Presentation is loading. Please wait.

Presentation is loading. Please wait.

Bart M. P. Jansen Jesper Nederlof

Similar presentations


Presentation on theme: "Bart M. P. Jansen Jesper Nederlof"β€” Presentation transcript:

1 Computing the Chromatic Number Using Graph Decompositions via Matrix Rank
Bart M. P. Jansen Jesper Nederlof Dagstuhl Seminar New Horizons in Param. Complexity January 21st 2018, Dagstuhl, Germany

2 Graph π‘ž-Coloring Input: An undirected graph 𝐺 Question: Can we assign each vertex a color 𝑓 𝑣 ∈ 1,…, π‘ž , such that 𝑓 𝑒 ≠𝑓(𝑣) for all edges 𝑒,𝑣 ∈𝐸(𝐺) ? NP-complete for π‘žβ‰₯3

3 Coloring algorithms using vertex separators
Parameterize the problem by measures of graph complexity Can π‘ž-Coloring be solved efficiently on graphs that are structurally simple? Parameterized by the treewidth 𝑑𝑀 of the graph, for π‘žβ‰₯3: π‘ž-Coloring can be solved in time 𝑂 βˆ— ( π‘ž 𝑑𝑀 ) π‘ž-Coloring can not be solved in time 𝑂 βˆ— (π‘žβˆ’πœ€ 𝑑𝑀 ) (under SETH) [Lokshtanov, Marx & TALG’18] – 𝑂 βˆ— suppresses π‘π‘œπ‘™π‘¦(𝑛) factors Solid algorithmic understanding for parameterizations based on vertex separators Treewidth lower-bound extends to pathwidth & feedback vertex number. So complexity is understood quite well for parameterizations based on the width/size of vertex separators of the graph.

4 Coloring algorithms using edge separators
Parameterize the problem by measures of graph complexity Can π‘ž-Coloring be solved efficiently on graphs that are structurally simple? What about graphs decomposable by small edge separators? If the input graph has maximum degree 𝑑, is given with a path decomposition of width 𝑝𝑀, and has at most 𝑠 proper π‘ž-colorings, then with high probability: SWAT ’16] Maximum degree 𝑑, pathwidth 𝑝𝑀 implies cutwidth ≀ 𝑑⋅𝑝𝑀 [Thilikos, Serna, J. Algorithms β€˜05] π‘ž-Coloring can be solved in randomized time 𝑂 βˆ— 𝑑/2 +1 𝑝𝑀 ⋅𝑠 Setting of bounded-degree graphs of bounded pathwidth is related to the graph width parameter cutwidth.

5 Decomposing graphs by small edge-cuts
The cutwidth of a linear ordering of 𝑉(𝐺) is: max. # edges intersected by a vertical line between vertices The cutwidth of a graph is min. cutwidth of any ordering of 𝑉(𝐺) For any graph 𝐺: π’•π’“π’†π’†π‘€π‘–π‘‘π‘‘β„Ž 𝐺 β‰€π’‘π’‚π’•π’‰π‘€π‘–π‘‘π‘‘β„Ž 𝐺 β‰€π’„π’–π’•π‘€π‘–π‘‘π‘‘β„Ž(𝐺) Cutwidth gives an upper-bound on the size of edge cuts that decompose the graph linearly, similarly as how pathwidth upper-bounds the size of vertex separators that decompose the graph linearly. So: how fast can q-Coloring be solved parameterized by cutwidth? cutwidth 3

6 Results for cutwidth If the input graph is given with a layout of cutwidth 𝑐𝑑𝑀, then: πœ”β‰ˆ is the matrix multiplication constant; key ingredient is a new rank bound one-sided error caused by Isolation Lemma; key ingredient is communication protocol for any πœ€>0, unless SETH fails Note: running time is independent of π‘ž using cutwidth running time 𝑂 βˆ— π‘ž 𝑑𝑀 is SETH-optimal for treewidth π‘ž-Coloring can be solved deterministically in time 𝑂 βˆ— ( 2 πœ”β‹…π‘π‘‘π‘€ ) π‘ž-Coloring can be solved in time 𝑂 βˆ— ( 2 𝑐𝑑𝑀 ) w. high probability Planar 3-Coloring cannot be solved in time 𝑂 βˆ— 2βˆ’πœ€ 𝑐𝑑𝑀 SETH lower bound holds in the same setting, where a layout is given in the input. Isolation lemma is by Mulmuley, Vazirani and Vazirani [Combinatorica ’87].

7 Results for bounded degree and pathwidth
If the input graph has maximum degree 𝑑, and is given with a path decomposition of width 𝑝𝑀, then: Removes dependence on solution space from BjΓΆrklund’s algorithm [SWAT ’16] Let 𝑑β‰₯5 be odd, let π‘ž 𝑑 ≔ 𝑑/2 +1. Then: for any πœ€>0, unless SETH fails π‘ž-Coloring can be solved in time 𝑂 βˆ— 𝑑/2 +1 𝑝𝑀 w.h.p. π‘ž 𝑑 -Coloring cannot be solved in time 𝑂 βˆ— 𝑑/2 +1βˆ’πœ€ 𝑝𝑀 on graphs of maximum degree 𝑑

8 Deterministic ideas: a rank bound
Uses rank-based approach for dynamic programming [Bodlaender,Cygan,Kratsch & N ’15; Cygan,Kratsch & N ’13; Fomin,Lokshtanov,Panolan & Saurabh ’16] Nr. of partial solutions needed in dynamic-programming table, can be reduced using rank bounds for β€˜compatibility matrices’ Our high-level insight: Compatibility matrix for partial π‘ž-colorings on two sides of a size-π‘˜ vertex separator can have rank β‰ˆ π‘ž π‘˜ size-π‘˜ edge separator has rank ≀ 2 π‘˜ Yields 𝑂 βˆ— ( 2 πœ”β‹…π‘π‘’π‘‘π‘€ ) algorithm using rank-based approach Bodlaender Cygan Kratsch Nederlof; Cygan Kratsch Nederlof; Fomin Lokhstanov Panolan Saurabh.

9 A compatibility matrix for π‘ž-colorings
Consider a bipartite graph 𝐻 on π‘˜ edges with bipartition 𝑋βˆͺπ‘Œ Corresponds to the edges crossing a cut in the linear layout Fix an integer π‘ž and define a matrix 𝑀 𝐻 as follows: For each π‘ž-coloring π’™βˆˆ π‘ž 𝑋 of 𝑋, there is a row in 𝑀 𝐻 For each π‘ž-coloring π’šβˆˆ π‘ž π‘Œ of π‘Œ, there is a column in 𝑀 𝐻 𝑀 𝐻 [𝒙,π’š]= 1 0 𝑀 𝐻 is useful to prune partial solutions, but has too large rank if 𝒙βˆͺπ’š is a proper π‘ž-coloring of 𝐻 otherwise ⬀ ⬀ … ⬀ ⬀ ⬀ 1 ... 𝑋 Intuitively, this is the compatibility matrix for partial solutions to q-coloring on two sides of a small edge separator. π‘Œ

10 A related integer-valued matrix
Consider a bipartite graph 𝐻 on π‘˜ edges with bipartition 𝑋βˆͺπ‘Œ Orient edges from 𝑋 to π‘Œ Fix an integer π‘ž and define a matrix 𝑀 𝐻 β€² as follows: For each π‘ž-coloring π’™βˆˆ π‘ž 𝑋 of 𝑋, there is a row in 𝑀 𝐻 β€² For each π‘ž-coloring π’šβˆˆ π‘ž π‘Œ of π‘Œ, there is a column in 𝑀 𝐻 β€² 𝑀 𝐻 β€² [𝒙,π’š]= (𝑖,𝑗)∈𝐸 𝐻 π‘₯ 𝑖 βˆ’ 𝑦 𝑗 𝑀 𝐻 β€² 𝒙,π’š β‰ 0 ⇔ (𝒙βˆͺπ’š) is a proper π‘ž-coloring of 𝐻 For all π‘ž: π‘Ÿπ‘˜ 𝑀 𝐻 β€² ≀ 2 π‘˜ for bipartite graphs 𝐻 on π‘˜ edges 3 2 1 3 3 3 … 1 2 3 2 1 2 2 -2 3 1 1 ... 𝑋 π‘Œ = 1βˆ’3 β‹… 2βˆ’3 β‹… 3βˆ’2 = βˆ’2 βˆ’1 1 =2 = 1βˆ’3 β‹… 2βˆ’3 β‹… 2βˆ’2 = βˆ’2 βˆ’1 0 =0 = 3βˆ’3 β‹… 1βˆ’3 β‹… 1βˆ’2 = 0 βˆ’2 βˆ’1 =0

11 Exploiting the rank bound algorithmically
Dynamic programming over a given linear layout From left to right, compute proper π‘ž-colorings of first 𝑖 vertices Can safely restrict to a row-basis of the partial colorings At most π‘Ÿπ‘˜( 𝑀′ 𝐻 )≀ 2 π‘˜ rows π‘ž-Coloring can be solved deterministically in time 𝑂 βˆ— ( 2 πœ”β‹…π‘π‘‘π‘€ ) π‘ž-Coloring can be solved deterministically in time 𝑂 βˆ— ( 2 πœ”β‹…π‘π‘‘π‘€ ) e f … b d β‰ 0 ... Running time dominated by the maximum number of colorings stored simultaneously at any point in time.

12 Intermezzo: The lower bounds
I’ve heard Daniel Marx say that every talk should have a picture of a reduction no-one understands, and since he’s one of the organizers I feel obliged to comply. Reduction from CNF-SAT with n variables to Planar 3-Coloring instance of cutwidth n + O(1).

13 Randomized ideas: the graph polynomial
Let 𝐺 be an undirected graph on vertex set 𝑛 The 𝑛-variate graph polynomial 𝑓 𝐺 is defined as follows: 𝑓 𝐺 π‘₯ 1 ,…, π‘₯ 𝑛 ≔ {𝑖,𝑗}∈𝐸(𝐺) 𝑖<𝑗 π‘₯ 𝑖 βˆ’ π‘₯ 𝑗 Extensively studied, i.e. in context of Alon-Tarsi Theorem [Combinatorica β€˜92] For any 𝒙= π‘₯ 1 ,…, π‘₯ 𝑛 ∈ π‘ž 𝑛 , we have: We analyze π‘ž-colorability of 𝐺 by summing the graph polynomial If π’™βˆˆ π‘ž 𝑛 𝑓 𝐺 (𝒙) β‰ 0, then 𝐺 is π‘ž-colorable; but converse may fail 𝒙 is a proper π‘ž-coloring of 𝐺⇔ 𝑓 𝐺 𝒙 β‰ 0 If π‘₯ is proper, then for each edge the endpoints have different colors, so 𝑝 𝐺 (𝒙) is a product of nonzero terms, so nonzero. If π‘₯ is not proper, then there is an edge whose endpoints have the same color, which yields a 0-term in the product.

14 Isolating a coloring Pick random weight ∈[2π‘›π‘ž] for each (π‘£βˆˆπ‘‰(𝐺),π‘βˆˆ π‘ž ) pair Algorithmic implication: To get a Monte Carlo algorithm for π‘ž-Coloring, pick random weight function and test whether βˆƒπ‘§βˆˆβ„• such that: 𝑃 𝐺 𝑧 ≔ π’™βˆˆ π‘ž 𝑛 weight 𝒙 =𝑧 𝑓 𝐺 (𝒙) β‰ 0 If there is a unique weight-𝑧 proper π‘ž-coloring, then 𝑃 𝐺 𝑧 β‰ 0 If 𝑃 𝐺 𝑧 β‰ 0, then there is a proper π‘ž-coloring of weight 𝑧 Isolation Lemma [Mulmuley, Vazirani, and Vazirani] If there is a proper π‘ž-coloring, then with probability β‰₯ 1 2 there is a unique proper π‘ž-coloring of minimum total weight

15 Rewriting the sum 𝑃 𝐺 𝑧 ≔ π’™βˆˆ π‘ž 𝑛 weight 𝒙 =𝑧 𝑓 𝐺 𝒙 = π’™βˆˆ π‘ž 𝑛 weight 𝒙 =𝑧 𝑖,𝑗 ∈𝐸 𝐺 𝑖<𝑗 π‘₯ 𝑖 βˆ’ π‘₯ 𝑗 = π’™βˆˆ π‘ž 𝑛 weight 𝒙 =𝑧 π‘ŠβŠ†πΈ(𝐺) 𝑖,𝑗 βˆˆπ‘Š 𝑖<𝑗 π‘₯ 𝑖 𝑖,𝑗 ∈𝐸 𝐺 βˆ–π‘Š 𝑖<𝑗 βˆ’ π‘₯ 𝑗 For fixed π‘Š, degree of π‘₯ π‘˜ in (∏ π‘₯ 𝑖 )( βˆ’ π‘₯ 𝑗 ) equals sum of: number of edges βˆˆπ‘Š to neighbors 𝑗>π‘˜ number of edges βˆ‰π‘Š to neighbors 𝑖<π‘˜ Let’s look at the graph polynomial, summed over all colorings of weight z. For each way of choosing the left or right endpoint of an edge, you get a term in the expansion of the product. Degree consideration hints at connection between degree of (partial evaluations of) this polynomial, and cutwidth of the graph.

16 Evaluate the sum using a linear layout
𝑃 𝐺 𝑧 = π’™βˆˆ π‘ž 𝑛 weight 𝒙 =𝑧 π‘ŠβŠ†πΈ(𝐺) 𝑖,𝑗 βˆˆπ‘Š 𝑖<𝑗 π‘₯ 𝑖 𝑖,𝑗 ∈𝐸 𝐺 βˆ–π‘Š 𝑖<𝑗 βˆ’ π‘₯ 𝑗 To solve π‘ž-Coloring on a graph with a bounded-cutwidth layout: Left-to-right dynamic programming over the layout, computing partial sums over colorings of the first 𝑖 vertices Leave some terms un-evaluated to allow partial sum for prefix of length 𝑖+1 to be computed from partial sums for prefixes of length 𝑖 Small cutwidth layout β‡’ degree of un-evaluated part is small

17 Main idea for the dynamic program
Communication protocol for evaluating polynomial sums Table entries stored in DP correspond to message that allows other player to find the answer Alice has univariate polynomial 𝑃 𝐴 (π‘₯) of degree 𝑑 𝐴 Bob has univariate polynomial 𝑃 𝐡 π‘₯ of degree 𝑑 𝐡 Both parties know 𝑑 𝐴 , 𝑑 𝐡 , and π‘ž Goal: Using little communication from Alice to Bob, enable Bob to output π‘₯∈ π‘ž 𝑃 𝐴 π‘₯ β‹… 𝑃 𝐡 π‘₯ Straight-forward communication strategy: Consider coefficients of 𝑃 𝐴 π‘₯ = 𝑐 0 π‘₯ 0 + 𝑐 1 π‘₯ 1 +…+ 𝑐 𝑑 𝐴 π‘₯ 𝑑 𝐴 Alice sends her 𝑑 𝐴 +1 coefficients ( 𝑐 0 , 𝑐 1 ,…, 𝑐 𝑑 𝐴 ) to Bob Summing evaluations of a polynomial that is split into two bounded-degree factors (one corresponding to edges already encountered in the linear layout, other corresponding to edges to-be encountered in the future).

18 Alternative communication strategy
Goal: Using little communication from Alice to Bob, enable Bob to output π‘₯∈ π‘ž 𝑃 𝐴 π‘₯ β‹… 𝑃 𝐡 π‘₯ Straight-forward communication strategy: Alice sends the 𝑑 𝐴 +1 coefficients ( 𝑐 0 , 𝑐 1 ,…, 𝑐 𝑑 𝐴 ) of 𝑃 𝐴 to Bob Alternative communication strategy: Consider coefficients of 𝑃 𝐡 π‘₯ = 𝑐 0 β€² π‘₯ 0 + 𝑐 1 β€² π‘₯ 1 +…+ 𝑐 𝑑 𝐡 β€² π‘₯ 𝑑 𝐡 Desired output can be rewritten as: π‘₯∈ π‘ž 𝑃 𝐴 π‘₯ 𝑐 0 β€² π‘₯ 0 + 𝑐 1 β€² π‘₯ 1 +…+ 𝑐 𝑑 𝐡 β€² π‘₯ 𝑑 𝐡 = 𝑐 0 β€² π‘₯∈ π‘ž 𝑃 𝐴 π‘₯ π‘₯ 0 + 𝑐 1 β€² π‘₯∈ π‘ž 𝑃 𝐴 π‘₯ π‘₯ 1 +β‹―+ 𝑐 𝑑 𝐡 β€² π‘₯∈ π‘ž 𝑃 𝐴 π‘₯ π‘₯ 𝑑 𝐡 First split the single sum into 𝑑 𝐡 +1 sums, one for every power; within one such sum, we can move the coefficient in front. These 𝑑 𝐡 +1 partial evaluations allow Bob to compute the output!

19 Alternative communication strategy
Goal: Using little communication from Alice to Bob, enable Bob to output π‘₯∈ π‘ž 𝑃 𝐴 π‘₯ β‹… 𝑃 𝐡 π‘₯ Straight-forward communication strategy: Alice sends the 𝑑 𝐴 +1 coefficients ( 𝑐 0 , 𝑐 1 ,…, 𝑐 𝑑 𝐴 ) of 𝑃 𝐴 to Bob Alternative communication strategy: Alice sends 𝑑 𝐡 +1 evaluations π‘₯∈ π‘ž 𝑃 𝐴 π‘₯ π‘₯ 𝑖 for 0≀𝑖≀ 𝑑 𝐡 Best protocol depends on which polynomial has larger degree Extends to multivariate polynomials for Alice and Bob Mix the two strategies, choosing most efficient one for each variable First split the single sum into 𝑑 𝐡 +1 sums, one for every power; within one such sum, we can move the coefficient in front.

20 From communication protocol to π‘ž-Coloring
From left to right in the linear layout, compute partial evaluations of 𝑃 𝐺 (𝑧) about the first 𝑖 vertices 𝑉 𝑖 Un-evaluated variables in the polynomial for 𝑉 𝑖 correspond to: vertices incident to an edge in the cut after 𝑖 Independently pick communication strategy for each vertex 𝑣: does the majority of the neighbors occur at positions ≀𝑖, or at positions >𝑖? π‘ž-Coloring can be solved in time 𝑂 βˆ— ( 2 𝑐𝑑𝑀 ) w. high probability π‘ž-Coloring can be solved in time 𝑂 βˆ— 𝑑/2 +1 𝑝𝑀 w.h.p.

21 THANK YOU! Conclusion PhD & Postdoc
SETH-optimal algorithms for π‘ž-Coloring, using several parameterizations concerning edge separators Uniform approach works for both cutwidth / degree+pathwidth Running times for cutwidth parameterization independent of π‘ž Running time 𝑂 βˆ— ( 2 𝑐𝑑𝑀 ) is already optimal for Planar 3-Coloring Open problems: Derandomize the algorithms Experiment with the cut-based reduction of partial solutions PhD & Postdoc THANK YOU!


Download ppt "Bart M. P. Jansen Jesper Nederlof"

Similar presentations


Ads by Google