Almost Everywhere: Naturally Occurring Arsenic in Wisconsinβs Aquifers
Almost-Everywhere Lower Bounds - MIT
Transcript of Almost-Everywhere Lower Bounds - MIT
Almost-Everywhere Circuit Lower Bounds from Circuit-Analysis Algorithms
Ryan WilliamsMIT
But all βheavy-liftingβ done by: Xin Lyu (Tsinghua) and Lijie Chen (MIT)
Outline
β’ Prior Work and a βSubtleβ Issue
β’ What We Do
β’ A Little About How We Do It
β’ Conclusion
Algorithmic Approach to Lower Bounds: Interesting circuit-analysis algorithms
tell us about the limitations of circuits in modeling algorithms
β
β
βNon-Trivialβ Circuit Analysis
Algorithm(beating brute force) Circuit Lower Bounds
SAT? YES/NO ββinterestingβπ
β
β Circuits are not βblack-boxesβ to algs!
Inherently non-relativizing
approach
Turing Machine drawing by Tom Dunne for American Scientist
Circuit-Analysis Problem #1:Generalized Circuit Satisfiability
Let C be a class of Boolean circuits
C = {formulas}, C = {arbitrary circuits}, C = {3CNFs}
A very βsimpleβ circuit analysis problem
[CLβ70s] C-SAT is NP-complete for practically all interesting CC-SAT is solvable in O(2n |K|) time by brute force
The C-SAT Problem:Given a circuit K(x1,β¦,xn) from C, is there an
assignment (a1, β¦, an) β {0,1}n such that K(a1,β¦,an) =1?
Circuit-Analysis Problem #2:Gap Circuit Satisfiability
Let C be a class of Boolean circuits
C = {formulas}, C = {arbitrary circuits}, C = {3CNFs}
Even simpler! In randomized polynomial time
[Folklore?] Gap-Circuit-SAT β P β P = RP [Hirsch, Trevisan, β¦] Gap-kSAT β P for all k
Gap-C-SAT:Given π²(x1,β¦,xn) from C, and the promise that either
(a) π² β‘ 0, or (b) π·ππ π² π = π β₯ π/π,decide which is true.
Nontrivially Faster C-SAT βΉ Circuit Lower Bounds for C
Slightly Faster Circuit-SAT[R.W. β10,β11]
Deterministic algorithms for:β’ Circuit SAT in O(2n/n10) timewith n inputs and nk gates, for all k
β’ Formula SAT in O(2n/n10) time
β’ πͺ-SAT in O(2n/n10) time
β’ Gap-πͺ-SAT in O(2n/n10) time on nk size, for all k
(Easily solved w/ randomness!)
No βCircuits for NEXPβ
Would imply:
β’ NEXP P/poly
β’ NEXP Poly-size formulas
β’ NEXP poly-size πͺ
NEXP poly-size πͺ
Concrete LBs:πͺ = ACC[Wβ11]πͺ = ACC of THR[Wβ14]
Even Faster SAT βΉ Stronger Lower Bounds
Somewhat Faster Circuit SAT[Murray-W. β18]
Det. algorithm for some π > π:
β’ Circuit SAT in O(2πβππ) time
with π inputs and 2ππgates
β’ Formula SAT in O(2πβππ) time
β’ πͺ-SAT in O(2πβππ) time
β’ Gap-πΆ-SAT in O(2πβππ) time
on 2ππgates
No βCircuits for Quasi-NPβ
Would imply:
β’ NTIME[ππππππππ π] P/poly
β’ NTIME[ππππππππ π] NC1
β’ NTIME[ππππππππ π] πͺ
NTIME[ππππππππ π] πͺ
πͺ = ACC of THR[MWβ18]
βFine-Grainedβ SAT Algorithms[Murray-W. β18]
Det. algorithm for some π > π:
β’ Circuit SAT in O(2(1βπ)π) timeon π inputs and 2ππ gates
β’ FormSAT in O(2(1βπ)π) time
β’ πͺ-SAT in O(2(1βπ)π) time
β’ Gap-πͺ-SAT is in O(π πβπ π) time on 2ππ gates
(Implied by PromiseRP in P)
No βCircuits for NPβ
Would imply:
β’ NP SIZE(ππ) for all π
β’ NP Formulas of size ππ
β’ NP πͺ-SIZE(ππ) for all π
NP πͺ-SIZE(ππ) for all π
Even Faster SAT βΉ Stronger Lower Bounds
πͺ = SUM of THRπͺ = SUM of ReLUπͺ = SUM of low-
degree polys[Wβ18]
Note: Would refute
Strong ETH!
Strongly believed to
be trueβ¦
[R.Chen-Oliveira-Santhanamβ18, Chen-Wβ19, Chenβ19, Chen-Ren β20]
Det. algorithm for some π > π:
β’ #Circuit SAT in O(2πβππ) time
with π inputs and 2ππgates
β’ #Formula SAT in O(2πβππ) time
β’ #πͺ-SAT in O(2πβππ) time
β’ πͺ-CAPP in O(2πβππ) time
No Circuits for Computing Quasi-NP on Average
Would imply:
β’ NTIME[ππππππππ π] canβt be
(1/2 +1/poly)-approximated in P/poly
β’ Inapproximability in NC1
β’ Inapproximability in πͺ/poly
Faster #SAT and CAPP βΉ Average-Case Lower Bounds
πͺ = ACC of THR[Chen-Renβ20]
Given a circuit of size s, approximate its fraction of SAT
assignments to within +- 1/s
[R.Chen-Oliveira-Santhanamβ18, Chen-Wβ19, Chenβ19, Chen-Ren β20]
Det. algorithm for some π > π:
β’ #Circuit SAT in O(2πβππ) time
with π inputs and 2ππgates
β’ #Formula SAT in O(2πβππ) time
β’ #πͺ-SAT in O(2πβππ) time
β’ πͺ-CAPP in O(2πβππ) time
No Circuits for Computing Quasi-NP on Average
Would imply:
β’ NTIME[ππππππππ π] canβt be
(1/2 +1/poly)-approximated in P/poly
β’ Inapproximability in NC1
β’ Inapproximability in πͺ/poly
Faster #SAT and CAPP βΉ Average-Case Lower Bounds
πͺ = ACC of THR[Chen-Renβ20]
There is an π β NTIME[ππππππππ π] such that, for infinitely many π, every poly(π)-size
circuit πΆ fails to compute ππ on more than 1
2+
1
poly π2π inputs.
Given a circuit of size s, approximate its fraction of SAT
assignments to within +- 1/s
A Subtle (But Important) Issue!When we prove statements like NEXP β ACC0 via circuit-analysis algorithms,
we end up showing that, for NEXP-complete problems such as Succinct3SAT, there are infinitely many input lengths π such that Succinct3SAT fails to have the desired ACC circuits on length-π inputs.
Let π: π, π β β {π, π} and let ππ: π, π π β {π, π} be the restriction of π
An infinitely-often circuit lower bound only says βππ doesnβt have small circuitsββ for infinitely many π:
ππ, ππ, ππ, ππ, β¦ . , ππππ, β¦ β¦ β¦ . , πππππ, β¦ . . , ππππππ, β¦ . .
We would greatly prefer an βalmost-everywhereβ circuit lower bound, which holds for all but finitely many inputs!
ππ, ππ, ππ, ππ, β¦ . , ππππ, β¦ β¦ β¦ . , πππππ, β¦ . . , ππππππ, β¦ . .
All of the classical circuit lower bounds from the 1980s (PARITY β AC0, MAJORITY β AC0[2], etc.) yield almost-everywhere lower bounds.
A Subtle (But Important) Issue!
Why does the algorithmic approach only get infinitely-often lower bounds?
Prior work relies on other lower bounds such as the nondeterministic time hierarchy theorem or MA/1 circuit lower bounds, and neither results are known to hold almost-everywhere.
If we knew (for example)
NTIME[ππ] is not infinitely often in NTIME[ππ/ππππ(π)],
then we could conclude some kind of almost-everywhere lower bound.
But there are oracles relative to which NEXP is infinitely often in NP! [Buhrman-Fortnow-Santhanam β09]
A Subtle (But Important) Issue!
Why does the algorithmic approach only get infinitely-often lower bounds?
Prior work relies on other lower bounds such as the nondeterministic time hierarchy theorem or MA/1 circuit lower bounds, and neither results are known to hold almost-everywhere.
If we knew (for example)
NTIME[ππ] is not infinitely often in NTIME[ππ/ππππ(π)],
then we could conclude some kind of almost-everywhere lower bound.
But there are oracles relative to which NEXP is infinitely often in NP! [Buhrman-Fortnow-Santhanam β09]
βThere are functions in NTIME[ππ] that are so hard, no nondeterministicalgorithm running in ππ/ππππ π time can correctly compute the
function on any infinite sequence of input lengthsβ
Outline
β’ Prior Work and a βSubtleβ Issue
β’ What We Do
β’ A Little About How We Do It
β’ Conclusion
[R.Chen-Oliveira-Santhanamβ18, Chen-Wβ19, Chenβ19, Chen-Ren β20]
Det. algorithm for some π > π:β’ C-SAT (or Gap-C-SAT) with π
inputs and π π π 1 gates in
ππ/ππ π time
β’ #πͺ-SAT (or πͺ-CAPP) in
O(ππβππ) time on 2ππ
gates
A.E. Circuit Lower Boundsfor π¬π΅π· on Average
Would imply:
β’ π¬π΅π· does not have π(π/π) size
C-circuits, for almost every π
β’ π¬π΅π· canβt be π/π + π/πππ π-
approximated with πππ πsize
C-circuits, for a.e. π
This Work:Faster SAT βΉ Almost-Everywhere Lower Bounds
πͺ = ACC of THR
π(π) = πππ π
Almost-everywhere average-case
lower bounds for ACC of THR!
There is an π β TIME[ππΆ(π)]SAT such that, for all but finitely many π, every s(π)-size circuit πΆ fails to compute ππ on more than
1
2+
1
s π2π inputs.
Given a circuit of size s, approximate its fraction of SAT
assignments to within +- 1/s
[R.Chen-Oliveira-Santhanamβ18, Chen-Wβ19, Chenβ19, Chen-Ren β20]
Det. algorithm for some π > π:β’ C-SAT (or Gap-C-SAT) with π
inputs and π π π 1 gates in
ππ/ππ π time
β’ #πͺ-SAT (or πͺ-CAPP) in
O(ππβππ) time on 2ππ
gates
A.E. Circuit Lower Boundsfor π¬π΅π· on Average
Would imply:
β’ π¬π΅π· does not have π(π/π) size
C-circuits, for almost every π
β’ π¬π΅π· canβt be π/π + π/πππ π-
approximated with πππ πsize
C-circuits, for a.e. π
This Work:Faster SAT βΉ Almost-Everywhere Lower Bounds
πͺ = ACC of THR
π(π) = πππ π
Given a circuit of size s, approximate its fraction of SAT
assignments to within +- 1/s
More Almost-Everywhere GoodnessIn fact, we can extend all previous βENP lower boundsβ proved via the
algorithmic method to the almost-everywhere setting.
Rigid matrices in πππ: There is a PNP algorithm π such that, for all but finitely many π, π on input 1π outputs an π Γ π matrix ππ satisfying β
2log1βπ π ππ = Ξ© π2 .
Extends [Alman-Cβ19], [Bhangale-Harsha-Paradise-Talβ20]
Probabilistic degree lower bounds:There is an ENP language πΏ such that, for all but
finitely many π, πΏπ does not have π π /log2 π -degree probabilistic π 2-
polynomials. Extends [Violaβ20]
Strong average-case ππππ lower bounds: Extends [Chen-Wβ19], [Chen-Renβ20]
with better inapproximability parameters
Correlation bounds: For all π > 0, and for all but
finitely many π, πΏπ cannot be 1
2+
1
2πΞ© 1
approximated by π1βπ-degree π 2-polynomials. Extends [Violaβ20]
Theorem: There is an ENP function π, such that for all sufficiently large π,
ππ cannot be 1
2+ 2βππ 1
-approximated by 2ππ 1-size ACC0 circuits.
[Chen-Renβ20] There is a function π in QuasiNP such that, for infinitely many π, every ACC0 circuit πΆ of size poly π cannot
1
2+
1
poly π-approximate ππ.
Extension to Average Case
βNewβ XOR Lemma: Suppose there is no ππππ¦ π -size linear combination πΏ of πͺ-circuits for f such that πΈπ₯ πΏ π₯ β π π₯ < 1/10.
Then πβπ cannot be 1
2+
1
π -approximated by size-π πͺ-circuits.
(Follows Levinβs proof of the XOR Lemma)π₯1, β¦ , π₯π β¦ π π₯1 β β― β π π₯π
Outline
β’ Prior Work and a βSubtleβ Issue
β’ What We Do
β’
β’ Conclusion
A Little About How We Do It
A Little About How We Do It
β’ How NEXP β ACC0 Was Proved
β’ Another View of the Proof
β’ Extending to Almost-Everywhere
How NEXP β ACC0 Was Proved
Let β be a βtypicalβ circuit class (like ACC0)
Thm A [Wβ11] (algorithm design β lower bounds)If for all k, Gap-β-SAT on nk-size is in O(2n/nk) time, then NEXP does not have poly-size β-circuits.
Thm B [Wβ11] (algorithm)
β β°, #ACC0-SAT on ππβsize is in O(ππβπβ
) time.(Used a well-known representation of ACC0 from 1990, that people long suspected should imply lower bounds)
Note that Theorem B gives a stronger algorithm than necessary in the hypothesis of Theorem A.
(This is the starting point of [Murray-Wβ18] which proves Quasi-NP lower bounds, and other subsequent work)
Idea of Theorem ALet β be some circuit class (like ACC0)
Thm A [Wβ11] (algorithm design β lower bounds)If for all k, Gap β-SAT on nk-size is in O(2n/nk) time, then NEXP does not have poly-size β-circuits.
Idea. Show that if we assume both:
(1) NEXP has poly-size β-circuits, AND
(2) a faster Gap β-SAT algorithm
Then we can show NTIME[ππ] β NTIME[o(ππ)].This contradicts the nondeterministic time hierarchy:thereβs an π³ππππ in NTIME[ππ] β NTIME[o(ππ)]
Proof Ideas in Theorem AIdea. Assume
(1) NEXP has poly-size β-circuits, AND
(2) thereβs a faster Gap β-SAT algorithm
Show that NTIME[ππ] β NTIME[o(ππ)] (contradiction)
Take any problem πΏ in nondeterministic ππ time. Given an input π₯, we decide πΏ on π₯ by:
1. Guessing a witness π¦ of O(ππ) length.
2. Checking π¦ is a witness for π₯ in O(ππ) time.
Want to βspeed-upβ both parts 1 and 2, using the above assumptions
Proof Ideas in Theorem AIdea. Assume
(1) NEXP has poly-size β-circuits, AND
(2) thereβs a faster Gap β-SAT algorithm
Show that NTIME[ππ] β NTIME[o(ππ)]
Take any problem L in nondeterministic ππ time. Given an input π₯, we decide L on π₯ in a FASTER way:
1. Use (1) to guess a witness π of o(ππ) length (Easy Witness Lemma [IKW02]: if NEXP is in P/poly, then L has βsmall witnessesβ)
2. Use (2) to check π is a witness for π in o(ππ) timeTechnical: Use a highly-structured PCPs for NEXP [Wβ10, BVβ14] to reduce the check to Gap β-SAT
Extend to Almost-Everywhere?Idea. Assume
(1) NEXP has poly-size β-circuits, AND
(2) thereβs a faster Gap β-SAT algorithm
Show that NTIME[ππ] β NTIME[o(ππ)] ?
Take any problem L in nondeterministic ππ time. Given an input π₯, we decide L on π₯ in a FASTER way:
1. Use (1) to guess a witness π of o(ππ) length (Infinitely-Often Easy Witness Lemma [???]: if NEXP is in io-P/poly, then L has βsmall witnessesβ ?)
2. Use (2) to check π is a witness for π in o(ππ) timeTechnical: Use a highly-structured PCPs for NEXP [Wβ10, BVβ14] to reduce the check to Gap β-SAT
Even if we could proveNTIME[ππ] β io-NTIME[o(ππ)],
We still donβt know how to complete step 1!
INFINITELY OFTEN^
NT[ππ] β io-NT[o(ππ)] and
π¬πΏπ·π΅π· β io-βwould imply our desired
easy witnesses. We could infer a contradiction!
But such an NTIME hierarchy looks very hard to prove⦠what to do??
A Little About How We Do It
β’ How NEXP β ACC0 Was Proved
β’ Another View of the Proof
β’ Extending to Almost-Everywhere
Another View of the ProofNTIME hierarchy β There is a function πβπππ β πππππ ππ β πππππ ππ/π
Consider a βcanonicalβ algorithm for πβπππ:
ππ‘ππ«π(π):1. Guess a witness π¦ of O(ππ) length.2. Check π¦ is a witness for π₯ in O(ππ) time.
Consider an algorithm that tries to βcheatβ in the computation of πβπππ, by only
verifying witnesses that are βcompressibleβ by small ACC0 circuits.
ππππππ(π):
1. Guess a πππ π-size ACC0 circuit
πΆ: 0,1 π β 0,1 .2. Check the truth-table of πΆ is a
witness for π₯, in ππ/ππ π time.
NTIME hierarchy β πcheat fails to compute πβπππ on infinitely many inputs
β There are infinitely many π such that ππππππ π = π and πππππ π = π
For each such π₯, every valid witness for πhard π₯ is a hard function:it cannot be computed by small ππππ circuits!
Another View of the Proof
Can use this to construct an π¬π΅π·/π algorithm with no small ππππ circuits:
For each such π₯, every valid witness for πhard π₯ is a hard function:it cannot be computed by small ππππ circuits!
There are infinitely many π such that ππππππ π = π and πππππ π = π
Finally, we can βremoveβ the advice by just considering an π¬π΅π· algorithm that takes π, π as input. This will also have no small ππππ circuits.
What was gained by this perspective??? (We already had NEXP not in ππππ)
Vague Idea: Can we use another hierarchy? Can we βconstructβ these bad ππ?
Input: an π-bit index π β {π, π}π.
Advice: an π-bit string ππ such that πcheat ππ = 0, πhard π₯π = 1.Output: Repeatedly call an NP oracle to find the lexicographically first witness πsuch that πππππ ππ = π, and output the π-th bit of π.
A Little About How We Do It
β’ How NEXP β ACC0 Was Proved
β’ Another View of the Proof
β’ Extending to Almost-Everywhere
Extending to Almost-Everywhere
Recall: It is open if there is an π β NTIME 2π β io-NTIME π(2π)
Idea: Start from a restricted almost-everywhere NTIME hierarchy
NTIMEGUESS[π(π), π(π)]: languages that can be decided by nondeterministic
algorithms running in π π π time and guessing at most π π bits of witness.
Theorem [Fortnow-Santhanam 2016]
NTIME π π β io-NTIMEGUESS π π π , π πFor time-constructible π π , thereβs a function decidable in π(π π )nondeterministic time that cannot be decided, even infinitely often,
by any π π π -time algorithm using π(π) bits of guessing.
[FSβ16] There is a function πππππ β πππππ ππ β π’π¨βππππππππππ π ππ , π π
Consider a βcanonicalβ algorithm for πβπππ:
ππ‘ππ«π(π):1. Guess a witness π¦ of O(ππ) length.2. Check π¦ is a witness for π₯ in O(ππ) time.
As before, we consider an algorithm that tries to βcheatβ to compute πβπππβ¦
Let π = π log π .
ππππππ(π):
1. Guess a πππ π-size ACC0 circuit
πΆ: 0,1 π β 0,1 .2. Check the truth-table of πΆ is a
witness for π₯, in π ππ time.
[FSβ16] β for a.e. π, πcheat fails to compute πβπππ on some input of length π
β For a.e. π, thereβs an π β π, π π such that ππππππ π = π and πππππ π = π
For each such π₯, every valid witness for πhard π₯ is a hard function:it cannot be computed by small ππππ circuits!
Does it Just Work??
2ππ 1β€ π π
π 2π β€ π ππ
Does it Just Work??
What happens when we try the same π¬π΅π· algorithm again?
For each such π₯, every valid witness for πhard π₯ is a hard function:it cannot be computed by small ππππ circuits!
Input: an π-bit index π β {π, π}π, recall π = π log π
Advice: an π-bit string ππ such that πcheat π₯π = 0, πhard π₯π = 1.Output: Repeatedly call an ππ oracle to find the lexicographically first witness πsuch that πππππ ππ = π, and output the π-th bit of π.
For a.e. π, thereβs an π β π, π π such that ππππππ π = π and πππππ π = π
Now the advice is insanely long! We canβt just remove it, as before!(And of course thereβs a function in π¬π΅π·/ππ/π without small ACC circuitsβ¦)
But now, the construction of such inputs ππ becomes an important problem!
If we could construct these βbadβ ππ in π¬π΅π· (given input ππ) weβd be done!
Advice has length π = ππ/π (!!)
Rough Idea: Using a variation on the proof of this time hierarchy, πΉ does a βbinary searchβ with its NP oracle,
making π(π) calls with queries of length about π(π π ), to find a bad input π₯π.
Theorem: There is a πππππ π π» π ππ algorithm πΉ (a refuter)
such that for every NTIMEGUESS π π π , π π algorithm π,
πΉ(1π, π) outputs an π-bit π₯π such that πβπππ π₯π β π π₯π , for every sufficiently large π.
Theorem: [Fortnow-Santhanam 2016]
Thereβs an πhard β NTIME π π β io-NTIMEGUESS π π π , π π
One More (New!) Ingredient
One More Tryβ¦
The πππ algorithm computing an almost-everywhere hard function:
For each such π₯, every valid witness for πhard π₯ is a hard function:it cannot be computed by small ππππ circuits!
Input: π-bit index π β {π, π}π, recall π = π log π
Algorithm: Set π β 2π/π and run refuter πΉ(ππ, ππππππ) in πππ, obtaining (for all
but finitely many π) an π-bit string ππ such that πcheat ππ β πhard ππ . Repeatedly call an ππ oracle to find the lexicographically first witness π such that
πππππ ππ = π, and output the π-th bit of π.
For a.e. π, thereβs an π β π, π π such that ππππππ π = π and πππππ π = π
Conclusion: πππ β π’π¨-ππππ
ConclusionWe have managed to prove several almost-everywhere lower
bounds for functions in π¬π΅π·, even for the average case.
What about NEXP? Or Quasi-NP? Or NP?
Can we prove ππππ β π’π¨-ππππ ?
What other lower bounds can be made a.e.?(e.g. πΊππ· β πΊπ°ππ¬ ππ )
Thanks for watching!