1. The Complexity of Quantum Entanglement Fernando G.S.L. Brandão ETH Zürich Based on joint work with M. Christandl and J. Yard Journees Deferation de Reserche en Mathematiques de Paris Centre/GT Informatique Quantique Paris, 09/05/2012

2. Problem 1: For M in H(C d ) (d x d matrix) compute Very Easy! Problem 2: For M in H(C d C l ), compute Quadratic vs Biquadratic Optimization This talk: Best known algorithm (and best hardness result) using ideas from Quantum Information Theory

3. Problem 1: For M in H(C d ) (d x d matrix) compute Very Easy! Problem 2: For M in H(C d C l ), compute Quadratic vs Biquadratic Optimization This talk: Best known algorithm (and best hardness result) using ideas from Quantum Information Theory

4. Problem 1: For M in H(C d ) (d x d matrix) compute Very Easy! Problem 2: For M in H(C d C l ), compute Quadratic vs Biquadratic Optimization This talk: Best known algorithm (and best hardness result) using ideas from Quantum Information Theory

5. Outline The Problem Quantum States Quantum Entanglement The Algorithm Parrilo-Lasserre Relaxation Monogamy of Entanglement Quantum de Finetti Theorem Applications A new characterization of Quantum NP Small Set Expansion Proof Ideas

6. Quantum States Pure States: norm-one vector in C d : Mixed States: positive semidefinite matrix of unit trace: Dirac notation reminder:

7. Quantum Measurements To any experiment with d outcomes we associate d positive matrices {M k } such that and calculate probabilities as E.g. For pure states,

8. Quantum Entanglement Pure States: If, it’s separable otherwise, it’s entangled. Mixed States: If it’s separable otherwise, it’s entangled.

9. Quantum Entanglement Pure States: If, it’s separable otherwise, it’s entangled. Mixed States: If, it’s separable otherwise, it’s entangled.

10. A Physical Definition of Entanglement LOCC: Local quantum Operations and Classical Communication Separable states can be created by LOCC: Entangled states cannot be created by LOCC: non-classical correlations

11. The Separability Problem Given is it separable or entangled? (Weak Membership: W SEP (ε, ||*||) Given ρ AB determine if it is separable, or ε-way from SEP SEP D

12. The Problem (for experimentalists)

15. Relevance Quantum Cryptography Security only if state is entangled Quantum Communication Advantage over classical (e.g. teleportation, dense coding) only if state is entangled Computational Physics Entanglement responsible for difficulty of simulation of quantum systems

16. The problem of deciding whether a state is entangled has been considered since the early days of the field of quantum information theory is regarded as a computationally difficult problem In this talk I’ll discuss the fastest known algorithm for this problem Deciding Entanglement

17. The Separability Problem (again) Given is it separable or entangled? (Weak Membership: W SEP (ε, ||*||) Given ρ AB determine if it is separable, or ε-way from SEP SEP D

18. The Separability Problem (again) Given is it separable or entangled? (Weak Membership: W SEP (ε, ||*||) Given ρ AB determine if it is separable, or ε-way from SEP SEP D Which norm should we use?

19. Norms on Quantum States How to quantify the distance in Weak-Membership? Euclidean Norm (Hilber-Schmidt): ||X|| 2 = tr(X T X) 1/2 Trace Norm ||X|| 1 = tr((X T X) 1/2 ) Obs: ||X|| 1 ≥||X|| 2 ≥d -1/2 ||X|| 1

20. The LOCC Norm Operational interpretation trace norm: ||ρ – σ|| 1 = 2 max 0<M<I tr(M(ρ – σ)) optimal bias of distinguishing the two states by quantum measurements For ρ AB, σ AB define ||ρ – σ|| LOCC = 2 max 0<M<I tr(M(ρ – σ)) : {M, I - M} in LOCC LOCC: Local quantum Operations and Classical Communication

21. The LOCC Norm Operational interpretation trace norm: ||ρ – σ|| 1 = 2 max 0<M<I tr(M(ρ – σ)) optimal bias of distinguishing the two states by quantum measurements For ρ AB, σ AB define the LOCC norm ||ρ – σ|| LOCC = 2 max 0<M<I tr(M(ρ – σ)) : {M, I - M} in LOCC Optimal bias of distinguishing two states by LOCC measurements E.g. (one-way LOCC)

22. Optimization Over Separable States (Best Separable State BSS(ε)) Given estimate to additive error ε

23. Previous Work When is ρ AB entangled? - Decide if ρ AB is separable or ε-away from separable Beautiful theory behind it (PPT, entanglement witnesses, etc) Horribly expensive algorithms State-of-the-art: 2 O(|A|log|B|log (1/ε)) time complexity for either ||*|| 2 or ||*|| 1 norms (Doherty, Parrilo, Spedalieri ‘04)

24. Hardness Results When is ρ AB entangled? - Decide if ρ AB is separable or ε-away from separable (Gurvits ‘02) NP-hard with ε=1/exp( |A||B| ) (Gharibian ‘08, Beigi ‘08) NP-hard with ε=1/poly( (|A||B|) 1/2 ) (Beigi&Shor ‘08) Favorite separability tests fail (Harrow&Montanaro ‘10) No exp(O(|A| 1-ν |A| 1-μ )) time algorithm for membership in any convex set within ε=Ω(1) trace distance to SEP and any ν+μ>0, unless ETH fails ETH (Exponential Time Hypothesis): SAT cannot be solved in 2 o(n) time (Impagliazzo&Paruti ’99)

25. Hardness Results When is ρ AB entangled? - Decide if ρ AB is separable or ε-away from separable (Gurvits ‘02) NP-hard with ε=1/exp( |A||B| ) (Gharibian ‘08, Beigi ‘08) NP-hard with ε=1/poly( |A||B| ) (Beigi&Shor ‘08) Favorite separability tests fail (Harrow&Montanaro ‘10) No exp(O(|A| 1-ν |A| 1-μ )) time algorithm for membership in any convex set within ε=Ω(1) trace distance to SEP and any ν+μ>0, unless ETH fails ETH (Exponential Time Hypothesis): SAT cannot be solved in 2 o(n) time (Impagliazzo&Paruti ’99)

26. Hardness Results When is ρ AB entangled? - Decide if ρ AB is separable or ε-away from separable (Gurvits ‘02) NP-hard with ε=1/exp( |A||B| ) (Gharibian ‘08, Beigi ‘08) NP-hard with ε=1/poly( |A||B| ) (Beigi, Shor ‘08) Favorite separability tests fail (Harrow&Montanaro ‘10) No exp(O(|A| 1-ν |A| 1-μ )) time algorithm for membership in any convex set within ε=Ω(1) trace distance to SEP and any ν+μ>0, unless ETH fails ETH (Exponential Time Hypothesis): SAT cannot be solved in 2 o(n) time (Impagliazzo&Paruti ’99)

27. Hardness Results When is ρ AB entangled? - Decide if ρ AB is separable or ε-away from separable (Gurvits ‘02) NP-hard with ε=1/exp( |A||B| ) (Gharibian ‘08, Beigi ‘08) NP-hard with ε=1/poly( |A||B| ) (Beigi, Shor ‘08) Favorite separability tests fail (Harrow, Montanaro ‘10) No exp(O(log 1-ν |A|log 1-μ |B|)) time algorithm for membership in any convex set within ε=Ω(1) trace distance to SEP, and any ν+μ>0, unless ETH fails ETH (Exponential Time Hypothesis): SAT cannot be solved in 2 o(n) time (Impagliazzo&Paruti ’99)

28. Algorithms for BSS Estimate with additive error ε State-of-the-art: 2 O((|A|+|B|)log (1/ε)) time complexity Exhaustive search over ε-nets on A and B!

29. Hardness Results for BSS (Gurvits ‘02, Gharibian ‘08, Beigi ‘08) NP-hard with ε=1/poly( |A||B| ) (Harrow, Montanaro ’10, built on Aaronson et al ‘08) No exp(O(log 1- ν |A|log 1-μ |B|||M|| ∞ )) time algorithm for any ν+μ>0 and constant ε, unless ETH fails Estimate with additive error ε

30. Main Result 1: Weak Membership (B., Christandl, Yard ‘10) There is a exp(O(ε -2 log|A|log|B|)) time algorithm for W SEP (||*||, ε) (in ||*|| 2 or ||*| LOCC )

31. Main Result 1: Weak Membership (B., Christandl, Yard ‘10) There is a exp(O(ε -2 log|A|log|B|)) time algorithm for W SEP (||*||, ε) (in ||*|| 2 or ||*| LOCC ) Remind: NP-hard for ε = 1/poly(|A||B|) in ||*|| 2 (Gurvits ‘02, Gharibian ‘08, Beigi ‘08) Corollary: the problem in ||*|| 2 is not NP-hard for ε = 1/polylog(|A||B|), unless ETH fails

32. Main Result 2: Best Separable State (BCY ‘10) 1.There is a exp(O(ε -2 log|A|log|B|(||M|| 2 ) 2 )) time algorithm for BSS(ε) 2.For M in LOCC, there is a exp(O(ε -2 log|A|log|B|)) time algorithm for BSS(ε)

33. Main Result 2: Best Separable State (BCY ‘10) 1.There is a exp(O(ε -2 log|A|log|B|(||M|| 2 ) 2 )) time algorithm for BSS(ε) 2.For M in LOCC, there is a exp(O(ε -2 log|A|log|B|)) time algorithm for BSS(ε) Contrast with: (Harrow, Montanaro ‘10) No exp(O(log 1-ν |A|log 1-μ |B|||M|| ∞ )) time algorithm for any ν+μ>0 and constant ε, unless ETH fails, even for separable M:. Remember: Part 2 works for

34. Main Result 2: Best Separable State (BCY ‘10) 1.There is a exp(O(ε -2 log|A|log|B|(||M|| 2 ) 2 )) time algorithm for BSS(ε) 2.For M in LOCC, there is a exp(O(ε -2 log|A|log|B|)) time algorithm for BSS(ε) Contrast with: (Harrow, Montanaro ‘10) No exp(O(log 1-ν |A|log 1-μ |B|||M|| ∞ )) time algorithm for any ν+μ>0 and constant ε, unless ETH fails, even for separable M:. Remember: Part 2 works for Quantum Info Remark: The difficulty to show optimality of the algorithm is the existence of separable measurements that are not LOCC, a well studied phenomena in quantum information (e.g. Bennett et al ‘98). Here we have a new computational- complexity motivation for further studying the problem!

35. The Algorithm We consider the a Parrilo-Lasserre hierarchy of SDP relaxations to the problem introduced in (Doherty, Parrilo and Spedalieri ’01) We prove it converges to a good approximate solution in a O(log|B|) number of rounds. Previously convergence only in Ω(|B|) rounds was known.

36. Optimization Over Separable States (again) (Best Separable State BSS(ε)) Given estimate to additive error ε This is a polynomial optimization problem. One can calculate a sequence of SDP approximations to it following the approach of (Parrilo ‘00, Lasserre ’01) We’ll derive the SDP hierachy by a quantum argument

37. Entanglement Monogamy Classical correlations are shareable: Given separable state Consider the symmetric extension Def. ρ AB is k-extendible if there is ρ AB 1 …B k s.t for all j in [k], tr \ B j (ρ AB 1 …B k ) = ρ AB A B1B1 B2B2 B3B3 B4B4 BkBk …

38. Entanglement Monogamy Classical correlations are shareable: Def. ρ AB is k-extendible if there is ρ AB 1 …B k s.t for all j in [k], tr \ B j (ρ AB 1 …B k ) = ρ AB Separable states are k-extendible for every k A B1B1 B2B2 B3B3 B4B4 BkBk …

39. Entanglement Monogamy Quantum correlations are non-shareable: ρ AB separable iff ρ AB k-extendible for all k Follows from: Quantum de Finetti Theorem (Stormer ’69, Hudson & Moody ’76, Raggio & Werner ’89) Monogamy of entanglement: Very useful concept in general, application e.g. in quantum key distribution

41. Entanglement Monogamy Quantitative version: For any k-extendible ρ AB, -Follows from: Finite quantum de Finetti Theorem (Christandl, König, Mitchson, Renner ‘05)

42. Entanglement Monogamy Quantitative version: For any k-extendible ρ AB, -Follows from: Finite quantum de Finetti Theorem (Christandl, König, Mitchson, Renner ‘05) Close to optimal: there is a k-ext state ρ AB s.t. For other norms ( ||*|| 2, ||*|| LOCC, … ) no better bound known.

43. Exponentially Improved de Finetti type bound (B., Christandl, Yard ‘10) For any k-extendible ρ AB, with||*|| equals ||*|| 2 or ||*|| LOCC Bound proportional to the (square root) of the number of qubits: exponential improvement over previous bound

44. How long does it take to check if a k- extension exists? Search for a symmetric extension is a semidefinite program (Doherty, Parrilo, Spedalieri ‘04) Can be solved in poly(n) time in the number of variables n n = |A| 2 |B| 2k Our bound implies k = O(ε -2 log|A|) Time Complexity: poly(|A||B| 2k ) = exp(O(ε -2 log|A|log|B|))

45. Does it work for 1-norm? There are k-extendible states s.t. For such states the SDP hierarchy only gives good solutions for k = O(|B|), which requires exponential time But we know also: So, hard instances are always “data hiding” states, i.e.

46. Does it work for 1-norm? There are k-extendible states s.t. For such states the SDP hierarchy only gives good solutions for k = O(|B|), which requires exponential time But we know also: So, hard instances are always “data hiding” states, i.e.

47. Does it work for 1-norm? There are k-extendible states s.t. For such states the SDP hierarchy only gives good solutions for k = O(|B|), which requires exponential time But we know also: So, hard instances are always “data hiding” states, i.e.

48. Does it work for 1-norm? There are k-extendible states s.t. For such states the SDP hierarchy only gives good solutions for k = O(|B|), which requires exponential time But we know also: So, hard instances are always “data hiding” states, i.e.

49. Algorithm for Best Separable State The idea Optimize over k=O(log|A|ε -2 (||X|| 2 ) 2 ) extension of ρ AB by SDP This is precisely the Parrilo-Lasserre hierarchy for the problem! (written in a somewhat different form) By Cauchy Schwartz: By de Finetti Bound:

50. Application 1: Quantum NP

51. QMA

52. - Quantum analogue of NP (or MA) - Local Hamiltonian Problem, N-representability, … Is QMA a robust complexity class? (Aharonov, Regev ‘03) superverifiers don’t help (Marriott, Watrous ‘05) Exponential amplification with fixed proof size (Beigi, Shor, Watrous ‘09) logarithmic size interaction doesn’t help

53. New Characterization QMA Corollary QMA doesn’t change allowing k = O(1) different proofs if the verifier can only apply LOCC measurements in the k proofs

54. New Characterization QMA Corollary QMA doesn’t change allowing k = O(1) different proofs if the verifier can only apply LOCC measurements in the k proofs Def QMA m (k): analogue of QMA with k proofs and proof size m

55. New Characterization QMA Corollary QMA doesn’t change allowing k = O(1) different proofs if the verifier can only apply LOCC measurements in the k proofs Def QMA m (k): analogue of QMA with k proofs and proof size m Def LOCCQMA m (k): analogue of QMA with k proofs, proof size m and LOCC verification procedure along the k proofs.

56. QMA(k) Def QMA m (k): A language L is in QMA m (k) if there is a quantum poly-time circuit that for every instance x implements the measurement {A x, I - A x } such that Completeness: If x in L, there exists k proofs, each of m qubits, s.t. Soundness: If x not in L, for any k states, Def 2 LOCCQMA m (k): Likewise, but {A x, I - A x } must be LOCC

57. New Characterization QMA Corollary QMA = LOCCQMA(k), k = O(1) LOCCQMA m (2) contained in QMA O(m 2 ) Contrast: QMA m (2) not in QMA O(m 2-δ ) for any δ>0 unless Quantum ETH * fails And: SAT has a LOCCQMA O(log(n)) (n 1/2 ) protocol * Quantum ETH: SAT cannot be solved in 2 o(n) quantum time (Harrow and Montanaro ’10) -- based on Aaronson et al ‘08 (Chen and Drucker ’10)

58. New Characterization QMA Corollary QMA = LOCCQMA(k), k = O(1) LOCCQMA m (2) contained in QMA O(m 2 ) Contrast: QMA m (2) not in QMA O(m 2-δ ) for any δ>0 unless Quantum ETH * fails Follows from QMA n 1/2 (2) protocol for SAT with n clauses And: SAT has a LOCCQMA O(log(n)) (n 1/2 ) protocol * Quantum ETH: SAT cannot be solved in 2 o(n) quantum time (Harrow and Montanaro ’10 – built on Aaronson et al ’08) (Chen and Drucker ’10)

59. New Characterization QMA Corollary QMA = LOCCQMA(k), k = O(1) LOCCQMA m (2) contained in QMA O(m 2 ) Idea to simulate LOCCQMA m (2) in QMA: Arthur asks for proof ρ on AB 1 B 2… B k with k = mε -2 He symmetrizes the B systems and applies the original verification prodedure to AB 1 Correcteness de Finetti bound implies:

60. Application 2: Small Set Expansion Small Set Expansion Problem: Given a graph determine whether all sets of sublinear size expand almost perfectly. Introduced in (Raghavendra, Steurer ’09), where it was conjectured to be a hard problem. It’s closely related to Khot’s Unique Games Conjecture

61. Application 2: Small Set Expansion (Barak, B., Harrow, Kelner, Steurer, and Zhou ‘12) connection of the Small Set Expansion Problem to the Best-Separable-State Problem for a LOCC operator (via the 2->4 norm of a projector) Can show that the SDP hierarchy gives a subexponential-time algorithm for the small set expansion problem, matching the performance of the algorithm of (Arora, Barak and Steurer ‘10) Small Set Expansion Problem: Given a graph determine whether all sets of sublinear size expand almost perfectly. Introduced in (Raghavendra, Steurer ’09), where it was conjectured to be a hard problem. It’s closely related to Khot’s Unique Games Conjecture

62. Proof Techniques Coding Theory Strong subadditivity of von Neumann entropy as state redistribution rate (Devetak, Yard ‘06) Large Deviation Theory Hypothesis testing of separable states (B., Plenio ‘08) Entanglement Measure Theory Squashed Entanglement (Christandl, Winter ’04)

63. I(A:B|E) Conditional Mutual Information: Measures the correlations of A and B relative to E in ρ ABE I(A:B|E) ρ := S(AE) ρ + S(BE) ρ – S(ABE) ρ – S(E) ρ Always positive: I(A:B|E) ρ ≥ 0 (strong-subadditivity of entropy) When does it vanish? I(A:B|E) ρ = 0 iff ρ ABE is a “Quantum Markov Chain State” E.g. Approximate version??? … (Hayden, Jozsa, Petz, Winter ‘04) (Lieb, Ruskai ‘73)

64. New Inequality for I(A:B|E) Thm: (B., Christandl, Yard ’10) Either LOCC or 2-norm Obs: The statement fails badly for 1-norm! The monogamy bound follows from this inequality and the chain rule (via an entanglement measure called squashed entanglement)

65. Summary Testing separability is rather easy Family of Parrilo-Lasserre SDP relaxations converge in log(n) rounds; proof by a quantum argument – new approach to proving fast convergence of SDP hierarchies. New Pinsker type lower bound for I(A:B|E) QMA is robust

66. Open Problems Is there a polynomial algorithm in 2-norm? Can we close the LOCC norm vs. trace norm gap in the results? (hardness vs. algorithm, LOCCQMA(k) vs QMA(k)) Are there more applications of the bound on the convergence of the SDP relaxation? Can we prove a quasipolynomial time algorithm for Small set Expansion? And for unique games or other UG-hard prioblems? Can we put new problems in QMA using QMA = LOCCQMA(k)? Are there more application of the inequality for I(A:B|E)?

67. Thank you!

68. Proof Outline

69. Relative Entropy of Entanglement The proof is largely based on the properties of the following entanglement measure: Def Relative Entropy of Entanglement (Vedral, Plenio ‘99)

70. Entanglement Hypothesis Testing Given (many copies) of ρ AB, what’s the optimal probability of distinguishing it from a separable state?

71. Entanglement Hypothesis Testing Given (many copies) of ρ AB, what’s the optimal probability of distinguishing it from a separable state? Def Rate Function: D(ρ AB ) is maximum number r s.t. there exists {M n, I-M n }, 0 < M n < I, D LOCC (ρ AB ) : defined analogously, but now {M, I-M} must be LOCC

72. Entanglement Hypothesis Testing Given (many copies) of ρ AB, what’s the optimal probability of distinguishing it from a separable state? Def Rate Function: D(ρ AB ) is maximum number r s.t there exists {M n, I-M n }, 0 < M n < I, D LOCC (ρ AB ) : defined analogously, but now {M, I-M} must be LOCC (B., Plenio ‘08) D(ρ AB ) = E R ∞ (ρ AB ) Obs: Equivalent to reversibility of entanglement under non-entangling operations (B., Plenio ‘08)

73. Proof in 1 Line

74. (i) Quantum Shannon Theory: State redistribution Protocol (ii) Large Deviation Theory: Entanglement Hypothesis Testing (iii) Entanglement Theory: Faithfulness bounds Relative entropy of Entanglement plays a triple role: (Devetak and Yard ‘07) (B. and Plenio ‘08)

75. First Inequality Non-lockability: (Horodecki 3 and Oppenheim ‘04) State Redistribution: How much does it cost to redistribute a quantum system? ABE F AE BF ½ I(A:B|E) Proof (i): Apply non-lockability to and use state redistribution to trace out B at a rate of ½ I(A:B|E) qubits per copy

76. Second Inequality Equivalent to: Monogamy relation for entanglement hypothesis testing Proof (ii) Use optimal measurements for ρ AE and ρ AB achieving D(ρ AE ) and D LOCC (ρ AB ), resp., to construct a measurement for ρ A:BE achieving D(ρ A:BE )

77. Third Inequality Pinsker type inequality for entanglement hypothesis testing Proof (iii) minimax theorem + martingale like property of the set of separable states

78. Thank you!