+ arXiv:1501.01715 + Richard Cleve & Rolando Somma Andrew - - PowerPoint PPT Presentation

Hamiltonian simulation with nearly

• ptimal dependence on all parameters

Dominic Berry

+

Andrew Childs & Robin Kothari

+ Richard Cleve & Rolando Somma

arXiv:1501.01715

Why is this important?

Why is this important?

Aharonov & Ta-Shma 2003: Algorithm to simulate sparse Hamiltonians Harrow, Hassidim, Lloyd 2009: Quantum algorithm to solve linear systems Childs, Cleve, Jordan, Yonge-Mallo 2009: Quantum algorithm for NAND trees Clader, Jacobs, Sprouse 2013: Quantum algorithm for scattering problems Berry 2014: Quantum algorithm for differential equations Wang 2014: Quantum algorithm for effective electrical resistance

The simulation problem

Problem: Given a Hamiltonian 𝐼, simulate 𝑒 𝑒𝑢′ 𝜔(𝑢′) = −𝑗𝐼 𝜔(𝑢′) for time 𝑢 and error no more than 𝜁. Inputs: 𝐼, 𝑢, 𝜁 Parameters of 𝐼:

• 𝑒 – sparseness
• 𝑂 – dimension
• 𝐼 or 𝐼 max – norms of the Hamiltonian

Progression of results

Standard method: Advanced methods: New method: Quantum walks 𝑃(𝑒 𝐼 max𝑢/ 𝜗) Compressed product formula

• r Taylor series

𝑃(𝑒2 𝐼 max𝑢 × polylog) Product formula 𝑃(𝑒4 𝐼 𝑢 1+𝜀/𝜗𝜀) Combined approach 𝑃(𝑒 𝐼 max𝑢 × polylog)

Main results

𝑃 𝑒 𝐼 max𝑢 × polylog



Near-linear in 𝑒, like quantum walk approach.



Polylogarithmic in 𝜁, like compressed product formulae.

Ω 𝑒 𝐼 max𝑢 + polylog

Lower bound: Complexity:

What is the polylog factor? Queries: polylog ≡

log(𝜐/𝜁) log log(𝜐/𝜁)

Gates: polylog ≡

log2(𝜐/𝜁) log log(𝜐/𝜁)

𝜐 = 𝑒 𝐼 max𝑢

Model

Sparse Hamiltonians

 Query: An efficient algorithm to determine the positions and

values of non-zero entries.

Standard method

 Use decomposition as

𝐼 = 𝐼𝑙

𝑁 𝑙=1  Divide time into 𝑠 intervals and use product formula:

𝑓−𝑗𝐼𝑢 ≈ 𝑓−𝑗𝐼𝑙𝑢/𝑠

𝑁 𝑙=1 𝑠

• S. Lloyd, Science 273, 1073 (1996).

Advanced methods

1.

Compressed product formulae

2.

Implementing Taylor series

3.

Quantum walks

4.

Superposition of quantum walk steps

• D. W. Berry, A. M. Childs, R. Cleve, R. Kothari, R. D. Somma, STOC ‘14; arXiv:1312.1414 (2013).
• D. W. Berry, A. M. Childs, R. Cleve, R. Kothari, R. D. Somma, arXiv:1412.4687 (2014).
• D. W. Berry, A. M. Childs, Quantum Information and Computation 12, 29 (2012).
• D. W. Berry, A. M. Childs, R. Kothari, arXiv:1501.01715 (2015).

Compressed product formulae

Crucial ideas we use in new work:

1.

Break evolution into segments.

2.

In each segment use controlled operations.

3.

Apply oblivious amplitude amplification to achieve result deterministically.

|𝜔〉 |𝜔′〉

𝑓−𝑗𝐼𝑢/𝑠

Break into segments

1 2 . . . . . . . . . . . . . t

1 4 2 4 3 4 5 4 6 4 7 4 9 4 10 4

𝑓−𝑗𝑉𝑢|𝜔〉 〈0| 𝑉 |𝜔〉 𝑆 𝑄 𝑆

𝑉 is self-inverse

Evolution using control qubits

• R. Cleve, D. Gottesman, M. Mosca, R. Somma,
• D. Yonge-Mallo, STOC ‘09; arXiv:0811.4428 (2008).

|𝜔〉 |𝜔′〉

𝑉

𝑘

𝑉

𝑘

𝑉

𝑘

𝑋

|𝑑 0𝑛 〉

𝑋†

Oblivious amplitude amplification

1 2 . . . . . . . . . . . . . t

1 4 2 4 3 4 5 4 6 4 7 4 9 4 10 4

measure

|𝜔〉 |𝜔′〉

𝑉

𝑘

𝑉

𝑘

𝑉

𝑘

𝑋

|𝑑 0𝑛 〉

𝑋†

Oblivious amplitude amplification

1 2 . . . . . . . . . . . . . t

1 4 2 4 3 4 5 4 6 4 7 4 9 4 10 4

|𝜔〉 |𝜔′〉

𝑉

𝑘

𝑉

𝑘

𝑉

𝑘

𝑋 𝑋†

Oblivious amplitude amplification

1 2 . . . . . . . . . . . . . t

1 4 2 4 3 4 5 4 6 4 7 4 9 4 10 4

success!

Advanced methods

1.

Compressed product formulae

2.

Implementing Taylor series

3.

Quantum walks

4.

Superposition of quantum walk steps

• D. W. Berry, A. M. Childs, R. Cleve, R. Kothari, R. D. Somma, STOC ‘14; arXiv:1312.1414 (2013).
• D. W. Berry, A. M. Childs, R. Cleve, R. Kothari, R. D. Somma, arXiv:1412.4687 (2014).
• D. W. Berry, A. M. Childs, Quantum Information and Computation 12, 29 (2012).
• D. W. Berry, A. M. Childs, R. Kothari, arXiv:1501.01715 (2015).



Implementing Taylor series

 Break Hamiltonian evolution into 𝑠 segments and use

𝑓−𝑗𝐼𝑢/𝑠 ≈ 1 𝑙! −𝑗𝐼𝑢/𝑠 𝑙

𝐿 𝑙=0  Aim to perform using controlled operations.

𝑓−𝑗𝐼𝑢/𝑠|𝜔〉 〈0|

𝐼𝑙

|𝜔〉

𝑙

𝑋 𝑋†

Implementing Taylor series

 Expand 𝐼 as sum of unitaries

𝐼 ≈ 𝛿 𝑉ℓ

𝑁 ℓ=1  Exponential is then

𝑓−𝑗𝐼𝑢/𝑠 ≈ ⋯ −𝑗𝑢/𝑠 𝑙 𝑙!

𝑁 ℓ𝑙=1 𝑁 ℓ2=1 𝑁 ℓ1=1 𝐿 𝑙=0

𝑉ℓ1𝑉ℓ2 ⋯ 𝑉ℓ𝑙

measure

|𝜔〉 𝑓−𝑗𝐼𝑢/𝑠|𝜔〉

𝑉ℓ1 𝑉ℓ2 𝑉ℓ𝐿

𝑋

|0〉

𝑋†

𝑙 ℓ1 ℓ2 ℓ𝐿

Advanced methods

1.

Compressed product formulae

2.

Implementing Taylor series

3.

Quantum walks

4.

Superposition of quantum walk steps

• D. W. Berry, A. M. Childs, R. Cleve, R. Kothari, R. D. Somma, STOC ‘14; arXiv:1312.1414 (2013).
• D. W. Berry, A. M. Childs, R. Cleve, R. Kothari, R. D. Somma, arXiv:1412.4687 (2014).
• D. W. Berry, A. M. Childs, Quantum Information and Computation 12, 29 (2012).
• D. W. Berry, A. M. Childs, R. Kothari, arXiv:1501.01715 (2015).

 

Quantum walks

Classical walk



Position is integer 𝑦.



Step is map 𝑦 → 𝑦 ± 1 with equal probability. Standard quantum walk



Quantum position and coin registers |𝑦, 𝑑〉.



Alternates coin and step operators, 𝐷 𝑦, ±1 = 𝑦, −1 ± 𝑦, 1 / 2 𝑇 𝑦, 𝑑 = |𝑦 + 𝑑, 𝑑〉 Szegedy quantum walk



Two subsystems with arbitrary dimension.



Step is controlled reflection.

Szegedy quantum walk

 Controlled reflections:

𝑘 〈𝑘| ⊗ (2|𝑑

𝑘〉〈𝑑 𝑘| − 𝕁) 𝑘  After doing this we swap the two systems.  Step operation is

𝑉 = 𝑗 × SWAP × controlled reflection

 Controlled reflection can be achieved with controlled

preparation: 𝑈 = 𝑘 〈𝑘| ⊗ |𝑑

𝑘〉 𝑘

• M. Szegedy, FOCS ‘04; arXiv:quant-ph/0401053 (2004).

|𝑑

𝑘〉

controlled on 𝑘 reflect about |𝑑

𝑘〉

Szegedy walk for Hamiltonians

Three part process:

1.

Start with state in one of the subsystems, and perform controlled state preparation 𝑈.

2.

Perform steps of quantum walk 𝑉 to approximate Hamiltonian evolution.

3.

Invert controlled state preparation, so final state is in one of the subsystems.

• A. M. Childs, Commun. Math. Phys. 294, 581 (2009).

𝑉 𝑈 𝑈† Each 𝑉 or 𝑈 uses 𝑃(1) calls to 𝐼.

Eigenvalues of walk

 Hamiltonian 𝐼 has eigenvalues 𝜇.  Step 𝑉 has eigenvalues

𝜈± = ±𝑓±𝑗 arcsin 𝜇

 Evolution under the Hamiltonian has

eigenvalues 𝑓−𝑗𝜇𝑢

 Given knowledge of + or − we can

correct to 𝑉𝑑 with eigenvalues 𝜈 = 𝑓−𝑗 arcsin 𝜇

𝜌 − arcsin 𝜇 arcsin 𝜇 𝜇

|𝜈−〉 |𝜈+〉

Eigenvalues of walk

 Step 𝑉𝑑 has eigenvalues

𝜈 = 𝑓−𝑗 arcsin 𝜇

 We aim for

𝑓−𝑗𝜇𝑢

 Try superposition of operations

𝑉sup = 𝛽𝑙𝑉𝑑

𝑙 𝐿 𝑙=−𝐿

𝑉sup|𝜔〉 〈0|

𝑉𝑑

𝑙

|𝜔〉 𝑋 𝑋†

𝜌 − arcsin 𝜇 arcsin 𝜇 𝜇

|𝜈−〉 |𝜈+〉

Choosing values for 𝛽𝑙

 We aim to find 𝛽𝑙 such that

𝛽𝑙𝜈𝑙

𝐿 𝑙=−𝐿

≈ 𝑓−𝑗𝜇𝑢

 The formula for 𝜈 gives

𝑓−𝑗𝜇𝑢 = exp 𝑢 2 𝜈 − 1 𝜈

 But this is the generating function for Bessel functions!

𝐾𝑙 𝑢 𝜈𝑙

∞ 𝑙=−∞

= exp 𝑢 2 𝜈 − 1 𝜈

 We can choose 𝛽𝑙 just from Bessel functions.

Without correcting the step

 We aim to find 𝛽𝑙 such that

𝛽𝑙𝜈±

𝑙 𝐿 𝑙=−𝐿

≈ 𝑓−𝑗𝜇𝑢

 The formula for 𝜈± gives

𝑓−𝑗𝜇𝑢 = exp − 𝑢 2 𝜈± − 1 𝜈±

 But this is the generating function for Bessel functions!

𝐾𝑙 −𝑢 𝜈±

𝑙 ∞ 𝑙=−∞

= exp − 𝑢 2 𝜈± − 1 𝜈±

 We can choose 𝛽𝑙 just from Bessel functions.  We don’t need to distinguish + from − or correct the step!

The complete algorithm

 Map into doubled Hilbert space.  Divide the time into 𝑠 = 𝑒 𝐼 max𝑢 segments.  For each segment:

• 1. Perform the superposition.

 Map back to original Hilbert space.

Total complexity: 𝑒 𝐼 max𝑢 × 𝐿

2.

Use amplitude amplification to

• btain success deterministically.

𝑉sup|𝜔〉 〈0|

𝑉𝑙

|𝜔〉 𝑋 𝑋†

𝐾𝑙(𝑦)

t

𝐾𝑙(1) 𝑙

Choosing the value of 𝐿

 Bessel function may be bounded as

𝐾𝑙 𝑦 ≤ 1 𝑙! 𝑦 2

𝑙  Scaling is the same as for Taylor series!  We can choose 𝐿 to be polylog

𝐿 ∼ log(𝜐/𝜁) log log(𝜐/𝜁)

 Overall scaling is

𝑃(𝑒 𝐼 max𝑢 × polylog)

Single-segment approach

𝐾𝑙(𝑒 𝐼 max𝑢) 𝑙 𝑒 𝐼 max𝑢 polylog Choosing segment sizes 𝜐𝛽 gives complexity 𝜐1+𝛽/2 + 𝜐1−𝛽/2 log(1/𝜁)

Lower bound

 Computing parity of a bit string 𝑦1, … , 𝑦𝑂 has complexity Ω(𝑂).  Can define a Hamiltonian such that evolving under the

Hamiltonian gives transition from |0〉 to |𝑂〉.

 Can define a Hamiltonian such that the states are connected

according to values of bits. Evolving gives parity.

 𝑂 ∝

𝐼 max𝑢 gives Ω( 𝐼 max𝑢) lower bound.

 Take 𝑒 copies of each node, and use superposition.  𝑂 ∝ 𝑒 𝐼 max𝑢 gives Ω(𝑒 𝐼 max𝑢) lower bound.

|0〉 |𝑂〉 |0,0〉 |𝑂, parity〉 |0,0,∗〉 |𝑂, parity,∗〉

Conclusions



We have complexity of sparse Hamiltonian simulation scaling as 𝑃(𝑒 𝐼 max𝑢 × polylog)



The lower bound is scaling as Ω(𝑒 𝐼 max𝑢 + polylog)



The method combines the quantum walk and compressed product formula approaches.

arXiv:1501.01715