Download presentation
Presentation is loading. Please wait.
1
Switching Lemmas and Proof Complexity
Paul Beame University of Washington
2
AC0 circuits and restrictions
Unbounded fan-in AND, OR gates, NOT gates Constant depth Restrictions Partial assignment Ο on X : πβ π,π,β πΏ π = π βπ π,π Vars(π) = π βπ β Given π:πΏβ{π,π} define π β π : Vars(π) β{π,π} by π β π πβ² =π π where π π = π π β² πβVars(π) π π πβVars(π)
3
Circuit lower bounds via restrictions
To show lower bound for function π choose family/distribution on restrictions πΉ π and show that: For any small, shallow circuit πͺ, Pr πβΌ πΉ π πͺ β π is not π πππππ is small P r πβΌ πΉ π π β π is not π πππππ is large π πππππ = computed by small height decision tree Write π« π for the smallest height of any decision tree computing πΉ.
4
Decision Tree for π·πππππ(ππ, ππ, ππ, ππ)
1 π π π π π π π π π π
5
Lower bound for Parity For every restriction π, π« πππ«π’ππ² β π =|Vars π | Let πΉ π β be the set of restrictions on π,π π with Vars π =β. Theorem (HΓ₯stad): If πͺ is an AC0 circuit of size πΊ and depth π
then for β=π/ (ππ log πΊ) π
βπ there is a πβ πΉ π β s.t. π« πͺ β π β€ππ log πΊ Corollary: Depth d Parity circuits require π π π/π
ππ size.
6
HΓ₯stadβs Switching Lemma
Lemma (HΓ₯stad): Let π be a π-DNF formula on π,π π . For β=ππ we have ππ« πβΌ πΉ π β π« π β π β₯π β€ πππ π Note: If πβ² has a decision tree π» of height π then both πβ and π β² are expressible as π-DNF formulas Circuit lower bound follows by repeatedly setting π= log πΊ , π= π ππ log πΊ , and π=π+π.
7
HΓ₯stadβs Switching Lemma
Lemma (HΓ₯stad): Let π be a π-DNF formula on π,π π . For β=ππ we have ππ« πβΌ πΉ π β π« π β π β₯π β€ πππ π Proof idea (Razborov): Consider a canonical way of converting π β π to a decision tree. Use existence of a path of length β₯π in canonical decision tree for π (plus knowledge of π) to find a more efficient description of Ο. Conclude that restrictions with this property are rare.
8
Applying a restriction to a DNF
For a DNF π and restriction π, π β π is a DNF that contains a reduced form of each term of π not falsifed π by removing all literals set to 1, keeps the same order among all the terms that survive, and has no other terms.
9
Canonical decision tree for a DNF
Given an ordered set of variables π½= π π , π π ,β¦, π π the full tree on π½ is: A full binary tree of height π that queries the πth variable in π½ at every node on the πth level. with a branch for each total assignment π
to π½. π π π π π π π π
10
Canonical decision tree for a DNF
Let DNF π=π»β¨πβ² where π» is the first term of π. The canonical decision tree for π consists of: the full tree π« π» for Vars(π») at the root. for the unique assignment π to Vars(π») that sets π» to 1, the branch of π« π» labelled π has leaf label 1 for every other assignment π
to Vars π» , the branch of π« π» labelled π
has the canonical decision tree for πβ² β π
at its leaf.
11
Switching Lemma Proof Suppose β a path of length π in canonical decision tree for π β π Pick leftmost such path π
= π
π π
π β¦ π
π Say that 1st term of π that survives in π β π is π» π π π
π assigns to Vars( π» π π β π ) π π sets (π» π π β π ) β π π =π Repeat for π=π,β¦,π Canonical decision tree for π β π π
π π
π π
π π
π π
π π π π π π π π π π π 1
12
Switching Lemma Proof Say that 1st term of π that survives in π β π π
π is π» π π π
π assigns to Vars( π» π π β π π
π ) π π sets (π» π π β π π
π ) β π π =π β¦ Obtain π π ,π π ,β¦, π π and indices of the terms in π, π» π π ,π» π π ,β¦, π» π π π π π
π 1 π π π
π π
π 1 π π π
π 1 π π π
π 1 π π 1 Canonical decision tree for π β π
13
Switching Lemma Proof Obtain π π ,π π ,β¦, π π and indices of the terms in π, π» π π ,π» π π ,β¦, π» π π Specify π β² =π π π β¦ π π Specify where the π vars set by π
, π are in π» π π ,π» π π ,β¦, π» π π Specify values set by π
Claim: π plus 1, 2, & let us reconstruct π Canonical decision tree for π β π π
π π
π π
π π
π π
π π π π π π π π π π π 1
14
Switching Lemma Proof Specify π β² =π π π β¦ π π
Specify where the π vars set by π, π
are in π» π π ,π» π π ,β¦, π» π π Specify values set by π
Claim: π plus 1, 2, & 3 let us reconstruct π 1. π β² β πΉ π ββπ so | πΉ π ββπ | possibilities Bit vector of length πβπ for partition of [π] into π π , π π ,β¦, π π plus elt of [π] π to indicate positions of each var in its term. Total π πβπ π π Bit vector of length π. Total π π So: less than ππ π | πΉ π ββπ | of the π have such a bad path
15
Switching Lemma Proof So: ππ« πβΌ πΉ π β π« π β π β₯π β€ ππ π πΉ π ββπ πΉ π β = ππ π π ββπ π πββ+π π β π πββ = ππ π ββ¦(ββπ+π) πββ+π β¦(πββ+π) β€ ππ π π πβπ π Can do better with a sharper analysis since β=ππ
16
Switching Lemma Proof Specify π β² =π π π β¦ π π
Specify where the π vars set by π, π
are in π» π π ,π» π π ,β¦, π» π π Specify values set by π
Claim: π plus 1, 2, & 3 let us reconstruct π Define π π =π π
π β¦ π
πβπ π π β¦ π π π π = π β² π» π π is 1st term of π surviving in π β π π
π β¦ π
πβπ and π» π π β π π =π Can build π π+π from π π and π plus π π , 2, & 3 Once all π π β¦ π π are known they can be removed to get π from πβ²
17
Proof Complexity Switching Lemmas
18
Pigeonhole formulas and matchings
For PH P π : variables π ππ for πβ π+π , πβ π Β¬( π ππ β¨β¦β¨ π ππ ) for πβ[π+π] Β¬( π ππ β¨ π ππ ) for π,πβ π+π , πβ[π] Take disjunction, add dual terms, get bijective PHP tautology Restrictions that do not simplify PH P π too much are matching restrictions. For π β partial bipartite matching π΄(π) s.t.: π sets π ππ to 1 for all π,π βπ΄(π) π sets π ππ to 0 for all π,π matched in π΄(π) but not to each other π leaves all other variables unset Define π = π΄ π Vars π ={ π ππ :π,π both not matched in π΄(π)} For matching restriction π, PH P π β π β‘PH P πβ π
19
Matching decision trees
Each path or branch π
is a matching restriction. Each node queries a currently unmatched vertex on the left or right and asks which vertex it is matched to One child node for each possible match Queried vertex will always be matched Leaf labels 0 or 1 Unlike Boolean decision trees, not all assignments have branches consistent with them
20
A matching decision tree on 4 Γ[3]
?π π? ππ ππ ππ ππ 1 ?π ππ ππ ππ ππ ππ π? ππ π? ππ ?π ππ Associated DNF (π ππ π ππ β¨ π ππ π ππ π ππ β¨ π ππ π ππ β¨ π ππ π ππ β¨ π ππ π ππ β¨ π ππ π ππ β¨ π ππ π ππ β¨ π ππ π ππ π ππ )
21
Matching disjunctions
A π-matching disjunction is a monotone π-DNF each of whose terms is a (bipartite) matching For π» a matching decision tree Let B r π π» = set of branches of π» with leaf label π. Let π·ππΉ π» = π
βB r π (π») π,π βπ΄(π
) π ππ If π» has height π then π·ππΉ(π») is a π-matching disjunction Let π» π be the same decision tree as π» but with leaf labels π and π swapped. A matching decision tree π» represents a matching disjunction π iff every branch π
in π» has π β π
=π iff π
has leaf label π
22
Full matching decision trees
Suppose that π« and πΉ are ordered sets. For a subset π½βπ«βͺπΉ, the full matching tree π» π½ for π½ on π«ΓπΉ is given by the following: If π½=β
then π» π½ consists of a single root node. If π½β©π«β β
, let π be 1st node of π« in π½; π» π½ has root with query π? and edges for all possible πβπΉ; edge labelled ππ leads to root of π» π½β π,π . Else π½β©πΉβ β
so let π be 1st node of πΉ in π½; then π» π½ has root with query ?π and edges for all possible πβπ«; edge labelled ππ leads to root of π» π½β π,π If π΄ is a matching, write π» π΄ for full matching tree for set π½ of nodes matched by π΄
23
Full matching tree on 1,2 β 4 Γ[3]
1? ππ ππ ππ ?π ?π ππ ππ ππ ππ ππ ππ
24
Full matching tree on 1,2 ,(2,1) β 4 Γ[3]
1? ππ ππ ππ π? ππ ππ π? ππ ππ π? ππ ππ ?π ππ ππ ππ ππ ?π ππ ππ ?π ?π ππ ππ
25
Canonical matching decision tree
Given a matching disjunction π=π»β¨πβ² where π»= π,π βπ΄ π ππ is the 1st term of π, the canonical matching decision tree for π consists of: the full matching tree π» π΄ for matching π΄ at the root for the matching restriction π with π΄ π =π΄, the branch of π» π΄ labelled π has leaf label 1 for every other matching restriction π
labelling a branch of π» π΄ , the branch labelled π
has the canonical matching decision tree for πβ² β π
at its leaf. Note: The canonical matching decision tree for π represents π
26
Intuition for PH P π lower bound
Assume wlog proof connectives are unbounded β¨,Β¬ Associate a height π matching decision tree π» π¨ for each subformula π¨ appearing in the proof s.t.: π» π = π» π = π» π ππ is the full matching tree on π,π with leaf label 1 at level 1 (ππ) and leaf label 0 at level 2. π» Β¬π¨ = π» π¨ π If π¨= β π¨ β and each π¨ β does not have an β¨ at its root then represents β π·ππΉ( π» π¨ β ) βπ-evaluationβ. Any tree with all leaves labelled 1 approximates true; all leaves labelled 0 β false If inference rule size β€π and π< π/π then all lines are β true but PH P π is β false 1
27
Matching Switching Lemma
Let π΄ π β denote the set of matching restrictions π on π+π Γ π with π =πββ. i.e., Vars π =π«ΓπΉ where πΉ =β For any matching disjunction π, let π« π΄ (π) be the minimum height of any matching decision tree representing π. Lemma (BIKPPW): Let π be any π-matching disjunction on [π+π]Γ[π]. Then for ππβ β+π β€πββ Pr πβΌ π΄ π β π« π΄ π β π β₯π β€ ππ β π (β+π) πββ π/π
28
Matching Switching Lemma Proof
Suppose β a path of length π in canonical matching decision tree for π β π Pick leftmost such path π
= π
π π
π β¦ π
π Say that 1st term of π that survives in π β π is π» π π π
π assigns to all vertices of matching in π» π π β π π π sets (π» π π β π ) β π π =π Repeat for π=π,β¦,π π π π
π 1 π π π
π π
π 1 π π π
π 1 π π π
π 1 π π 1 Canonical matching decision tree for π β π
29
Matching Switching Lemma Proof
Obtain π π ,π π ,β¦, π π and indices of the terms in π, π» π π ,π» π π ,β¦, π» π π Specify π β² =π π π β¦ π π Specify where the β€π vars set by π are in π» π π ,π» π π ,β¦, π» π π Specify values set by π
Claim: π plus 1, 2, & let us reconstruct π Canonical matching decision tree for π β π π
π π
π π
π π
π π
π π π π π π π π π π π 1
30
Matching Switching Lemma Proof
Specify π β² =π π π β¦ π π Specify where the β€π vars set by π are in π» π π ,π» π π ,β¦, π» π π Specify values set by π
Claim: π plus 1, 2, & 3 let us reconstruct π Define π π =π π
π β¦ π
πβπ π π β¦ π π π π = π β² π» π π is 1st term of π surviving in π β π π
π β¦ π
πβπ and π» π π β π π =π Can build π π+π from π π and π plus π π , 2, & 3 Once all π π β¦ π π are known they can be removed to get π from πβ²
31
Matching Switching Lemma Proof
Specify π β² =π π π β¦ π π Specify where the πβ²β€π vars set by π are in π» π π ,π» π π ,β¦, π» π π Specify values set by π
Claim: π plus 1, 2, & 3 let us reconstruct π 1. # of vars π β² set by π satisfies π π β€ π β² β€π and π β² β π΄ π ββ π β² so π/πβ€ π β² β€π | π΄ π ββ π β² | possibilities Bit vector of length πβ²βπ for partition of [πβ²] into π π , π π ,β¦, π π plus elt of [π] π β² to indicate positions of each var in its term. Total β€π π β² βπ π π β² Matching of length π touching all vertices of π. Total β π So: less than π βπ β π π/πβ€ π β² β€π (ππ) π β² | π΄ π ββ π β² | of the π have such a bad path
32
Matching Switching Lemma Proof
| π΄ π β |= π β π+π β¦(β+π) π΄ π ββ π β² π΄ π β = ββ¦ ββ π β² +π πββ+ π β² β¦ πββ+π β+π β¦(ββ π β² +π) β€ β(β+π) πββ π β² So since ππβ β+π β€πββ the probability of height at least π is at most π βπ β π π/πβ€ π β² β€π ππ π β² | π΄ π ββ π β² | π΄ π β β€ β π ππβ(β+π) πββ π/π β€ ππ β π (β+π) πββ π/π
33
Beyond Bipartite Matching
Similar switching lemmas for Matching restrictions on the variables of the complete graph π² ππ+π [B-Pitassi 1996] π-hypergraph matching restrictions on complete π-regular hypergraphs over [ππ+π] for πβ€π<π [B-Riis 1998] Switching lemmas for πΉππ(π): [Buss, Impagliazzo, Segerlind 2004] Set a polynomially-small fraction of inputs. [Razborov 2017] Switching lemmas for Tseitin formulas: [Pitassi, Rossman, Servedio, Tan 2016] [HΓ₯stad 2017]
34
Multi-Switching Lemma
35
Using Hastadβs Switching Lemma
Lemma (HΓ₯stad): Let π be a π-DNF formula on π,π π . For β=ππ we have ππ« πβΌ πΉ π β π« π β π β₯π β€ πππ π Use it for π different π-DNFs at once by setting π=π/(πππ) and π=logβ‘π+π to get failure probability at most π πππ π β€π/π. To get better failure probability need larger π. Alternative: A multiswitching lemma.
36
What multiswitching produces
A single decision tree π» of height π is a common π-partial decision tree for π-DNF formulas π π , π π ,β¦, π π iff for every branch π of π», for every πβ π , π« π π β π <π. Define π·( π π ,β¦, π π , π,π) to be the property that no such tree exists. Multiswitching Lemma (HΓ₯stad 2014*): Let π π , π π ,β¦, π π be π-DNF formulas on π,π π . For β=ππ, integer π, and πβ₯ π₯π¨π π we have ππ« πβΌ πΉ π β [ π·(π π β π , β¦, π π β π ,π,π)]β€π (ππππ) π . *related lemma shown in [Impagliazzo, Matthews, Paturi 2012]
37
What multiswitching produces
Multiswitching Lemma (HΓ₯stad 2014): Let π π , π π ,β¦, π π be π-DNF formulas on π,π π . For β=ππ, integer π, and πβ₯ π₯π¨π π we have ππ« πβΌ πΉ π β [ π·(π π β π , β¦, π π β π ,π,π)]β€π (ππππ) π . Corollary: Let π π , π π ,β¦, π π be π-DNF formulas on π,π π . For β=ππ and πβ₯ π₯π¨π π there is a distribution on restrictions π depending on π π , π π ,β¦, π π such that: βπβπ. |Vars(π)|=β ππ« πβΌπ βπβ π .π«(π π β π β₯π]β€π (ππππ) β .
38
Proof Multiswitching Lemma
The original proof used conditional probability using arbitrary downward-closed conditions and random restrictions with independent probabilities over the bits. We give a bit simpler proof using similar ideas to the proof of the basic lemma. Suppose that π·(π π β π , β¦, π π β π ,π,π) holds. We will give a specific construction for a decision tree π» and bound that probability that it fails to be a height-π π-partial decision tree for π π β π , β¦, π π β π .
39
Multiswitching Lemma Proof
Given π,π π , β¦, π π let π π be the first index such that |π« π» π π π β π |β₯π Let π
π be the lexicographically first path in the canonical decision tree for π π π β π of length π. Define π π = π π π β¦ π π π as before, except make sure that π π π reaches the 1-leaf. Then π β² = π π β₯π and π β² <π+π. π
π is enough to determine π π . Begin building tree π» by with complete tree of height πβ² on the variables set by π π For every path π π path of length πβ in π» continue as follows:
40
Multiswitching Lemma Proof
Given π, π π ,π π , β¦, π π let π π be the first index such that |π« π» π π π β π π π |β₯π We have π π β₯ π π Let π
π be the lexicographically first path in the canonical decision tree for π π π β π π π of length π. Define π π = π π π β¦ π π β² π as before, except make sure that π π β² π reaches the 1-leaf. Then π β²β² = π π β₯π and π β²β² <π+π. π
π is enough to determine π π . Continue tree π» at leaf π π with complete tree of height πβ²β² on the variables set by π π For every path π π π π in π» continue and repeat until no more indices.
41
Multiswitching Lemma Proof
Failure occurs iff there is a path of length β₯π in π» Let π be lex first such path Let π π ,β¦, π π correspond to π π ,β¦, π π . Define π β² =π π π β¦ π π β πΉ π ββπ As before we include info to decode π from πβ²: π π ,β¦, π π π vars within terms of π π π set by π
π for each πβ[π] Bit vector of values set by π
π β¦ π
π Bit vector of values set by π π β¦ π π π π π π π π π π π π π π π π π π π π π π Tree π»
42
Multiswitching Lemma Proof
Define π β² =π π π β¦ π π β πΉ π ββπ . Info to decode π from πβ²: π π ,β¦, π π β[π] π vars within terms of π π π set by π
π for each πβ[π] Bit vector of values set by π
π β¦ π
π Bit vector of values set by π π β¦ π π Claim: this is enough to decode given π π ,β¦, π π ,π,π: Look at first term π» π π set to 1 in π π π β π β² (previous terms will be set to 0). This yields π π π . Use the information from 2 to determine which vars of π» π π are unset by π and hence set by π
π π . Use 3 to determine π
π π . Continue as in the ordinary switching lemma decoding by replacing π π π by π
π π in π β² to get π β²β² . Then look at 1st term π» π π set to 1 in π π π β π β²β² etc. Continue replacing until get all of π π . Now go back and replace π π by π π in πβ² to get π π π π π β¦ π π , using 4.
43
Multiswitching Lemma Proof
Define π β² =π π π β¦ π π β πΉ π ββπ . Info to decode π from πβ²: π π ,β¦, π π β[π] π vars within terms of π π π set by π
π for each πβ[π] Bit vector of values set by π
π β¦ π
π Bit vector of values set by π π β¦ π π Counts: π π β€ π πβπ π πβ€πβ
π π since πβ₯ π₯π¨π π and π πβπ β€π π π β² βπ π π β² where π β² β€π is the total length of π
π β¦ π
π Total β€ π π π π π π β² π π Total fraction of restrictions where π·( π π ,β¦, π π , π,π) holds is at most: πππ π |πΉ π ββπ | πΉ π β =π πππ π π ββπ π πββ+π π β π πββ = π πππ π ββ¦(ββπ+π) πββ+π β¦(πββ+π) β€π ππππ πβπ π
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.