Quantum advantage with shallow circuits. arxiv: Sergey Bravyi (IBM) David Gosset (IBM) Robert Koenig (Munich)

Transcription

1 Quantum advantage with shallow circuits arxiv: ergey Bravyi (IBM) David Gosset (IBM) Robert Koenig (Munich)

2 In this talk I will describe a provable, non-oracular, quantum speedup which is attained by constant-depth quantum circuits in a 2D architecture.

3 I. Overview

4 A depth-d quantum circuit consists of d time steps. Each time step contains one- and two-qubit gates acting on disjoint qubits. H

5 A depth-d quantum circuit consists of d time steps. Each time step contains one- and two-qubit gates acting on disjoint qubits. H Time step 1

6 A depth-d quantum circuit consists of d time steps. Each time step contains one- and two-qubit gates acting on disjoint qubits. H Time step 2

7 A depth-d quantum circuit consists of d time steps. Each time step contains one- and two-qubit gates acting on disjoint qubits. H Time step 3

8 A depth-d quantum circuit consists of d time steps. Each time step contains one- and two-qubit gates acting on disjoint qubits. H

9 A depth-d quantum circuit consists of d time steps. Each time step contains one- and two-qubit gates acting on disjoint qubits. H We are interested in constant-depth quantum circuits, for which d = O(1).

10 H Constant-depth quantum circuits

11 Constant-time quantum computation How much does parallelism buy us if we only have a fixed computation time? H Constant-depth quantum circuits

12 Constant-time quantum computation How much does parallelism buy us if we only have a fixed computation time? H Quantum algorithms for small quantum computers Constant-depth quantum circuits

13 Constant-time quantum computation How much does parallelism buy us if we only have a fixed computation time? H Quantum algorithms for small quantum computers tructure/imulation Cannot prepare codewords of good quantum codes [Eldar, Harrow 2015 ] Constant-time quantum computation Efficient classical simulation of depth-2 circuits [Terhal, Divincenzo 2002] Constant-depth quantum circuits General simulation algorithms (superpolynomial) [Aaronson, Chen 2016]

14 Constant-time quantum computation How much does parallelism buy us if we only have a fixed computation time? H Quantum algorithms for small quantum computers tructure/imulation Cannot prepare codewords of good quantum codes [Eldar, Harrow 2015 ] Constant-time quantum computation Efficient classical simulation of depth-2 circuits [Terhal, Divincenzo 2002] General simulation algorithms (superpolynomial) [Aaronson, Chen 2016] Constant-depth quantum circuits Beat poly-time classical computation? Constant-depth unlikely to be classically simulable. [Terhal, Divincenzo 02] Beat the best classical computers for some task? [Gao et al. 17] [Bermejo-Vega et al. 17] uses IQP results [Bremner, Montanaro, hepherd 16]

15 H OR AND AND OR This talk: Are constant-depth quantum circuits more powerful than constantdepth classical circuits?

16 Classical circuits A classical gate computes a boolean function f: {0,1} k {0,1} x 1 x 2 x k f f(x) f(x)

17 Classical circuits A classical gate computes a boolean function f: {0,1} k {0,1} Number of input bits k is called the fan-in x 1 x 2 f f(x) f(x) x k

18 Classical circuits A classical gate computes a boolean function f: {0,1} k {0,1} Number of input bits k is called the fan-in x 1 x 2 f f(x) f(x) Number of times the output is used is called the fan-out x k

19 Classical circuits A classical gate computes a boolean function f: {0,1} k {0,1} Number of input bits k is called the fan-in x 1 x 2 f f(x) f(x) Number of times the output is used is called the fan-out x k We consider circuits composed of bounded fan-in gates, i.e., k = O(1). We do not restrict the fan-out.

20 Classical circuits A classical gate computes a boolean function f: {0,1} k {0,1} Number of input bits k is called the fan-in x 1 x 2 f f(x) f(x) Number of times the output is used is called the fan-out x k We consider circuits composed of bounded fan-in gates, i.e., k = O(1). We do not restrict the fan-out.

21 Constant-depth classical circuits A depth-d classical circuit consists of d layers (time steps) of gates. OR AND AND OR We consider constant-depth circuits composed of bounded fan-in gates. This class of circuits is known as NC 0. We also allow the circuit to be probabilistic (random input bits are provided).

22 H OR AND AND OR Our result: We describe a computational problem that is solved with certainty by a constantdepth quantum circuit. We prove that any classical circuit which solves the problem with high probability must have depth increasing logarithmically with input size.

23 H Quantum circuit depth d = O(1) Classical OR needs AND depth d clog n AND OR Our result: We describe a computational problem that is solved with certainty by a constantdepth quantum circuit. We prove that any classical circuit which solves the problem with high probability must have depth increasing logarithmically with input size.

24 H Quantum circuit depth d = O(1) Classical OR needs AND depth d clog n AND OR Our result: We describe a computational problem that is solved with certainty by a constantdepth quantum circuit (also: all gates act locally on a 2D grid). We prove that any classical circuit which solves the problem with high probability must have depth increasing logarithmically with input size.

25 H Quantum circuit depth d = O(1) Classical OR needs AND depth d clog n AND OR Our result: We describe a computational problem that is solved with certainty by a constantdepth quantum circuit (also: all gates act locally on a 2D grid). We prove that any classical circuit which solves the problem with high probability must have depth increasing logarithmically with input size. The quantum speedup is unconditional (does not rely on complexity-theoretic conjectures and is non-oracular)

26 II. The 2D Hidden Linear Function Problem

27 Quadratic form on a grid Let G = (V, E) be an N N grid graph. Write n = N 2 = V

28 Quadratic form on a grid Let G = (V, E) be an N N grid graph. Write n = N 2 = V A e b v Choose coefficients A e {0,1} for each edge and b v {0,1} for each vertex.

29 Quadratic form on a grid Let G = (V, E) be an N N grid graph. Write n = N 2 = V A e b v Choose coefficients A e {0,1} for each edge and b v {0,1} for each vertex. Any choice of coefficients defines a quadratic form q: {0,1} n Z 4 q x = e= v,w E 2A e x v x w b v x v v V

30 The quadratic form hides a linear function Define a set L q = {x F 2 n : q x y = q x + q y for all y F 2 n }

31 The quadratic form hides a linear function Define a set L q = {x F 2 n : q x y = q x + q y for all y F 2 n } Lemma The set L q is a linear subspace of F 2 n. Furthermore, there is a secret bit string z {0,1} n such that q x = 2z T x x L q

32 The quadratic form hides a linear function Define a set L q = {x F 2 n : q x y = q x + q y for all y F 2 n } Lemma The set L q is a linear subspace of F 2 n. Furthermore, there is a secret bit string z {0,1} n such that q x = 2z T x x L q Define a computational problem where the goal is to find a secret bit string

33 The 2D Hidden Linear Function Problem Input: Coefficients A {0,1} E and b {0,1} V. pecifies a quadratic form q(x) and a subspace L q F 2 n

34 The 2D Hidden Linear Function Problem Input: Coefficients A {0,1} E and b {0,1} V. pecifies a quadratic form q(x) and a subspace L q F 2 n Output: A secret bit string z {0,1} n such that q x = 2z T x x L q

35 The 2D Hidden Linear Function Problem Input: Coefficients A {0,1} E and b {0,1} V. pecifies a quadratic form q(x) and a subspace L q F 2 n Output: A secret bit string z {0,1} n such that q x = 2z T x x L q Can be viewed as a non-oracular version of the Bernstein-Vazirani problem. [Bernstein Vazirani 1993] In general each instance of the 2D HLF has many valid solutions z.

36 Quantum algorithm b A b A A qubit for each vertex of the N N grid 0 n H n CZ(A) (b) H n Output z {0,1} n

37 Quantum algorithm b A b A A qubit for each vertex of the N N grid 0 n H n CZ(A) (b) H n Output z {0,1} n Apply CZ v,w for edges e = v, w with A e = 1

38 Quantum algorithm b A b A A qubit for each vertex of the N N grid 0 n H n CZ(A) (b) H n Output z {0,1} n Apply CZ v,w for edges e = v, w with A e = 1 Apply phase gate v to qubits with b v = 1

39 Quantum algorithm b A b A A qubit for each vertex of the N N grid 0 n H n CZ(A) (b) H n Output z {0,1} n Apply CZ v,w for edges e = v, w with A e = 1 Apply phase gate v to qubits with b v = 1 = i

40 Quantum algorithm b A b A A qubit for each vertex of the N N grid 0 n H n CZ(A) (b) H n Output z {0,1} n Apply CZ v,w for edges e = v, w with A e = 1 Apply phase gate v to qubits with b v = 1 = i Fact: The output z is a uniformly random solution to the 2D HLF Problem.

41 The algorithm can be implemented in constant-depth b A b A 0 n H n CZ(A) (b) H n 1 time-step 1 time-step 1 time-step Four layers of CCZ gates. (even/odd vertical/horizontal edges) Decompose CCZ gates into 1- and 2-qubit gates.

42 it only requires classically controlled Clifford gates between nearest neighbor qubits on a 2D grid. Example: Place a qubit at each vertex Place input bits on vertices and edges: : Edge with A e = 1 : Vertex with b v = 1

43 it only requires classically controlled Clifford gates between nearest neighbor qubits on a 2D grid. Example: Place a qubit at each vertex Place input bits on vertices and edges: : Edge with A e = 1 : Vertex with b v = 1

44 it only requires classically controlled Clifford gates between nearest neighbor qubits on a 2D grid. Example: H H H H H H H H H H : Edge with A e = 1 H H H H H : Vertex with b v = 1 H H H H H H H H H H

45 it only requires classically controlled Clifford gates between nearest neighbor qubits on a 2D grid. Example: CZ : Edge with A e = 1 CZ : Vertex with b v = 1 CZ

46 it only requires classically controlled Clifford gates between nearest neighbor qubits on a 2D grid. Example: CZ : Edge with A e = 1 : Vertex with b v = 1

47 it only requires classically controlled Clifford gates between nearest neighbor qubits on a 2D grid. Example: CZ CZ : Edge with A e = 1 : Vertex with b v = 1

48 it only requires classically controlled Clifford gates between nearest neighbor qubits on a 2D grid. Example: CZ : Edge with A e = 1 : Vertex with b v = 1 CZ

49 it only requires classically controlled Clifford gates between nearest neighbor qubits on a 2D grid. Example: : Edge with A e = 1 : Vertex with b v = 1

50 it only requires classically controlled Clifford gates between nearest neighbor qubits on a 2D grid. Example: H H H H H H H H H H : Edge with A e = 1 H H H H H : Vertex with b v = 1 H H H H H H H H H H

51 it only requires classically controlled Clifford gates between nearest neighbor qubits on a 2D grid. Example: : Edge with A e = 1 : Vertex with b v = 1

52 The 2D HLF problem is solved by a constant-depth quantum circuit with gates acting locally in 2D. Next we show that it cannot be solved by a constant-depth classical circuit

53 III. Classical lower bound

54 Theorem: The following holds for all sufficiently large N. Let C N be a classical probabilistic circuit composed of gates of fan-in K which solves size-n instances of the 2D HLF Problem with probability greater than 7/8. Then depth C N log(n) 8log(K)

55 Input (instance on N N grid) Random bits (drawn from any joint distribution) A {0,1} E b {0,1} V r {0,1} l C N Circuit must have depth Ω(log N ) Output z {0,1} V olution with probability > 7/8

56 Proof Ideas Locality in shallow classical circuits Each output bit can only depend on O(1) input bits. Vs. Quantum nonlocality Measurement statistics of entangled quantum states cannot be reproduced by local hidden variable models

57 Locality in classical circuits x 1 x 2 OR AND z 1 input x 3 x 4 AND OR z 2 output x 5 x 6 AND The lightcone L z k connected to z k. of an output bit z k is the set of input bits x i that are causually

58 Locality in classical circuits x 1 x 2 OR AND z 1 input x 3 x 4 AND OR z 2 output x 5 x 6 AND The lightcone L z k connected to z k. of an output bit z k is the set of input bits x i that are causually

59 Locality in classical circuits x 1 x 2 OR AND z 1 input x 3 x 4 AND OR z 2 output x 5 x 6 AND The lightcone L z k connected to z k. of an output bit z k is the set of input bits x i that are causually

60 Locality in classical circuits x (Includes random bits) C depth d, fan-in K z = F(x) Constant-depth locality : Lightcones of output bits have constant size L z k K d We ll see that the 2D Hidden Linear Function problem cannot be solved by constantdepth local circuits. First consider simpler forms of locality

61 Quantum nonlocality beats completely local circuits [Greenburger et al. 1990][Mermin 1990] GHZ = 1 P GHZ = GHZ satisfies: P {XXX, XYY, YXY, YYX}

62 Quantum nonlocality beats completely local circuits [Greenburger et al. 1990][Mermin 1990] GHZ = 1 P GHZ = GHZ satisfies: P {XXX, XYY, YXY, YYX} Choose bits b 1, b 2, b 3 and then measure each qubit of GHZ in either the X basis (if b j = 0) or the Y basis (if b j = 1). Outcomes z j 1, +1 satisfy: i b 1+b 2 +b 3 z 1 z 2 z 3 = 1 whenever b 1 b 2 b 3 = 0 GHZ relation

63 Quantum nonlocality beats completely local circuits [Greenburger et al. 1990][Mermin 1990] GHZ = 1 P GHZ = GHZ satisfies: P {XXX, XYY, YXY, YYX} Choose bits b 1, b 2, b 3 and then measure each qubit of GHZ in either the X basis (if b j = 0) or the Y basis (if b j = 1). Outcomes z j 1, +1 satisfy: i b 1+b 2 +b 3 z 1 z 2 z 3 = 1 whenever b 1 b 2 b 3 = 0 GHZ relation The GHZ relation cannot be satisfied by a completely local classical probabilistic circuit where each output bit z j is correlated with at most one of the input bits b k.

64 Quantum nonlocality beats geometrically local circuits [Barrett et al. 2007] Graph state on an M-cycle (M even): Φ M = M j=1 CZ j,j+1 H M 0 M

65 Quantum nonlocality beats geometrically local circuits [Barrett et al. 2007] Graph state on an M-cycle (M even): Φ M = M j=1 Choose 3 qubits u, v, w on the even sublattice. Measure u, v, w in X or Y basis and all other qubits in X basis. CZ j,j+1 H M 0 M

66 Quantum nonlocality beats geometrically local circuits [Barrett et al. 2007] Graph state on an M-cycle (M even): Φ M = M j=1 Choose 3 qubits u, v, w on the even sublattice. Measure u, v, w in X or Y basis and all other qubits in X basis. CZ j,j+1 H M 0 M u X/Y X X X X X/Y X v X X X X X/Y w

67 Quantum nonlocality beats geometrically local circuits [Barrett et al. 2007] Graph state on an M-cycle (M even): Φ M = M j=1 Choose 3 qubits u, v, w on the even sublattice. Measure u, v, w in X or Y basis and all other qubits in X basis. Input b u, b v, b w {0,1} Measurement bases Output z {0,1} M Measurement outcomes CZ j,j+1 H M 0 M u X/Y X X X X X/Y X v X X X X X/Y w

68 Quantum nonlocality beats geometrically local circuits [Barrett et al. 2007] Graph state on an M-cycle (M even): Φ M = M j=1 Choose 3 qubits u, v, w on the even sublattice. Measure u, v, w in X or Y basis and all other qubits in X basis. Input b u, b v, b w {0,1} Measurement bases Output z {0,1} M Measurement outcomes CZ j,j+1 H M 0 M u X/Y X X X X X/Y X v X X X Fact: Input/output satisfy a cycle relation R b u, b v, b w, z = 1 similar to the GHZ relation. X X/Y w

69 Quantum nonlocality beats geometrically local circuits [Barrett et al. 2007] Graph state on an M-cycle (M even): Φ M = M j=1 Choose 3 qubits u, v, w on the even sublattice. Measure u, v, w in X or Y basis and all other qubits in X basis. Input b u, b v, b w {0,1} Measurement bases Output z {0,1} M Measurement outcomes CZ j,j+1 H M 0 M u X/Y X X X X X/Y X v X X X Fact: Input/output satisfy a cycle relation R b u, b v, b w, z = 1 similar to the GHZ relation. X X/Y w Lemma: uppose a classical circuit satisfies the cycle relation with probability > 7/8. Then some output bit z k is correlated with a distant input bit b u, b v or b w. (this means it is not the nearest vertex of the triangle)

70 How is this related to the 2D Hidden Linear Function Problem?

71 How is this related to the 2D Hidden Linear Function Problem? b b A A 0 n H n CZ(A) (b) H n z {0,1} n

72 How is this related to the 2D Hidden Linear Function Problem? b b A A 0 n H n CZ(A) (b) H n z {0,1} n Prepare graph state for graph with adjacency matrix A

73 How is this related to the 2D Hidden Linear Function Problem? b b A A 0 n H n CZ(A) (b) H n z {0,1} n Prepare graph state for graph with adjacency matrix A Measure jth qubit in X or Y basis depending on b j

74 How is this related to the 2D Hidden Linear Function Problem? b b A A 0 n H n CZ(A) (b) H n z {0,1} n Prepare graph state for graph with adjacency matrix A Measure jth qubit in X or Y basis depending on b j Choosing A to describe the adjacency matrix of a cycle and choosing b appropriately we infer (from Barrett et al.) a cycle relation satisfied by input/output.

75 How is this related to the 2D Hidden Linear Function Problem? b b A A 0 n H n CZ(A) (b) H n z {0,1} n Prepare graph state for graph with adjacency matrix A Measure jth qubit in X or Y basis depending on b j Choosing A to describe the adjacency matrix of a cycle and choosing b appropriately we infer (from Barrett et al.) a cycle relation satisfied by input/output. A classical circuit which solves the 2D HLF problem must also satisfy all such cycle relations.

76 Quantum nonlocality beats constant-depth local circuits We use constant-depth locality (every output bit has constant-sized lightcone) and a probabilistic argument to prove the following:

77 Quantum nonlocality beats constant-depth local circuits We use constant-depth locality (every output bit has constant-sized lightcone) and a probabilistic argument to prove the following: Lemma: uppose a classical circuit has depth less than log(n) 8log(K). Then we can find 3 vertices u, v, w on the even sublattice of the N N grid and a cycle Γ which passes through them, such that input bits b u, b v, b w are not correlated with any distant output bits on Γ. v u Γ w

78 Quantum nonlocality beats constant-depth local circuits We use constant-depth locality (every output bit has constant-sized lightcone) and a probabilistic argument to prove the following: Lemma: uppose a classical circuit has depth less than log(n) 8log(K). Then we can find 3 vertices u, v, w on the even sublattice of the N N grid and a cycle Γ which passes through them, such that input bits b u, b v, b w are not correlated with any distant output bits on Γ. v u Γ w The circuit does not w.h.p satisfy the cycle relation for Γ

79 Quantum nonlocality beats constant-depth local circuits We use constant-depth locality (every output bit has constant-sized lightcone) and a probabilistic argument to prove the following: Lemma: uppose a classical circuit has depth less than log(n) 8log(K). Then we can find 3 vertices u, v, w on the even sublattice of the N N grid and a cycle Γ which passes through them, such that input bits b u, b v, b w are not correlated with any distant output bits on Γ. v u Γ w The circuit does not w.h.p satisfy the cycle relation for Γ It does not w.h.p solve instances of 2D HLF problem where A is the adjacency matrix of Γ.

80 Quantum nonlocality beats constant-depth local circuits We use constant-depth locality (every output bit has constant-sized lightcone) and a probabilistic argument to prove the following: Lemma: uppose a classical circuit has depth less than log(n) 8log(K). Then we can find 3 vertices u, v, w on the even sublattice of the N N grid and a cycle Γ which passes through them, such that input bits b u, b v, b w are not correlated with any distant output bits on Γ. v u Γ w The circuit does not w.h.p satisfy the cycle relation for Γ It does not w.h.p solve instances of 2D HLF problem where A is the adjacency matrix of Γ. This provides our lower bound on the depth of any classical circuit which solves the 2D HLF problem with probability greater than 7/8.

81 Open problems Recursive HLF problems? The recursive version of Bernstein-Vazirani gives a superpolynomial speedup in query complexity. Noisy constant-depth quantum circuits vs noiseless constant-depth classical circuits? ampling problems? Can constant-depth quantum circuits sample from a distribution that can t be sampled by classical constant depth circuits? A recent characterization of distributions sampled by NC 0 circuits might be useful [Viola 2014]. Polynomial speed-up? Constant-depth quantum algorithm solves the 2D HLF Problem in linear time. Best known classical algorithm takes time O(n 2 ).