Approximating the Tutte Polynomial (and the Potts...
Transcript of Approximating the Tutte Polynomial (and the Potts...
![Page 1: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/1.jpg)
Approximating the Tutte Polynomial (and thePotts partition function)
Leslie Ann Goldberg, University of Liverpool
(based on joint work with Mark Jerrum)
Counting, Inference and Optimization on Graphs, PrincetonNovember 2-5, 2011
![Page 2: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/2.jpg)
The Tutte polynomial of a graph G = (V , E)
T (G; x , y) =X
A✓E
(x � 1)(V ,A)�(V ,E)(y � 1)|A|�(|V |�(V ,A))
(V ,A) = number of connected components of the graph (V ,A)
1
![Page 3: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/3.jpg)
T (G; x , y) =X
A✓E
(x � 1)(V ,A)�(V ,E)(y � 1)|A|�(|V |�(V ,A))
If G is connected, T (G; 1, 1) counts spanning trees.
2
![Page 4: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/4.jpg)
T (G; x , y) =X
A✓E
(x � 1)(V ,A)�(V ,E)(y � 1)|A|�(|V |�(V ,A))
If G is connected, T (G; 2, 1) counts forests.
3
![Page 5: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/5.jpg)
Combinatorial interpretation of the Tutte polynomial
1 2 3 4 5 6 7 81 2 3 4 5 6 7 8
1
2
3
4
5
6
7
8
1
2
3
4
5
6
7
8
acyclic orientations
spanning trees
Potts
forests
spanning subsets
reliability polynomial
chromatic polynomial
flow polynomial
Partition function of the q-state Potts model at (x � 1)(y � 1) = q
4
![Page 6: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/6.jpg)
Complexity of evaluating the Tutte polynomial
For fixed rationals x and y , Jaeger, Vertigan and Welsh (1990)studied the complexity of the following problem.
Name. TUTTE(x , y).Input. A graph G = (V ,E).Output. T (G; x , y).
They showed that for all (x , y), TUTTE(x , y) is either #P-hard orcomputable in polynomial time.
T (G; x , y) =X
A✓E
(x � 1)(V ,A)�(V ,E)(y � 1)|A|�|V |+(V ,A)
5
![Page 7: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/7.jpg)
Complexity of evaluating the Tutte polynomial
1 2 3 4 5 6 7 81 2 3 4 5 6 7 8
1
2
3
4
5
6
7
8
1
2
3
4
5
6
7
8
spanning trees
2-colourings
2-flows
(x � 1)(y � 1) = 1.
6
![Page 8: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/8.jpg)
Approximate evaluation
Reminder from Mark’s talk:
Definition. An FPRAS is a randomised algorithm thatproduces a result that is correct to within relative error 1 ± "
with high probability. It must run in time poly(n, "�1), where n isthe input size.
7
![Page 9: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/9.jpg)
Approximate evaluation
1 2 3 4 5 6 7 81 2 3 4 5 6 7 8
1
2
3
4
5
6
7
8
1
2
3
4
5
6
7
8
8
![Page 10: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/10.jpg)
FPRASable points (green)
Points where exact evaluation is possible in polynomialtime
Points on the upper branch of the hyperbola(x � 1)(y � 1) = 2. This is due to Jerrum and Sinclair’sFPRAS (1993) for the partition function of the Ising modelin the ferromagnetic case. (Mark already talked about this)The Ising model is the 2-state Potts model.
9
![Page 11: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/11.jpg)
The random cluster formulation of the Tutte polynomial
The multivariate Tutte polynomial of G is
ZTutte(G; q,�) =X
F✓E
q(V ,F )Y
e2F
�e,
where q and � = {�e}e2E are commuting indeterminates.
If (x � 1)(y � 1) = q and �e = y � 1 for all e then
T (G; x , y) = q�(V ,E)��|V |+(V ,E)ZTutte(G; q,�).
10
![Page 12: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/12.jpg)
The partition function of the q-state Potts model
If q is positive integer then (Fortuin and Kastelyn) ZTutte(G; q,�)is equal to
ZPotts(G; q,�) =X
�:V![q]
Y
e2E
�1 + �e�e(�)
�,
where [q] = {1, . . . , q} is a set of q spins or colours, and �e(�)
is 1 if e is monochromatic in � and 0 otherwise.
The Ising model is the case q = 2. Jerrum and Sinclair’sFPRAS is for �e = � > 0 (the ferromagnetic case).
11
![Page 13: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/13.jpg)
Non-FPRASable pointsAt every grey point, we (2008) showed that there is no FPRASunless RP=NP.
1 2 3 4 5 6 7 81 2 3 4 5 6 7 8
1
2
3
4
5
6
7
8
1
2
3
4
5
6
7
8
12
![Page 14: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/14.jpg)
x < �1 except q = 0, 1
y < �1 except q = 1, 2
In the vicinity of the origin in the triangle y < �1 � 2x
In the vicinity of the origin in the triangle x < �1 � 2y
In the vicinity of the origin and q > 1.5
This is straightforward for some points, for example,T (G;�2, 0) is the number of proper 3-colourings of G (so thedecision problem is NP-hard). On the other hand, T (G; 0,�5)is the number of nowhere-zero 6-flows, and Seymour hasshown that there is a nowhere-zero 6-flow iff G has no cutedge, so the decision problem is in P.
13
![Page 15: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/15.jpg)
Implementing edge weightsusing series and parallel compositions(see Brylawski, JVW, Sokal)
s t s t�1
�2
�⇤ 1 + �⇤ = (1 + �1)(1 + �2)
s t s t�⇤ �1 �2 1 + q�⇤=
⇣1 + q
� 1
⌘⇣1 + q
� 2
⌘
Our key tool (for q 62 {0, 1, 2}): If copies of � can be used toimplement some �⇤ 62 [�2, 0] (y 62 [�1, 1]) and also some�⇤ 2 (�2, 0) then there is no FPRAS for evaluatingZTutte(G; q,�) (where � is the constant function mapping everyedge to weight �).
14
![Page 16: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/16.jpg)
What about the other points?
1 2 3 4 5 6 7 81 2 3 4 5 6 7 8
1
2
3
4
5
6
7
8
1
2
3
4
5
6
7
8
15
![Page 17: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/17.jpg)
On the blue hyperbola segment (q = 2), approximateevaluation is equivalent in difficulty to approximatelycounting perfect matchings in a graph.
Alon, Frieze and Welsh (1995) gave an FPRAS for theregion x � 1, y � 1 when G is “dense” (minimum degree⌦(n)).
Recently (2010), we gave a negative result for the redregion, subject to the hardness of the complexity class#RH⇧1, which Mark told you about.
16
![Page 18: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/18.jpg)
Reminders from Mark’s talk
The problems in #RH⇧1 can be expressed in terms of countingthe number of models of a logical formula from a certainsyntactically restricted class.
Complete for #RH⇧1 wrt Approximation-preserving(AP)-reductions:
ProblemName: #BIS.Instance: A bipartite graph B.Output: The number of independent sets in B.
17
![Page 19: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/19.jpg)
ProblemName: TUTTE(q, �).Instance: A graph G.Output: ZTutte(G; q,�), where � is the constant function with�e = � for all e.
The result in the red region: TUTTE(q, �) is hard for #RH⇧1
with respect to Approximation-Preserving (AP)-reductions whenq > 2 and � > 0.
1 2 3 4 5 6 7 81 2 3 4 5 6 7 8
1
2
3
4
5
6
7
8
1
2
3
4
5
6
7
8
18
![Page 20: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/20.jpg)
The Tutte polynomial of a hypergraph
The (multivariate) Tutte polynomial of H = (V, E) is defined as
ZTutte(H; q,�) =X
F✓Eq(V,F)
Y
f2F�f ,
where (V,F) denotes the number of connected componentsin the sub-hypergraph (V,F).
19
![Page 21: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/21.jpg)
From #BIS to hypergraph Tutte
ProblemName: UNIFORMHYPERTUTTE(q, �).Instance: A uniform hypergraph H = (V, E).Output: ZTutte(H; q,�), where � is the constant function with� f = � for all f .
Using standard techniques:
Lemma. Suppose q > 1. Then
#BIS AP UNIFORMHYPERTUTTE(q, q � 1).
20
![Page 22: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/22.jpg)
From hypergraph Tutte to graph Tutte
Lemma. Suppose that q > 2 and �, �0 > 0. ThenUNIFORMHYPERTUTTE(q, �) AP TUTTE(q, �0).
We go via the intermediate problem of approximatingZTutte(G, q,�) when � = {�e} contains two different edgeweights (which are part of the problem instance).
21
![Page 23: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/23.jpg)
The Random cluster distribution
For a graph G = (V ,E), associate every edge e 2 E with aquantity p(e) 2 [0, 1]. For a set of edges A ✓ E define
eP(G;A, q, p) = q(V ,A)Y
e2A
p(e)Y
e2E\A
(1 � p(e)).
Then the probability of edge-set A in the random cluster modelis given by
P(G;A, q, p) = eP(G;A, q, p)/Zrc(G; q, p).
where Zrc(G; q, p) is the appropriate normalising factor(essentially just an alternative parameterisation of themultivariate Tutte polynomial).
22
![Page 24: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/24.jpg)
First-order phase transitionFor the complete graph KN and q > 2, the random clustermodel has a first-order phase transition with edge probabilityp = �c/N [Bollobás, Grimmett and Janson, 1996].
If p is carefully tuned, two phases coexist: one with allcomponents O(log n); one with a giant component containing aconstant fraction of the vertices.
23
![Page 25: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/25.jpg)
Simulating a hyperedgeGadget for simulating a hyperedge with t = |T | vertices.
"Terminals"
K T
Large complete graph,
edge probability p
Cross edges, probability p'
Salient features:
t ⌧ N, where N = |K |,p ⇡ �c/N,
�c/N ⌧ p0 ⌧ 1.24
![Page 26: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/26.jpg)
Simulating a hyperedge (continued)
Gadget in “hyperedge excluded” state: With high probability alledges crossing from T hit a different connected componentin K .
"Terminals"
K T
Effect: all terminals in T find themselves in differentcomponents.
25
![Page 27: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/27.jpg)
Simulating a hyperedge (continued)
Gadget in “hyperedge included” state: With high probability atleast one edge from each terminal hits the giant componentin K .
"Terminals"
K T
Effect: all terminals in T find themselves in the samecomponents.
26
![Page 28: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/28.jpg)
Getting #BIS hardness for the red region
So far we have reduced #BIS to a multivariate version ofZTutte(G; q,�) in which q is fixed and � = {�e} contains just twovalues (not under our control).
We need to get from there to TUTTE(q, �), where � is a desiredconstant edge weight. This can be using implementations(series-parallel compositions) as before. Putting it all together:
Theorem. Suppose q > 2 and � > 0. Then#BIS AP TUTTE(q, �).
27
![Page 29: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/29.jpg)
1 2 3 4 5 6 7 81 2 3 4 5 6 7 8
1
2
3
4
5
6
7
8
1
2
3
4
5
6
7
8
28
![Page 30: Approximating the Tutte Polynomial (and the Potts ...people.csail.mit.edu/andyd/CIOG_slides/goldberg_talk_ciog2011.pdf(x 1)(y 1)=2. This is due to Jerrum and Sinclair’s FPRAS (1993)](https://reader036.fdocuments.us/reader036/viewer/2022081409/60850d18d8d099705c5a6143/html5/thumbnails/30.jpg)
Does TUTTE(q, �) AP #BIS for q > 2 and � > 0?
Bordewich 2010 showed that if any problem in #P fails to havean FPRAS, then there is an infinite approximation hierarchywithin #P.
29