SHORTER GATE SEQUENCES FOR QUANTUM COMPUTING BY MIXING UNITARIES PHYSICAL REVIEW A 95, (2017) ARXIV:

Transcription

1 SHORTER GATE SEQUENCES FOR QUANTUM COMPUTING BY MIXING UNITARIES PHYSICAL REVIEW A 95, (2017) ARXIV:

2 BIG PICTURE better compiling Can Randomisation help? if then do X, but if then do Y

3 OUTLINE 1 Gate-synthesis / compiling 2 Measuring noise and coherence 3 The RESULT 4 The PROOF

4 Gate-synthesis /compiling

5 GATE SYNTHESIS / COMPILING G Gate set A collection of unitaries used to build circuits G = {A, B, C} ABC, ABBC, CBC, e.g. from we can build etc Universality - Informal statement G is universal if it can implement any unitary (upto finite precision) Universality - Formal statement G V > 0 U 2hGi d(u, V ) apple is universal if for any target unitary and there exists a finite circuit such that Example Clifford+T or Clifford+Toffoli

6 GATE SYNTHESIS / COMPILING C : G! R + Cost model Each elementary gate given a positive valued cost This induces a circuit cost if U = Y i G i then C(U) = X i C(G i ) Example Uniform cost model: for all C(G) =1 G 2 G Magic state cost model / T-count: and for all in the Clifford group. C(T )=1 C(C) =0 C

7 GATE SYNTHESIS / COMPILING blackbox software V > 0 (G, C) Unitary compiler U 2hGi with U V 1 apple usually with some promise C(U) apple f( ) For efficient compilers The promise function f( ) is often polylog f( ) apple A log(1/ ) An optimal compiler will have the lowest possible C(U) and f( )

8 GATE SYNTHESIS / COMPILING Solovay-Kitaev Consider any universal gate set G (generating a group) with uniform cost. We can efficiently solve the compiling problem using where is a constant dependent on G C(U) apple O(log(1/ ) ) Comments: For Clifford+T, the Solovay-Kitaev method gives See: Dawson & Nielsen QIC, 6 81 (2006) apple 3.97 Solovay-Kitaev is efficient but not optimal.

9 GATE SYNTHESIS / COMPILING Modern methods Let G be Clifford+T and consider the T-count. We can efficiently solve the compiling problem using C(U) apple O(log(1/ )) Refs: Kliuchnikov, Maslov, Mosca, QIC (2013) arxiv: Ross and Selinger QIC (2016) arxiv: Comments: Solver is optimal (under mild assumptions) Asymptotically C(U) apple O(log(1/ )) is much better than C(U) apple O(log(1/ ) 3.97 ) Practically means 100s of gates rather than 10,000s of gates Convenient command line tool

10 GATE SYNTHESIS / COMPILING

11 GATE SYNTHESIS / COMPILING Random compiling problem Given a and output a probability distribution of circuits, realising V > 0 such that E( ) = X p i U i U i i d(e, V) apple and minimise C(E) Def: U( ) := U U Cost models for random circuits Our results for worst-case But average cost also interesting C(E) := max(c(u i )) C(E) := X i p i C(U i )

12 Measuring noise & Coherence

13 MEASURING NOISE & COHERENCE Desiderata for d(e, V) : must compose nicely d( V, E ) = d( V, E ) V (3) V (1) E (3) E (1) d(, E (2) ) V (4) V (2) E (4) apple X j ( V (j), E (j) ) to leading order

14 MEASURING NOISE & COHERENCE What distance we use matters! Average infidelity d F (E,U) := 1 Z F (E( ),U i) =1 Z h U E( ih )U i looks simple, often used to report experimental results. fails our desiderata!

15 MEASURING NOISE & COHERENCE What distance we use matters! Diamond norm distance d (E, U) := 1 2 E where U = 1 2 max X 1 := Tr[ p X X] (E I)( ) (U I)( ) 1 1 is the Schatten 1-norm composed nicely (yes to desiderata); due to nice properties, used in proofs, e.g. of threshold theorem. So <1% threshold statements refer to diamond distance!

16 MEASURING NOISE & COHERENCE Algorithm level coherent noise : Consider gate set G = {A, B, C} When we compile a target gate each individual gate is perfect V we perform U ABACABABABA = Ve i The approximation error (due to finite sequence length) is a coherent error. e i For coherent noise d (U, V) p d F (U, V) so choice of distance measures is important!

17 MEASURING NOISE & COHERENCE Contrast physical level coherent noise Consider a scenario where we want to implement sequence from gate set but our experiment systematically performs G = {A, B, C} G = {Ã, B, C} Ã = Ae i A B = Be i B we have, where are small noise terms. Randomised compiling can help say U is target unitary is unitary with coherent noise Ũ E. Knill arxiv:quant-ph/ Wallman & Emerson, Phys Rev A 94, (2016) so that d (Ũ, U) = and d F (Ũ, U) 2 then there exist twirling schemes E( ) = X i p i U i U i where U i = P i ŨP 0 i with d (E, U) 2

18 The RESULT

19 THE RESULT Preamble Let G be a universal gate set with cost measure C assume some blackbox unitary compiler V 0 < apple 0.01 equipped with promise C(U) apple f( ) Theorem then there exists a random sequence with and cost E( ) = X j p j U j U j d (V, E) apple 10 2 C(E) apple f( ) 2nd Theorem (paraphrased) For single qubit axial rotations: the assumptions can be relaxed and inequalities tightened slightly.

20 THE RESULT same cost gets you better error suppression Theorem then there exists a random sequence with and cost E( ) = X j p j U j U j d (V, E) apple 10 2 C(E) apple f( ) same error suppression for lower cost? Corollary if the unitary cost is polylog C(U) apple f( ) =A log 2 (1/ ) then there exists a random circuit E( ) = X j p j U j U j with d (V, E) apple and cost C(E) apple C f( ) (1/2) 1 C = 1+ log(a) 1!!0 2 log(1/ ) 2

21 THE RESULT WE SAVE ~ 2 Known gamma values: For single qubit Clifford+T gate set and T-count cost metric C(U) apple f( ) = 9 log(1/ ) =1 so and roughly 2x saving = = = 10 5 optimal unitary random compiling 314 or less 165 or less 90 or less Using Solovay-Kitaev for single qubits For optimal multiqubit synthesis =? 3.97, so save unknown, conjecture

22 THE RESULT Is derandomisation possible? p (1) 1 p(2) 1 p(3) 1 p(4) 1 p (1) 2 p(2) 1 p(3) 1 p(4) 1 p (1) 2 p(2) 2 p(3) 1 p(4) U (3) 1 U (4) 1 U (3) 1 U (4) 1 U (3) 1 U (4) 1 U (2) 1 U (2) 1 U (2) 2 U (1) 1 U (1) 2 U (1) 2 i i i TRUE THAT: For fixed input there is one best choice of unitaries that maximise fidelity BUT: best unitary depends on choice of input state; can t be calculated without simulating the whole computation!.

23 The PROOF

24 THE PROOF STEP 1. Show from suitable unitaries/hamiltonians we can find random circuit with quadratically reduced error suppression. STEP 2. Give concrete algorithm finding suitable Hamiltonians.

25 THE PROOF STEP 1. Show from suitable unitaries/hamiltonians we can find random circuit with quadratically reduced error suppression.* STEP 2. Give concrete algorithm finding suitable Hamiltonians. * Similar step 1 result found simultaneously by Hastings arxiv: QIC (2017)

26 THE PROOF STEP 1. Show from suitable Hamiltonians we can find random circuit V U j = Ve ih j Lemma. Given target unitary and a set of unitaries such that (1) All Hamiltonians within of origin H j 1 apple δ H 1 H 1 H 3 δ (2) Hamiltonians enclose the origin X 9p j 0 apple p j pj =1 X p j H j =0 j H 3 H 2 δ H 1 H 2 then E( ) = X j p j U j U j satisfies d (E, V) apple 1.1 2

27 THE PROOF Sketch start with d (E, V) =d (V E, identity) and V E( ) = X j p j e ih j e ih j Expand exponential and find all first order terms cancel due to V E( ) = + O(H 2 j, ) Subtracting identity gives V E( ) = O(H 2 j, ) X j p j H j =0 Leaving H 2 j and higher order terms Take 1-norm, use triangle inequality, Hölder's inequality, carefully bound higher order terms V E( ) 1 = O( H j 2 1)=O( 2 )

28 THE PROOF STEP 2. Give concrete algorithm finding suitable Hamiltonians. assume some blackbox unitary compiler V U equipped with promise C(U) apple f( ) It follows U 1 = Ve ih 1 where H j 1 apple = This gives 1 nearby Hamiltonian, δ H 1 but in a D variable problem we typically need D+1 points to enclose the origin! e.g. For 2 variable problems we need 3 points/hamiltonians H 1 H 3 H 2

29 THE PROOF STEP 2. Give concrete algorithm finding suitable Hamiltonians. (a) (b) τ 2 (c) τ 2 H 2 H 1 H 1 = µ 2 H 1 Find near origin H 1 Select new target 2 / H 1 Find H 2 near 2 (d) τ 3 (e) H 3 τ 3 (f) H 3 H 2 H 2 H 2 µ 3 H 1 H 1 H 1 Find µ 3 = p 1 H 1 + p 2 H 2 near origin and 3 / µ 3 Find H 3 near 3 µ 4 Find near origin EXIT when origin enclosed or continue looping

30 THE PROOF STEP 2. Give concrete algorithm finding suitable Hamiltonians. We prove that: Each iteration give a new within of origin. Each oracle call gives a new mixture µ n = nx j=1 p j H j H j When µ n 1 =0 the program EXITs, but can also EXIT when µ n 1 2 and convergence is exponentially fast µ n 1 apple 6 e 0.62n Furthermore (unproven), Geometric intuition hints that algorithm terminates in constant number of steps

31 THE PROOF STEP 1. Show from suitable Hamiltonians we can find random circuit showed E( ) = X j p j U j U j satisfiesi d (E, V) apple STEP 2. Give concrete algorithm finding suitable Hamiltonians. found unitaries with H j 1 apple = COMBINED RESULT d (E, V) apple 10 2

32 The END bit

33 THE END BIT We reviewed how Unitary compilers have coherent noise on the algorithm level. Coherent noise at the physical level can be quadratically reduced using random circuits. This work showed Our random circuits give free quadratic error suppression; Free quadratic error suppression can be swapped for shorter gate sequences; For single qubit gates we save a factor 2x; Savings could be larger for multiqubit unitaries. Future work Implement and numerically test. Classical runtime convergence in constant time? Better algorithms?

34 THE END BIT HIRING 4 POSTDOCS QCDA - Quantum Code Design & Architecture Sheffield - London - Dan Browne Paris - Anthony Leverrier & Jean Pierre Tillich Delft - Barbara Terhal, Koen Bertels & Carmina Garcia Almudever Munich - Robert Koenig Funded though

35 THANK YOU!