Inverting Well Conditioned Matrices in Quantum LogSpace Amnon Ta-Shma Tel-Aviv University.

Slides:



Advertisements
Similar presentations
Pretty-Good Tomography Scott Aaronson MIT. Theres a problem… To do tomography on an entangled state of n qubits, we need exp(n) measurements Does this.
Advertisements

Quantum Circuit Decomposition
University of Queensland
3D Geometry for Computer Graphics
Quantum Computing MAS 725 Hartmut Klauck NTU
Divide and Conquer. Recall Complexity Analysis – Comparison of algorithm – Big O Simplification From source code – Recursive.
Theoretical Program Checking Greg Bronevetsky. Background The field of Program Checking is about 13 years old. Pioneered by Manuel Blum, Hal Wasserman,
Nattee Niparnan. Recall  Complexity Analysis  Comparison of Two Algos  Big O  Simplification  From source code  Recursive.
The number of edge-disjoint transitive triples in a tournament.
Advanced Topics in Algorithms and Data Structures
(Omer Reingold, 2005) Speaker: Roii Werner TexPoint fonts used in EMF. Read the TexPoint manual before you delete this box.: AA A AA A A A A AA A.
Quantum Phase Estimation using Multivalued Logic.
Computer Graphics Recitation 5.
Randomized Algorithms and Randomized Rounding Lecture 21: April 13 G n 2 leaves
Undirected ST-Connectivity 2 DL Omer Reingold, STOC 2005: Presented by: Fenghui Zhang CPSC 637 – paper presentation.
University of Queensland
Quantum Automata Formalism. These are general questions related to complexity of quantum algorithms, combinational and sequential.
Undirected ST-Connectivity in Log-Space By Omer Reingold (Weizmann Institute) Year 2004 Presented by Maor Mishkin.
EECS 598: Background in Theory of Computation Igor L. Markov and John P. Hayes
Preference Analysis Joachim Giesen and Eva Schuberth May 24, 2006.
Complexity 1 Mazes And Random Walks. Complexity 2 Can You Solve This Maze?
Study Group Randomized Algorithms Jun 7, 2003 Jun 14, 2003.
Anuj Dawar.
6 1 Linear Transformations. 6 2 Hopfield Network Questions.
1 Recap (I) n -qubit quantum state: 2 n -dimensional unit vector Unitary op: 2 n  2 n linear operation U such that U † U = I (where U † denotes the conjugate.
Randomness in Computation and Communication Part 1: Randomized algorithms Lap Chi Lau CSE CUHK.
Undirected ST-Connectivity In Log Space
Quantum Counters Smita Krishnaswamy Igor L. Markov John P. Hayes.
Undirected ST-Connectivity In Log Space Omer Reingold Slides by Sharon Bruckner.
Dynamic Programming Introduction to Algorithms Dynamic Programming CSE 680 Prof. Roger Crawfis.
Dominant Eigenvalues & The Power Method
1 Introduction to Quantum Information Processing QIC 710 / CS 768 / PH 767 / CO 681 / AM 871 Richard Cleve QNC 3129 Lecture 18 (2014)
CS 461 – Nov. 21 Sections 7.1 – 7.2 Measuring complexity Dividing decidable languages into complexity classes. Algorithm complexity depends on what kind.
Alice and Bob’s Excellent Adventure
One Complexity Theorist’s View of Quantum Computing Lance Fortnow NEC Research Institute.
1 Introduction to Quantum Information Processing QIC 710 / CS 678 / PH 767 / CO 681 / AM 871 Richard Cleve DC 2117 / QNC 3129 Lectures.
Algorithms for a large sparse nonlinear eigenvalue problem Yusaku Yamamoto Dept. of Computational Science & Engineering Nagoya University.
Stochastic Algorithms Some of the fastest known algorithms for certain tasks rely on chance Stochastic/Randomized Algorithms Two common variations – Monte.
Theory of Computing Lecture 15 MAS 714 Hartmut Klauck.
Quantum Computing MAS 725 Hartmut Klauck NTU
Short course on quantum computing Andris Ambainis University of Latvia.
Quantum Factoring Michele Mosca The Fifth Canadian Summer School on Quantum Information August 3, 2005.
Week 10Complexity of Algorithms1 Hard Computational Problems Some computational problems are hard Despite a numerous attempts we do not know any efficient.
Chapter 5 Eigenvalues and Eigenvectors 大葉大學 資訊工程系 黃鈴玲 Linear Algebra.
Quantum Computing MAS 725 Hartmut Klauck NTU
1 Dorit Aharonov Hebrew Univ. & UC Berkeley Adiabatic Quantum Computation.
Umans Complexity Theory Lectures Lecture 1a: Problems and Languages.
06/12/2015Applied Algorithmics - week41 Non-periodicity and witnesses  Periodicity - continued If string w=w[0..n-1] has periodicity p if w[i]=w[i+p],
Sorting: Implementation Fundamental Data Structures and Algorithms Klaus Sutner February 24, 2004.
1 Introduction to Quantum Information Processing CS 467 / CS 667 Phys 467 / Phys 767 C&O 481 / C&O 681 Richard Cleve DC 3524 Course.
1 Introduction to Quantum Information Processing CS 467 / CS 667 Phys 667 / Phys 767 C&O 481 / C&O 681 Richard Cleve DC 653 Lecture.
1 Introduction to Quantum Information Processing QIC 710 / CS 667 / PH 767 / CO 681 / AM 871 Richard Cleve DC 2117 Lectures
The NP class. NP-completeness Lecture2. The NP-class The NP class is a class that contains all the problems that can be decided by a Non-Deterministic.
Unsupervised Learning II Feature Extraction
Space Complexity Guy Feigenblat Based on lecture by Dr. Ely Porat Complexity course Computer science department, Bar-Ilan university December 2008.
Beginner’s Guide to Quantum Computing Graduate Seminar Presentation Oct. 5, 2007.
ALGEBRAIC EIGEN VALUE PROBLEMS
The NP class. NP-completeness
Richard Cleve DC 2117 Introduction to Quantum Information Processing CS 467 / CS 667 Phys 667 / Phys 767 C&O 481 / C&O 681 Lecture.
The complexity of the Separable Hamiltonian Problem
Introduction to Quantum Computing Lecture 1 of 2
Markov Chains Mixing Times Lecture 5
Streaming & sampling.
Background: Lattices and the Learning-with-Errors problem
Turnstile Streaming Algorithms Might as Well Be Linear Sketches
OSU Quantum Information Seminar
Undirected ST-Connectivity In Log Space
Maths for Signals and Systems Linear Algebra in Engineering Lectures 13 – 14, Tuesday 8th November 2016 DR TANIA STATHAKI READER (ASSOCIATE PROFFESOR)
Eigenvalues and Eigenvectors
On Solving Linear Systems in Sublinear Time
Presentation transcript:

Inverting Well Conditioned Matrices in Quantum LogSpace Amnon Ta-Shma Tel-Aviv University

Space Bounded Complexity Space complexity measures the memory size needed for solving a problem.

An Example: Multiplying two matrices Input: Two n  n matrices A,B. Output: C = AB. Algorithm: C i,j =  A i,k B k,j For i=1,..,n For j=1,..,n { c=0; For k=1,..,n c = c + A i,k B k,j ; output c; } We do not count the input as working area, because we are not allowed to change it. We do not count the output as working area, because we are not allowed to read or change it. We view it as sending the output to a printer. We only count memory elements that we can read, write and change (i,j,c,k) as working area.

An Example: Multiplying two matrices Input: Two n  n matrices A,B. Output: C = AB. Algorithm: C i,j =  A i,k B k,j The input is not counted The output is not counted The only thing we count is memory we can read, write and change. The algorithm above runs in O(log n) space.

Another Example: Undirected connectivity Input: An undirected graph G=(V,E). Output: Is the graph connected? Can be solved with linear space and time. Omer Reingold showed the problem can be solved with logarithmic space and polynomial time.

Problems not known to be in Log Connectivity of directed graphs. Determinant of an integer matrix. Inverting an integer matrix. NL – complete. DET – complete. NL – Non-deterministic Logspace. DET – all languages that are LogSpace reducible to integer determinant.

What is known Log  NL  DET  DSPACE(log 2 n) STCON is NL complete Matrix inversion, int determinant are DET complete.

Probabilistic space-bounded computation BPL – the class of languages that are solvable by space-bounded machines that have online access to an unbounded sequence of truly uniform bits. Log  BPL  DET BPL  DSPACE(log 1.5 n) [SaksZhou]

Quantum space-bounded computation BQL – all languages solvable by a LOG machine that may use O(log n) qubits. Counting the number of qubits is a natural complexity measure. The definition has several variants, and we will discuss it soon.

What do we know about BQL? Log  BPL  BQL  DSPACE(log 2 n) Not much else is known. No natural candidate for a language in BQL not known to be in BPL.

In this talk We will modify an algorithm of Harrow, Hassidim, Lloyd for approximated matrix inversion. HHL studied quantum time complexity. We will study quantum space cpmplexity. We will show the problem is in BQL The problem is not known to be in BPL.

Quadratic gap This is first natural candidate for a problem in BQL not in BPL. It Presents a quadratic gap between BQL and what we currently know in BPL, and this gap is best possible. Our work might lead to new classical algorithms.

Defining BQL

Deterministic Space TM Input tape: Read only, Head moves in all directions Output tape: Write only, Head moves Left Work tape: Read/Write, All directions.

Quantum space-bounded machines An additional quantum tape with O(log n) qubits. Two heads over the quantum tape. The allowed quantum operations are: HAD, CNOT, T plus measurements M in the standard basis. H H M H H TT MT TT T MXX X X X X Intermediate measurement

Classical control We use the usual  function mechanism. The  function only depends on the classical data.  : Q x  Input x  Work  Q x  Work x  out x {L,R} 4   (q CNOT ) applies CNOT on the qubits under the two heads  Similarly for  (q HAD ) and  (q T )   (q M ) measures the qubit under the first head in the standard basis. Moves to q M,0,q M,1 depending on answer.

BQL O(log n) classical bits and qubits. Classical control. Intermediate measurements. BQL without intermediate measurements is also interesting but possibly much weaker. H H M H H TT MT TT T MXX X X X X

Matrix inversion and the HHL algorithm

Time complexity of Matrix inversion Can be solved as fast as matrix multiplication. Current best time O(n  ),  Matrix inversion depends on all input bits and so the time complexity must be  (n 2 ).

The HHL problem The HHL algorithm studies a modified version of matrix inversion: Input: A matrix A, a vector b, Output: Approximation of certain predicates of x=A -1 b Since we deal with approximation, the input matrix has to be stable.

Stability – the condition number Matrix inversion is not stable, if there exists an eigenvalue close to 0. Matrix inversion is stable if all eigenvalues are far from zero. The condition number  (A) is defined to be  (A)= ||A|| / ||A -1 ||

HHL’09 Input: A matrix A, a vector b, condition number k Output: Approximation of certain predicates of x=A -1 b Quantum Time complexity: O(k log n). Exponentially faster than the classical time bound Ω(n).

HHL - Summary Only an approximation Only for well conditioned A Only for sparse matrices A Only for special b Only for certain predicates over x. Very nice idea! Surprising technique and result.

Our result Input: A matrix A, condition number k Output: Approximation of A -1 Quantum space complexity: O(log(kn)). Currently best classical bound O(log 2 n). Quadratic gap.

The technique

Basic idea: Sampling the spectrum using phase estimation

First observation: We can work with Hermitian matrices Given input A. We look for the SVD, A=UDV. Define H=, H is Hermitian. The SVD of H is And it so happens that one can read A’s decomposition from H’s decomposition. We also assume all eigenalues are well-separated. 0 0 A A†A† 0 0 V U†U† 0 0 D D†D† 0 0U V†V†

Basic approach Input: Hermitian A. U=e iA is unitary. Assume: We can simulate U t for t=1,…,T, and, We know an eigenvector v of U. Then, using phase estimation, we estimate the eigenvalue λ associated with v.

First challenge: How do we find an eigenvector? Classically: A big question. Once we know A and an eigenvector v, We can easily compute in small space.

Sampling instead of finding an eigenvector The completely mixed state I is the mixture Obtained by taking a uniform eigenvector of A. If we apply phase estimation on I we sample a random (eigenvector,eigenvalue) pair of A. We can generate the uniform distribution over the eigenvectors of A, even though we do not know any specific eigenvecctor.

2 nd Challenge: Simulate U=e iA {HAD, CNOT, T} is a universal basis, Hence any unitary U can be approximated by a circuit with these gates. The challenge is designing a deterministic Log space algorithm that given A produces a quantum circuit over {HAD, CNOT, T} That approximates U=e iA.

Reminder: Universality of {HAD, CNOT, T} Given a unitary U: 1.Decompose U to a product of 2-level unitaries. 2.Convert a 2-level unitary to a product of CNOT and 1-qubit unitaries. 3.Approximate any 1-qubit unitary by a short product of {HAD, T} A unitary that acts non-trivially only on a 2 dimensional subspace spanned by 2 standard basis vectors. Using the Solovay Kitaev Theorem.

Simulating U=e iA in small space Given a unitary U: 1.Approximately decompose it to a product of 2-level unitaries, using Trotter formula. 2.Convert a 2-level unitary to a product of CNOT and 1-qubit unitaries. 3.Approximate any 1-qubit unitary by a short product of {HAD, T} using a space- efficient version of the Solovay-Kitaev theorem, recently proved by [vM,W].

Altogether: Given A: Run phase estimation with U=e iA on the completely mixed state. This uniformly sample an approximation of an (eigenvector, eigenvalue) pair, in logarithmic space.

Approximating the whole spectrum

First attempt: Repeated sampling Assume all eigenvalues are in [-1,1]. Divide [-1,1] to small consecutive intervals. For each interval, pick poly(n) independent samples, and estimate the number of eigenvalues in the interval by the fraction of samples that fall into it. 1

A problem: eigenvalues close to a boundary Eigenvalues that lie close to an interval boundary might fall into both neighboring intervals and lead to wrong results. 1

The solution: Consistent estimation A probabilistic/quantum algorithm estimates a value z, if w.h.p. it outputs a value close to z. A probabilistic/quantum algorithm consistently estimates a value z, if w.h.p. it outputs a fixed value close to z. Consistent sampling solves the problem above.

Consistent Sampling using the shift & truncate method [SZ] Original accuracy:

Consistent Sampling using the shift & truncate method [SZ] Original accuracy: New accuracy: Pick uniformly a value 0 < k < 2 10 and fix it. Shift the eigenvalues by the fixed shift k* Now, w.h.p., all eigenvalues are far away from a boundary

Approximate the spectrum using consistent sampling Divide [-1,1] to small consecutive intervals. For each interval, pick poly(n) independent samples, and estimate the number of eigenvalues in the interval by the fraction of samples that fall into it. 1

Approximating the eigenvectors

Quantum state tomography Quantum tomography is the process of reconstructing the quantum state for a source by measurements on the systems coming from the source. Quantum tomography is possible if we can repeatedly and consistently generate the same state.

Estimating an eigenvector We saw we can consistently estimate an eigenvalue i. Each time we get i we have the n-dimensional eigenvector v i, represented with log(n) qubits. Using quantum state tomography we efficiently output the n coordinates of v i.

Quantum tomography in small space Where: E (1) projects onto |k>, E (2) projects onto | l >, E (3) projects onto |k>+| l >, and, E (4) projects onto |k>+i | l >, For each k, l:

Inverting a matrix

Inverting a matrix whose eigenvalues are well separated. Approximate the eigenvalues, D=Diag( 1,…, n ) Approximate the eigenvectors v 1,…,v n. V=(v 1,…,v n ) Then, A  VDV † A -1  VD -1 V †

Some reflections

BQL is surprisingly powerful Either: BQL is indeed stronger than BPL, or BPL is also surprisingly powerful. Reingold showed USTCON  L, So far, this was not extended to RL=L. An intriguing question : Can one approximately invert stochastic matrices in BPL?

     ≤    >  ×≥   llll† ρ    α      ≤    >  ×≥   

Space complexity Matrix inversion is in DSPACE(log 2 n) and is complete in DET NC 1  LogSpace  NL   DET   AC 1  NC 2  DSPACE(log 2 n)

A quick reminder: The reason it works If λ=(ω n ) t Applying inverse QFT returns t, and therefore λ.