Quick viewing(Text Mode)

Lower Bounds for Subgraph Isomorphism

Lower Bounds for Subgraph Isomorphism

Lower Bounds for Subgraph Isomorphism

Benjamin Rossman University of Toronto Outline

• Subgraph isomorphism problems • Two grand challenges in Circuit Complexity: P vs NP via the circuit size of k-CLIQUE NC1 vs P via the formula size of k-CYCLE • The average-case setting • Progress in restricted circuit classes Subgraph Isomorphism Problems k-CLIQUE Given a graph X, does it contain a k-clique (complete subgraph of size k)? k-CLIQUE Given a graph X, does it contain a k-clique (complete subgraph of size k)? k-CLIQUE Given a graph X, does it contain a k-clique (complete subgraph of size k)?

k is an arbitrary fixed constant n is the size of the input graph X k-CLIQUE Given a graph X, does it contain a or slow-growing function k-clique

(complete subgraph of size k(n) = O(log n)k)?

k is an arbitrary fixed constant n is the size of the input graph X k-CLIQUE Given a graph X, does it contain a k-clique (complete subgraph of size k)?

Time complexity of k-CLIQUE • exhaustive search O(nk)

• fastest known algorithm n0.79 k + O(1) k-CLIQUE Given a graph X, does it contain a k-clique (complete subgraph of size k)?

Time complexity of k-CLIQUE • exhaustive search O(nk)

• fastest known algorithm n0.79 k + O(1)

n(ω/3)k + O(1) where 2 ≤ ω < 2.37 is the exponent of matrix multiplication [Nesetril-Poljak ’85] k-CLIQUE Given a graph X, does it contain a k-clique (complete subgraph of size k)?

Time complexity of k-CLIQUE • exhaustive search O(nk)

• fastest known algorithm n0.79 k + O(1) • conjectured lower bound nΩ(k) (implies P ≠ NP) k-CYCLE (on layered graphs)

v0 v1 ⋯ vk–1 v0 k-CYCLE (on layered graphs)

v0 v1 ⋯ vk–1 v0

1 1 2 2 3 3 ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ n n k-CYCLE (on layered graphs)

v0 v1 ⋯ vk–1 v0

1 1 2 2 3 3 ⋮ ⋮ X ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ n n k-CYCLE (on layered graphs)

v0 v1 ⋯ vk–1 v0

1 1 2 2 3 3 ⋮ ⋮ X ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ n n k-CYCLE (on layered graphs)

Space complexity of k-CYCLE • best upper bound O(log k · log n) • conjectured lower bound Ω(log k · log n) (implies L ≠ NL)

G-SUBGRAPH ISOMORPHISM

G G-SUBGRAPH ISOMORPHISM Given a graph X, does it contain a subgraph isomorphic to G?

G X COLORED G-SUBGRAPH ISOMORPHISM (SUB(G)) Given a graph X and a vertex-coloring V(X) → V(G), does X contain a properly colored G-subgraph?

G X COLORED G-SUBGRAPH ISOMORPHISM (SUB(G)) Given a graph X and a vertex-coloring V(X) → V(G), does X contain a properly colored G-subgraph?

• Includes k-CLIQUE and k-CYCLE as special cases COLORED G-SUBGRAPH ISOMORPHISM (SUB(G)) Given a graph X and a vertex-coloring V(X) → V(G), does X contain a properly colored G-subgraph?

• Includes k-CLIQUE and k-CYCLE as special cases

• Complexity is minor-monotone: if F is a minor of G, then SUB(F) reduces to SUB(G) COLORED G-SUBGRAPH ISOMORPHISM (SUB(G)) Given a graph X and a vertex-coloring V(X) → V(G), does X contain a properly colored G-subgraph?

• Includes k-CLIQUE and k-CYCLE as special cases

• Complexity is minor-monotone: if F is a minor of G, then SUB(F) reduces to SUB(G)

• Conjectured to be hard-on-average for natural input distributions: (G-colored) Erdős-Rényi random graphs at an appropriate threshold

COLORED G-SUBGRAPH ISOMORPHISM (SUB(G)) Given a graph X and a vertex-coloring V(X) → V(G), does X contain a properly colored G-subgraph?

Understanding the complexity of SUB(G) with respect to different computational resources (time, space, …) leads to separations of important complexity classes (P ≠ NP, L ≠ NL, …) COLORED G-SUBGRAPH ISOMORPHISM (SUB(G)) Given a graph X and a vertex-coloring V(X) → V(G), does X contain a properly colored G-subgraph?

Understanding the complexity of SUB(G) with respect to different computational resources (time, space, …) leads to separations of important complexity classes (P ≠ NP, L ≠ NL, …)

We focus on complexity measures defined by Boolean circuits Circuit Complexity Circuit Complexity

• A concrete, combinatorial approach to the major questions of Complexity Theory including P vs. NP Circuit Complexity

• Studies the complexity of specific problems (e.g. k-CLIQUE, MATRIX MULTIPLICATION, etc.) in combinatorial models of computation, most importantly Boolean circuits Circuit Complexity

• Studies the complexity of specific problems (e.g. k-CLIQUE, MATRIX MULTIPLICATION, etc.) in combinatorial models of computation, most importantly Boolean circuits

• Goal is to prove unconditional lower bounds, which do not rely on any unproven assumptions Circuit Complexity

• Studies the complexity of specific problems (e.g. k-CLIQUE, MATRIX MULTIPLICATION, etc.) in combinatorial models of computation, most importantly Boolean circuits

• Goal is to prove unconditional lower bounds which a problem (i.e. decision problem) is do not reply on any unproven assumptions represented by a sequence of n boolean functions fn : {0,1} → {0,1} Circuit Complexity

• Studies the complexity of specific problems (e.g. k-CLIQUE, MATRIX MULTIPLICATION, etc.) in combinatorial models of computation, most importantly Boolean circuits

• Goal is to prove unconditional lower bounds which do not reply on any unproven assumptions represented by a seq. of functions ( )n {0,1} 2 (≈ {n-vertex graphs}) → {0,1}

Boolean Circuits

¬ ∨

∨ ∧ ¬

¬ ∨ ∧

∧ ∧ ¬ ∨

x1 x2 x3 x4 x5 Boolean Circuits

• Every n-variable computes an n- variable {0,1}n → {0,1} Boolean Circuits

• Every n-variable Boolean circuit computes an n- variable Boolean function {0,1}n → {0,1}

• We say that a problem is “solved” by a sequence of Boolean circuits C1, C2, …, Cn, … if Cn computes the appropriate function {0,1}n → {0,1}

Boolean Circuits

• Every n-variable Boolean circuit computes an n- variable Boolean function {0,1}n → {0,1}

• We say that a problem is “solved” by a sequence of Boolean circuits C1, C2, …, Cn, … if Cn computes the appropriate function {0,1}n → {0,1}

in contrast to uniform models of computation (e.g. Turing machines) where a single algorithm solves the problem on all instances Circuit Size

• The circuit size of a function f : {0,1}n → {0,1} is the minimum number of gates in a circuit computing f Circuit Size

• The circuit size of a function f : {0,1}n → {0,1} is the minimum number of gates in a circuit computing f

• Theorem [Shannon 1949] Almost all Boolean functions have circuit size Ω(2n/n) Circuit Size

• The circuit size of a function f : {0,1}n → {0,1} is the minimum number of gates in a circuit computing f

• Theorem [Shannon 1949] Almost all Boolean functions have circuit size Ω(2n/n)

• The goal in Circuit Complexity is proving lower bounds for explicit Boolean functions (e.g. k-CLIQUE) Circuit Size

• Theorem [Schnorr 1976, Fischer-Pippenger 1979] Turing mach. time T(n) ⟹ circuit size O(T(n)*log T(n))

• Corollary A super-polynomial lower bound on the circuit size of any function in NP implies P ≠ NP Circuit Size

• Theorem [Schnorr 1976, Fischer-Pippenger 1979] Turing mach. time T(n) ⟹ circuit size O(T(n)*log T(n))

• Corollary A super-polynomial lower bound on the circuit size of any function in NP implies P ≠ NP

Circuit Complexity is widely believed to be the most viable approach to P ≠ NP Circuit Size

• Grand challenge (P ≠ NP) Prove a super-polynomial lower bound on the circuit size of any problem in NP

Circuit Size

• Grand challenge (P ≠ NP) Prove a super-polynomial lower bound on the circuit size of any problem in NP

Concretely, show that k-CLIQUE requires circuit size nΩ(k) Circuit Size

• Grand challenge (P ≠ NP) Prove a super-polynomial lower bound on the circuit size of any problem in NP

• Best known lower bound

Circuit Size

• Grand challenge (P ≠ NP) Prove a super-polynomial lower bound on the circuit size of any problem in NP

• Best known lower bound

3n – O(1) 1976 [Schnorr] 4n – O(1) 1991 [Zwick] 4.5n – o(n) 2001 [Lachish-Raz] 5n – o(n) 2002 - today [Iwama-Morizumi] Boolean Formulas

Formulas are circuits with ∧ the structure of a tree ¬ ∨

∨ ∧ ∧

¬ ∨ x4 ¬ ¬ ∧

∧ ∧ ∨ x2 ¬ x3 x4

x3 x1 x2 x4 x5 x1 x5 Formulas lack “memory”: the result of each sub-computation is only used once Boolean Formulas

Formulas are circuits with ∧ the structure of a tree ¬ ∨

∨ ∧ ∧

¬ ∨ x4 ¬ ¬ ∧

∧ ∧ ∨ x2 ¬ x3 x4

x3 x1 x2 x4 x5 x1 x5 Formulas lack “memory”: the result of each sub-computation is only used once Boolean Formulas

Formulas are circuits with ∧ the structure of a tree ¬ ∨ Open: Are circuits more

powerful than formulas? ∨ ∧ ∧

¬ ∨ x4 ¬ ¬ ∧

∧ ∧ ∨ x2 ¬ x3 x4

x3 x1 x2 x4 x5 x1 x5 Formulas vs. Circuits

• Grand challenge #2 (NC1 ≠ P) Prove that polynomial-size circuits are strictly more powerful than polynomial-size formulas Formulas vs. Circuits

• Grand challenge #2 (NC1 ≠ P) Prove that polynomial-size circuits are strictly more powerful than polynomial-size formulas

NC1 = the class of problems solvable by polynomial-size formulas Formulas vs. Circuits

• Grand challenge #2 (NC1 ≠ P) Prove that polynomial-size circuits are strictly more powerful than polynomial-size formulas

Concretely, show that k-CYCLE requires formula size nΩ(log k) Formulas vs. Circuits

• Grand challenge #2 (NC1 ≠ P) Prove that polynomial-size circuits are strictly more powerful than polynomial-size formulas

• Best known formula size lower bound n1.5 – o(1) 1961 [Subbotovskaya] n2 1971 [Khrapchenko] n2.5 – o(1) 1991 [Andreev] n3 – o(1) 1998 - today [Hastad] Restricted Circuit Classes Restricted Circuit Classes

• Most research in Circuit Complexity studies restricted classes of circuits and formulas in the hope of developing stronger techniques Restricted Circuit Classes

• Most research in Circuit Complexity studies restricted classes of circuits and formulas in the hope of developing stronger techniques

• The two most natural and well-studied restrictions: AC0 circuits constant-depth, unbounded fan-in Boolean circuits monotone circuits Boolean circuits without negation (no NOT gates) AC0 Circuits

∨ ∨ ∨ ∨ ∨

∧ ∧ ∧ ∧ ∧ ∧

∨ ∨ ∨ ∨ ∨ ∨

x1 ¬x1 x2 ¬x2 x3 ¬x3 x4 ¬x4 x5 ¬x5 AC0 Circuits

∧ depth is bounded by a constant d

∨(or slow-growing function d(n) << log n) ∨ ∨ ∨ ∨

∧ ∧ ∧ ∧ ∧ ∧

∨ ∨ ∨ ∨ ∨ ∨

x1 ¬x1 x2 ¬x2 x3 ¬x3 x4 ¬x4 x5 ¬x5 AC0 Formulas

∨ ∨ ∨

∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧

… x5 ¬x8 AC0 Lower Bounds

• Exponential lower bounds known since the 1980’s: 0 Θ(n1/(d– )1) the depth-d AC circuit size PARITYn is 2 [Ajtai, Furst-Saxe-Sipser, Yao, Hastad]

AC0 Lower Bounds

• Exponential lower bounds known since the 1980’s: 0 Θ(n1/(d– )1) the depth-d AC circuit size PARITYn is 2 [Ajtai, Furst-Saxe-Sipser, Yao, Hastad]

This “size-depth tradeoff” is a limitation of AC0 lower bounds (which become trivial at depth d = log n) AC0 Lower Bounds

• Exponential lower bounds known since the 1980’s: 0 Θ(n1/(d– )1) the depth-d AC circuit size PARITYn is 2 [Ajtai, Furst-Saxe-Sipser, Yao, Hastad]

The true complexity of k-CLIQUE and k-CYCLE should not exhibit such a “size-depth tradeoff” (once d >> k) AC0 Lower Bounds

• Exponential lower bounds known since the 1980’s: 0 Θ(n1/(d– )1) the depth-d AC circuit size PARITYn is 2 [Ajtai, Furst-Saxe-Sipser, Yao, Hastad]

• Certain barriers (“Natural Proofs” [Razborov-Rudich ’97]) show that existing lower bounds methods cannot extend to unbounded-depth circuits and formulas AC0 Lower Bounds

• Exponential lower bounds known since the 1980’s: 0 Θ(n1/(d– )1) the depth-d AC circuit size PARITYn is 2 [Ajtai, Furst-Saxe-Sipser, Yao, Hastad]

• Certain barriers (“Natural Proofs” [Razborov-Rudich ’97]) show that existing lower bounds methods cannot extend to unbounded-depth circuits and formulas

We want lower bounds for AC0 circuits and formulas with potential to extend beyond depth log n The Monotone Setting

• k-CLIQUE (and other subgraph isomorphism problems) are monotone properties: if a graph X contains a k-clique, then so does any graph X’ ⊇ X. The Monotone Setting

• k-CLIQUE (and other subgraph isomorphism problems) are monotone properties: if a graph X contains a k-clique, then so does any graph X’ ⊇ X.

• Natural to study their complexity with respect to monotone circuits and formulas (= Boolean circuits and formulas without NOT gates) Monotone Lower Bounds

• Many strong lower bounds in monotone models!

Theorem [Razborov ‘86] k-CLIQUE has monotone circuit size nΩ(k)

Theorem [Karcher-Wigderson ‘88] k-CYCLE has monotone formula size nΩ(log k) Monotone Lower Bounds

• Many strong lower bounds in monotone models! mon-P ≠ mon-NP Theorem [Razborov ‘86] k-CLIQUE has monotone circuit size nΩ(k)

Theorem [Karcher-Wigderson ‘88] k-CYCLE has monotone formula size nΩ(log k)

mon-NC1 ≠ mon-P Monotone Lower Bounds

• Many strong lower bounds in monotone models!

Theorem [Razborov ‘86] k-CLIQUE has monotone circuit size nΩ(k)

Theorem [Karcher-Wigderson ‘88] k-CYCLE has monotone formula size nΩ(log k)

• However, it’s unclear what these lower bounds tell us about non-monotone circuits and formulas (∃ exponential gaps between mon vs. non-mon complexity of explicit monotone functions [Tardos ‘88]) k-CLIQUE

complete graph

n-vertex graphs (ordered by inclusion)

empty graph k-CLIQUE

YES instances (graphs with a k-clique)

NO instances no k-clique k-CLIQUE

YES instances (graphs with a k-clique) threshold edge density (50% of graphs have a k-clique)

NO instances no k-clique Theorem [Razborov 86] k-CLIQUE has monotone circuit size nΩ(k)

YES instances (graphs with a k-clique)

NO instances no k-clique Theorem Razborov’s[Razborov 86] Distribution k-CLIQUE has monotone circuit size nΩ(k)

relevant NO instances

relevant YES instances Theorem Razborov’s[Razborov 86] Distribution k-CLIQUE has monotone circuit size nΩ(k)

complete k−1 partite graphs (maxterms)

relevant NO instances

relevant YES instances

isolated k-cliques (minterms) Theorem Razborov’s[Karchmer-Wigderson Distribution 88] STCONN has monotone circuit size nΩ(log n)

st-cuts (maxterms)

relevant NO instances

relevant YES instances

st-paths (minterms) The relevant YES and NO instances are separable by density (# edges)

relevant NO instances

relevant YES instances …and therefore separable by small non-monotone circuits (size O(n2))

anti-monotone threshold function k-CLIQUE on G(n,p)

Erdős-Rényi random graph at the k-CLIQUE threshold k-CLIQUE on G(n,p)

Relevant YES and NO instances are not separable by density k-CLIQUE on G(n,p)

Relevant YES and NO instances are not separable by density

Open problem: Show that average-case k-CLIQUE on G(n,p) requires monotone circuit size nΩ(k) k-CLIQUE on G(n,m)

exactly m edges Theorem [Berkowitz 82]k-CLIQUE On any on sliceG( of the Hamming cuben,m) , monotone complexity = non-monotone complexity Theorem [Berkowitz 82]k-CLIQUE On any on sliceG( of the Hamming cuben,m) , monotone complexity = non-monotone complexity

Corollary If k-CLIQUE on G(n,m) has monotone circuit size nΩ(k), then P ≠ NP Summary, so far

Subgraph isomorphism problems k-CLIQUE k-CYCLE SUB(G) Two grand challenges in Circuit Complexity P vs. NP via the circuit size of k-CLIQUE NC1 vs. P via the formula size of k-CYCLE Two restricted settings & limitations of known results AC0 limited by “size-depth tradeoffs” monotone lack of average-case lower bounds (under product distributions, e.g. G(n,p)) Summary, so far

Subgraph isomorphism problems k-CLIQUE k-CYCLE SUB(G) Two grand challenges in Circuit Complexity P vs. NP via the circuit size of k-CLIQUE NC1 vs. P via the formula size of k-CYCLE Two restricted settings & limitations of known results AC0 limited by “size-depth tradeoffs” monotone lack of average-case lower bounds (under product distributions, e.g. G(n,p)) Average-Case k-CLIQUE k-CLIQUE on G(n,p)

• We consider G(n,p) where p = Θ(n−2/(k–1)) is the critical threshold for the existence of k-cliques

1

Pr[ G(n,p) contains a k-clique ] 1/2

0 pthreshold 1 edge probability p k-CLIQUE on G(n,p)

• We consider G(n,p) where p = Θ(n−2/(k–1)) is the critical threshold for the existence of k-cliques

• Best known algorithm for k-CLIQUE on G(n,p): (iterated, randomized) GREEDY algorithm GREEDY algorithm on G(n,p) Example k = 6 p = ϴ(n−2/5)

GREEDY algorithm on G(n,p) Example k = 6 p = ϴ(n−2/5)

GREEDY algorithm on G(n,p) Example k = 6 p = ϴ(n−2/5)

GREEDY algorithm on G(n,p) Example k = 6 p = ϴ(n−2/5)

GREEDY algorithm on G(n,p) Example k = 6 p = ϴ(n−2/5)

GREEDY algorithm on G(n,p) Example k = 6 p = ϴ(n−2/5)

terminates with 3-clique (no common neighbor)

GREEDY algorithm on G(n,p) Example k = 6 p = ϴ(n−2/5)

GREEDY algorithm on G(n,p) Example k = 6 p = ϴ(n−2/5)

GREEDY algorithm on G(n,p) Example k = 6 p = ϴ(n−2/5)

GREEDY algorithm on G(n,p) Example k = 6 p = ϴ(n−2/5)

again terminates with 3-clique (in fact, this happens almost surely)

GREEDY algorithm on G(n,p) Example k = 6 p = ϴ(n−2/5)

GREEDY algorithm on G(n,p) Example k = 6 p = ϴ(n−2/5)

GREEDY algorithm on G(n,p) Example k = 6 p = ϴ(n−2/5)

GREEDY algorithm on G(n,p) Example k = 6 p = ϴ(n−2/5)

GREEDY algorithm on G(n,p) Example k = 6 p = ϴ(n−2/5)

GREEDY algorithm on G(n,p) Example k = 6 p = ϴ(n−2/5)

almost surely: • after n1/5 iterations, we find a 4-clique • after n4/5 iterations, we find a 5-clique • after n9/5 iterations, we find a 6-clique (if one exists)..

GREEDY algorithm on G(n,p) Average-Case k-CLIQUE

• nε k2 iterations of GREEDY almost surely find a clique of size (½ + ε)k for any 0 ≤ ε < ½

• nk/4 + O(1) iterations of GREEDY solve k-CLIQUE almost surely on G(n,p) Average-Case k-CLIQUE

• nε k2 iterations of GREEDY almost surely find a clique of size (½ + ε)k for any 0 ≤ ε < ½

• nk/4 + O(1) iterations of GREEDY solve k-CLIQUE almost surely on G(n,p)

has implementation by monotone AC0 circuits [Amano ‘10] Karp’s Conjecture

Conjecture (essentially [Karp ‘76]) ITERATED GREEDY is the asymptotically optimal clique-finding algorithm in G(n,p) In particular, k-CLIQUE has average-case circuit size Ω(nk/4) for every constant k Karp’s Conjecture

Conjecture (essentially [Karp ‘76]) ITERATED GREEDY is the asymptotically optimal clique-finding algorithm in G(n,p) In particular, k-CLIQUE has average-case circuit size Ω(nk/4) for every constant k

even up to k = 2 log n, p = ½, and even with a planted clique of size n0.49 Karp’s Conjecture

Conjecture (essentially [Karp ‘76]) ITERATED GREEDY is the asymptotically optimal clique-finding algorithm in G(n,p) In particular, k-CLIQUE has average-case circuit size Ω(nk/4) for every constant k

• Implies P ≠ NP (in a strong way) • Has motivated a lot of research in Complexity/ Crypto/Algorithms

Lower Bounds for Average-Case SUB(G) Lower Bounds

[R. STOC’08], [Amano CCC’10], [R. FOCS’10], [Li-Razborov-R. FOCS’14], [R. STOC’14], [R. CCC’15], [Kawarabayashi-R. SODA’18]

Kazuyuki Amano Yuan Li Ken-ichi Kawarabayashi Gunma University Google University of Chicago National Institute of and Steklov Institute Informatics, Tokyo Lower Bounds

• Unified method giving average-case lower bounds on the circuit size and formula size of SUB(G) (for all patterns G) in both the AC0 and monotone settings Lower Bounds

• Unified method giving average-case lower bounds on the circuit size and formula size of SUB(G) (for all patterns G) in both the AC0 and monotone settings

Complexity of SUB(G) to tied to invariants of the graph G: Θ~(tree-width(G)) circuit size = n formula size = npoly(tree-depth(G)) Lower Bounds

• Unified method giving average-case lower bounds on the circuit size and formula size of SUB(G) (for all patterns G) in both the AC0 and monotone settings

• Answers basic open questions: ² size hierarchy theorem for uniform AC0 ² separation of AC0 formulas vs. circuits ² average-case lower bound against monotone NC1 Theorem [R. ‘08] AC0 circuits solving k-CLIQUE on G(n,p) with prob. 0.51 require size Ω(nk/4) Theorem [R. ‘08] AC0 circuits solving k-CLIQUE on G(n,p) with prob. 0.51 require size Ω(nk/4)

• Confirms Karp’s conjecture for AC0

2 • Holds (without degradation) to depth log n / k log log n Theorem [R. ‘08] AC0 circuits solving k-CLIQUE on G(n,p) with prob. 0.51 require size Ω(nk/4)

• Confirms Karp’s conjecture for AC0

2 • Holds (without degradation) to depth log n / k log log n • Previous lower bound: Ω(k/d 2 ) n for depth-d circuits [Beame ‘90] Theorem [R. ‘08] AC0 circuits solving k-CLIQUE on G(n,p) with prob. 0.51 require size Ω(nk/4)

0 • Confirms Karp’s conjecture for breaks out of the previous AC “size-depth trade-offs” 2 • Holds (without degradation) to depth log n / k log log n • Previous lower bound: Ω(k/d 2 ) n for depth-d circuits [Beame ‘90] Theorem [R. ‘10] Monotone circuits solving k-CLIQUE on both G(n,p) – and G(n,p + p1 o(1)) require size Ω(nk/4) Theorem [R. ‘10] Monotone circuits solving k-CLIQUE on both G(n,p) – and G(n,p + p1 o(1)) require size Ω(nk/4)

relevant NO instances

relevant YES instances Theorem [R. ‘10] Monotone circuits solving k-CLIQUE on both G(n,p) – and G(n,p + p1 o(1)) require size Ω(nk/4)

still (barely) separable by density L

relevant NO instances

relevant YES instances • [Amano’10], [Li-Razborov-R. ‘14] extend the lower bound method to SUB(G) for arbitrary patterns G

G COLORED G-SUBGRAPH ISOMORPHISM (SUB(G)) Given a graph X and a vertex-coloring V(X) → V(G), does X contain a properly colored G-subgraph?

G X COLORED G-SUBGRAPH ISOMORPHISM (SUB(G)) Given a graph X and a vertex-coloring V(X) → V(G), does X contain a properly colored G-subgraph?

• Upper bound: ntree-width(G) + O(1)

G

COLORED G-SUBGRAPH ISOMORPHISM (SUB(G)) Given a graph X and a vertex-coloring V(X) → V(G), does X contain a properly colored G-subgraph?

• Upper bound: ntree-width(G) + O(1)

Theorem [Li-Razborov-R. ‘14] Ω~(tree-width(G)) Nearly matching n lower bound for 0 AC circuits COLORED G-SUBGRAPH ISOMORPHISM Tightly connects the circuit size of (SUB(G)) Given a graph SUB(G) to a combinatorial X and a vertex-coloring V(X) → V(G), does X parameter of a simpler object: contain a properly colored G-subgraph? “union trees” • Upper bound: ntree-width(G) + O(1)

Theorem [Li-Razborov-R. ‘14] Ω~(tree-width(G)) Nearly matching n lower bound for 0 AC circuits G b a d c G b b a d union tree � = a d c c

b a d c

b a d a

d b a a G b a d c

b a d a a c hitting set � = b b d d c c

intersects every union tree G b X = G-colored Erdős-Rényi graph a d (with a threshold family of edge prob’s {pe}e∈E(G)) c

n−1 n−3/2 n−1

n−1/2 G b X = G-colored Erdős-Rényi graph a d (with a threshold family of edge prob’s {pe}e∈E(G)) c H

n−1 n−3/2 n−1

n−1/2 G b X = G-colored Erdős-Rényi graph a d (with a threshold family of edge prob’s {pe}e∈E(G)) c H

|V(H)| Ex[ # of H-subgraphs in X ] = n ∏ pe e∈E(H) n−1 n−3/2 n−1

n−1/2 G b X = G-colored Erdős-Rényi graph a d (with a threshold family of edge prob’s {pe}e∈E(G)) c

Theorem [LRR’14] AC0 circuits solving SUB(G) on X with probability 0.51 have size at least

min max Ex [ # of H-subgraphs in X ] union tree � H ∈ �

G b X = G-colored Erdős-Rényi graph a d (with a threshold family of edge prob’s {pe}e∈E(G)) c

Theorem [LRR’14] AC0 circuits solving SUB(G) on X with probability 0.51 have size at least

min max Ex [ # of H-subgraphs in X ] union tree � H ∈ �

= max min Ex [ # of H-subgraphs in X ] hitting set � H ∈ � G b • equals nXk = /4 + O(1)G-colored when G is a k-clique Erdős-Rényi graph a d (with a threshold family of edge prob’s {pe}e∈E(G)) c • tight up to nO(1) factor [Rosenthal’18]

Theorem [LRR’14] AC0 circuits solving SUB(G) on X with probability 0.51 have size at least

min max Ex [ # of H-subgraphs in X ] union tree � H ∈ �

= max min Ex [ # of H-subgraphs in X ] hitting set � H ∈ � G b X = G-colored Erdős-Rényi graph a d (with a threshold family of edge prob’s {pe}e∈E(G)) c

Theorem 2 [LRR’14] For every pattern graph G, there exists a “threshold family” {pe}e∈E(G) such that min max Ex [ # of H-subgraphs in X ] union tree � H ∈ �

= nΘ~(tree-width(G)) G 1/2 {p } ∈ obtained from a certain 1 e e E(G) 1/2 1/2 Markov chain on G witnessing 1/2 1/2 1/2 large tree-width [Grohe-Marx’07]

Theorem 2 [LRR’14] For every pattern graph G, there exists a “threshold family” {pe}e∈E(G) such that min max Ex [ # of H-subgraphs in X ] union tree � H ∈ �

= nΘ~(tree-width(G)) Average-Case k-CYCLE k-CYCLE

Ck

1 1 2 2 3 3 ⋮ ⋮ X ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ n n k-CYCLE

In the average-case setting, Ck X is a Ck-colored Erdős-Rényi 1 graph where every edge has 1 2 probability Θ(1/n) 2 3 3 ⋮ ⋮ X ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ n n k-CYCLE

To find all paths between left and right... k-CYCLE find all paths between left and middle... k-CYCLE find all paths between middle and right... k-CYCLE and merge the resulting sub-computations k-CYCLE and merge the resulting sub-computations k-CYCLE

• Solvable by: monotone AC0 circuits of size nO(1), depth O(log k) k-CYCLE

• Solvable by: monotone AC0 circuits of size nO(1), depth O(log k) monotone AC0 formulas of size nO(log k)

∧ ∧ k-CYCLE

• Solvable by: monotone AC0 circuits of size nO(1), depth O(log k) monotone AC0 formulas of size nO(log k)

Theorem [R. 14] AC0 formulas solving average-case k-CYCLE require size nΩ(log k) k-CYCLE

• Solvable by: monotone AC0 circuits of size nO(1), depth O(log k) monotone AC0 formulas of size nO(log k)

Theorem [R. 14] AC0 formulas solving average-case k-CYCLE require size nΩ(log k)

AC0 version of “NC1 ≠ P”: sharp separation between the power of AC0 formulas vs. circuits k-CYCLE

• Solvable by: monotone AC0 circuits of size nO(1), depth O(log k) monotone AC0 formulas of size nO(log k)

Theorem [R. 14] AC0 formulas solving average-case k-CYCLE require size nΩ(log k)

Theorem [R. 15] Mon. formulas solving average-case k-CYCLE require size nΩ(log k) first (and only) super-polynomial average-case lower bound against monotone formulas under a product distribution

Theorem [R. 15] Mon. formulas solving average-case k-CYCLE require size nΩ(log k) Formula Size of SUB(G)

• SUB(G) is solvable by monotone AC0 formulas of size O(ntree-depth(G)) Formula Size of SUB(G)

• SUB(G) is solvable by monotone AC0 formulas of size O(ntree-depth(G))

K4 K3,3 P7

Credit: Wikipedia (David Eppstein) Formula Size of SUB(G)

• SUB(G) is solvable by monotone AC0 formulas of size O(ntree-depth(G))

• Combining a result of [Kawarabayashi-R. ‘18] in Graph Minor Theory with lower bounds for

SUB(Pathk), SUB(Gridk), SUB(Treek), we show:

Theorem AC0 formulas and monotone formulas Ω(tree-depth(G) const. ) solving SUB(G) require size n Formula Size of SUB(G) The complexity of SUB(G) is a • SUB(G) minor-monotone graph parameter is solvable by monotone AC0 formulas of size O(ntree-depth(G))

• Combining a result of [Kawarabayashi-R. ‘18] in Graph Minor Theory with lower bounds for

SUB(Pathk), SUB(Gridk), SUB(Treek), we show:

Theorem AC0 formulas and monotone formulas Ω(tree-depth(G) const. ) solving SUB(G) require size n Pathset Formulas

• Our lower bounds for AC0 formulas and monotone formulas are via reductions to a simplified setting of pathset formulas Pathset Formulas

• Our lower bounds for AC0 formulas and monotone formulas are via reductions to a simplified setting of pathset formulas

We show that any AC0 formula (resp. monotone formula) solving average-case SUB(G) can be (randomly) transformed into a pathset formula computing a dense pathset Pathset Formulas

• Our lower bounds for AC0 formulas and monotone formulas are via reductions to a simplified setting of pathset formulas

• A lower bound on pathset formula size [R. ‘14] is the combinatorial heart of our results Pathset Formulas

• Our lower bounds for AC0 formulas and monotone formulas are via reductions to a simplified setting of pathset formulas

• A lower bound on pathset formula size [R. ‘14] is the combinatorial heart of our results

• Ambitious goal of future work: lower bounds for unrestricted Boolean formulas by reduction to the pathset framework Summary

• SUB(G) is an important family of problems, whose complexity is tied to major open questions (P vs NP, L vs NL, …), as well as natural invariants of G (tree-width and tree-depth)

• Conjectured (e.g. [Karp 76]) to be hard-on-average under natural input distributions: (G-colored) Erdős-Rényi graphs at an appropriate threshold Summary

• Grand challenges in Circuit Complexity: P vs NP via the circuit size of k-CLIQUE NC1 vs P via the formula size of k-CYCLE

• Progress in the AC0 and monotone settings via an approach (the pathset framework) with potential to extend beyond these restrictions Thank you!