Barriers in Quantum Computing And How to Smash

Barriers in Quantum Computing (And How to Smash Them Through Closer Interactions Between Classical and Quantum CS) Day Classical complexity theorists, remain in your seats. Do not leave the room. There is no need to panic. 9: 20 -10: 20 Scott Aaronson, The Computational Complexity of Linear Optics 10: 20 -11: 00 Break 11: 00 -12: 00 Umesh Vazirani, Challenges of Hamiltonian Complexity 12: 00 -2: 00 Lunch 2: 00 -2: 55 Ronald de Wolf, Quantum Proofs for Classical Theorems 3: 00 -3: 15 Break 3: 15 -4: 10 Chris Peikert, Lattices and Quantum Computing: Barriers and Opportunities 4: 15 -5: 15 Rump Session (Ben Reichardt, Yi-Kai Liu, ADD YOUR NAME HERE) 6: 00 -8: 30 Workshop Dinner at the Palmer House

The Computational Complexity of Linear Optics vs Scott Aaronson and Alex Arkhipov MIT

The Extended Church. Turing Thesis (ECT) Everything feasibly computable in the physical world is feasibly computable by a (probabilistic) Turing machine Shor’s Theorem: QUANTUM SIMULATION has no efficient classical algorithm, unless FACTORING does

So then what more barriers are there to smash in quantum computing theory? Building a QC able to factor large numbers is damn hard! After 16 years, no fundamental obstacle has been found (or even seriously proposed), but who knows? Can’t we “meet the physicists halfway, ” and show computational hardness for quantum systems closer to what actually exists now? FACTORING might be in BPP! At any rate, it’s an extremely “special” problem Wouldn’t it be great to show that if BPP=BQP, then (say) the polynomial hierarchy collapses?

Today: “A New Attack on the ECT” We define a model of computation based on linear optics: n identical photons traveling through a network of poly(n) beamsplitters, phase-shifters, etc. , then a measurement of where the photons ended up Crucial point: No entangling interactions between pairs of photons needed! Our model is contained in BQP, but seems unlikely to be BQP-complete. We don’t know if it solves any decision problems that are hard classically. But for sampling and search problems, the situation is completely different…

Theorem 1. Suppose that for every linear-optics network, the probability distribution over measurement outcomes can be sampled in classical polynomial time. Then P#P=BPPNP (so PH collapses) More generally, let O be any oracle that simulates a “OK, but isn’t the real description of linear-optics network A, given aquestion the A and a hardness of approximate sampling? random string r. Then After all, experiments are noisy, and not even the linear-optics network itself can be exactly!” So even if linear optics samplesimulated in BPPPH, that still collapses PH! (New evidence that QCs have capabilities beyond PH, complementing [A’ 10], [FU’ 10])

Theorem 2. Suppose two plausible conjectures are true: the permanent of a Gaussian random matrix is (1) #P-hard to approximate, and (2) not too concentrated around 0. Let O be any oracle takes as input a description of a linear-optics network A, a random string r, and 01/ , But to fully smash the barriers, and that samples from a distribution -close to A’s in we need some variation distance. Then classical help proving our conjectures… In other words: if our conjectures hold, then even simulating noisy linear-optics experiments is classically intractable, unless PH collapses

From Sampling to Search Problems Let FBPP and FBQP be the classes of search problems (like finding a Nash equilibrium) that are solvable by classical and quantum computers respectively, in poly(n, 1/ ) time and with success probability 1 - Then assuming our conjectures, even FBPP=FBQP would imply P#P=BPPNP Theorem (A. , ECCC 10 -128): FBPP=FBQP if and only if Samp. P=Samp. BQP. Proof Sketch: Samp. P=Samp. BQP FBPP=FBQP is immediate. For the other direction, use Kolmogorov complexity!

Particle Physics In One Slide There are two basic types of particle in the universe… All I can say is, the bosons got the harder job Indeed, [Valiant 2002, Terhal-Di. Vincenzo 2002] BOSONS FERMIONS showed that noninteracting fermion systems can be simulated in BPP. But, confirming Avi’s joke, we’ll argue that the analogous problem for bosons Their transition amplitudes aremuch harder… (such as photons) is given respectively by…

Linear Optics for Dummies We’ll be considering a special kind of quantum computer, which is not based on qubits The basis states have the form |S =|s 1, …, sm , where si is the number of photons in the ith “mode” We’ll never create or destroy photons. So if there are n photons, then s 1, …, sm are nonnegative integers summing to n For us, m=poly(n) Initial state: |I =|1, …, 1, 0, ……, 0

You get to apply any m m unitary matrix U If n=1 (i. e. , there’s only one photon, in a superposition over the m modes), U acts on that photon in the obvious way In general, there are ways to distribute n identical photons into m modes U induces an M M unitary (U) on the n-photon states as follows: Here US, T is an n n submatrix of U (possibly with repeated rows and columns), obtained by taking si copies of the ith row of U and tj copies of the jth column for all i, j

Example: The “Hong-Ou-Mandel Dip” Suppose U Then Pr[the two photons land in different modes] is Pr[they both land in the first mode] is

Beautiful Alternate Perspective The “state” of our computer, at any time, is a degree-n polynomial over the variables x=(x 1, …, xm) (n<<m) Initial state: p(x) : = x 1 xn We can apply any m m unitary transformation U to x, to obtain a new degree-n polynomial Then on “measuring, ” we see the monomial with probability

OK, so why is it hard to sample the distribution over photon numbers classically? Given any matrix A Cn n, we can construct an m m unitary U (where m 2 n) as follows: Suppose we start with |I =|1, …, 1, 0, …, 0 (one photon in each of the first n modes), apply U, and measure. Then the probability of observing |I again is

Claim 1: p is #P-complete to estimate (up to a constant factor) Claim 2: Suppose we had a fast classical algorithm for linear-optics sampling. Then Idea: Valiant proved that the we could estimate p in BPPNP PERMANENT is #P-complete. Idea: Let M be our classical Can use a classical reduction sampling algorithm, and let r to go from a multiplicative be its randomness. Use 2 approximation of |Per(A)| approximate counting to to Per(A) itself. estimate Conclusion: Suppose we had a fast classical algorithm for linear-optics sampling. Then P#P=BPPNP.

As I said before, I find this result unsatisfying, since it only talks about the classical hardness of exactly sampling the distribution over photon numbers What about sampling a distribution that’s 1/poly(n)close in variation distance? Difficulty: The sampler might adversarially refuse to output the one submatrix whose permanent we care about! That changes the output distribution by only exp(-n), so we still have an excellent sampler … but we can no longer use it to estimate |Per(A)|2 in BPPNP To get around this difficulty, it seems we need to “smuggle in” the matrix A that we about as a random submatrix of U

Main Result Consider applying a Haar-random m m unitary matrix U, to n photons in m=poly(n) modes: U Distribution DU over photon numbers Main technical lemma used in proof: Suppose there’s a classical algorithm to sample a Let m n 6. Then an n n submatrix of an m m distribution -close to DU in poly(n, 1/ ) time. Then for Haar unitary matrix is Õ(1/n)-close in variation all , 1/poly(n), there’s also a BPPNP algorithm to distance to a matrix of independent Gaussians. estimate |Per(X)|2 to within additive error n!, with probability 1 - over a Gaussian random matrix

So the question boils down to this: how hard is it to additively estimate |Per(X)|2, with high probability over a Gaussian random matrix We conjecture that it’s #P-hard—in which case, even approximate classical simulation of our linear-optics experiment would imply P#P=BPPNP We can decompose this conjecture into two plausible sub-conjectures: that multiplicatively estimating Per(X) is #P-hard for Gaussian X, and that Per(X) is “not too concentrated around 0”

The Permanent-of-Gaussians Conjecture (PGC) The following problem is #P-hard. Given a matrix X Cn n of i. i. d. Gaussian entries, together with 01/ and 01/ , output an approximation z such that We can prove #P-hardness if =0 or =0. So what makes the PGC nontrivial is really the combination of average-case with approximation…

The Permanent Anti-Concentration Conjecture (PACC) There exist constants C, D and >0 such that for all n and >0, Empirically true! Also, we can prove it with determinant in place of permanent

Experimental Prospects It seems well within current technology to do our experiment with (say) n=4 photons and m=20 modes (Current record: n=2 photons) Physicists we consulted: “Sounds hard! But If you can scale to n photons and error in variation not as hard as building a universal QC” distance, using “poly(n, 1/ ) experimental effort, ” then modulo our complexity conjectures, the ECT is false What Remark: take to scale to (say) n=20 photons and would it No point in scaling this experiment m=500 modes? much beyond 20 or 30 photons, since then a classical computer can’t even verify the answers! - Reliable single-photon sources (standard laser isn’t good enough!) - Reliable photodetector arrays - Stable apparatus to ensure that w. h. p. , all n photons arrive at photodetector arrays at exactly the same time

Open Problems Prove the PGC ($200) and PACC ($100)! Can our linear-optics model solve classically-intractable decision problems? What about problems for which a classical computer can verify the answers? Do BPP=BQP or Promise. BPP=Promise. BQP have interesting structural complexity consequences? Similar hardness results for other natural quantum systems (besides linear optics)? Bremner, Jozsa, Shepherd 2010: Another system for which exact classical simulation would collapse PH