorithm

• With high probability, p will behave like a generic fmwk

• With high probability, fmwks that satisfy w will satisfy all of W(p)

• Can be done with integer linear algebra• The exceptions satisfy a “low” degree

polynomial• No false positives• GGR in RP

One more breath…

Bonus result

• If a graph is generically globally flexible

• One can continuously flex in one higher dimension back down to second framework

Future work

• More algebraic: – Not just the L function on graphs– More general field– More general metric signature

• More combinatorial:– Deterministic efficient algorithm

Future work

• Given lengths, it is NP-hard to figure out the framework in Rd

• Semi-definite programming will typically find solution in Rv

• But sometimes it happens to give an answer in Rd

– These are frameworks for which higher dimensions don’t help

• Can these cases be nicely characterized?