Designing algorithms for near-term quantum

computers
Omar Shehab
Quantum Computer Scientist
IonQ/UMD

15 November 2019
shehab@ionq.com

Georgia Tech Quantum Alliance Colloquium

1
IONQ www.ionq.co Proprietary Material

Collaborators
IonQ ORNL UMD Stanford
● Matthew Kissan ● Raphael Pooser ● Kevin Landsman ● Isaac H Kim
● Yunseong Nam ● Eugene Dumitrescu ● Norbert Linke
JHU/APL
● Alex McCaskey ● Cinthia Huerta
● Greg Quiroz
● Pavel Lougovski Alderete
● Thomas Papenbrock ● Nhung H. Nguyen
● Daiwei Zhu
● Chris Monroe

2
Outline

● What is quantum computing?

● Building trapped-ion quantum computers
● Near-term quantum computing
● An example algorithm
● Quantum supremacy

3 3
Earlier this month

Microsoft CEO Satya Nadella announced IonQ’s quantum computer access

through Microsoft cloud service on November 4 at Microsoft Ignite 2019 4 4
What is quantum computing?

● Proposed by Feynman in
1985 with the motivation
of simulating quantum
systems
● Based on the principles of
quantum mechanics
● Potential advantage over IonQ IBM
classical computing
comes from quantum
phenomena like quantum
superposition, quantum
entanglement, and
quantum measurement

Google Rigetti
5 5
What is quantum computing?

● Not a silver bullet for all computationally hard (for

example NP-complete) problems
● Non-von Neumann architecture
● Data is stored in qubits residing in Hilbert space
● Logical operations are complex unitary matrices
● Readout is probabilistic
● Algorithms are often visualized as “quantum
circuits”

6 6
What is quantum computing?

Example: Quantum algorithm to find an element from an

unstructured database.

Source: Wikipedia page of the Grover’s algorithm 7 7

 Quantum computational complexity

n X n chess/Go PSPACE

Finding the ground state of

k-local Hamiltonian QMA

NP-complete
Traveling salesperson
P
Graph isomorphism NP
BQP
P

Integer factoring P
P
Primality checking

Extending the graphics from Scott Aaronson’s lecture notes

8 8
 Quantum algorithm timeline
Non-convex
optimization with
Algorithm for Quantum chemistry
Comparable
prime factoring Simulation with
approximation ratio
With exponential Exponential
speedup speedup

1985 1994 1996 2005 2008 2014 2015

Feynman proposes Solving special Quantum chemistry

quantum computing Unstructured systems of linear simulation on
database search equations with near-term quantum
with quadratic exponential computers
speedup speedup

9 9
* References in the back up slides
 Quantum algorithm timeline
Non-convex
optimization with
Algorithm for Quantum chemistry
Comparable
prime factoring Simulation with
approximation ratio
With exponential Exponential
speedup speedup

1985 1994 1996

Triggered2008
2005
funded
2014 2015
academic
Feynman proposes Solvingresearch
special Quantum chemistry
Unstructured systems of linear simulation on
quantum computing
database search
programs
equations with near-term quantum
with quadratic exponential computers
speedup speedup

10 10
* References in the back up slides
 Quantum algorithm timeline
Non-convex
optimization with
Algorithm for Quantum chemistry
Comparable
prime factoring Simulation with
approximation ratio
With exponential Exponential
speedup speedup
Triggered private
industrial effort
1985 1994 1996 2005 2008 2014 2015
mainly focusing on
chemistry/physics
Feynman proposes simulation Solving special Quantum chemistry
quantum computing Unstructured systems of linear simulation on
database search equations with near-term quantum
with quadratic exponential computers
speedup speedup

11 11
* References in the back up slides
 Quantum algorithm timeline
Non-convex
optimization with
Algorithm for Quantum chemistry
Potential forprime factoring Comparable
Simulation with
approximation ratio
commercialWith exponential Exponential
incentives stillspeedup
not speedup

clear for
1985
ML/DL/optimization 1994 1996 2005 2008 2014 2015

Feynman proposes Solving special Quantum chemistry

quantum computing Unstructured systems of linear simulation on
database search equations with near-term quantum
with quadratic exponential computers
speedup speedup

12 12
* References in the back up slides
Potential market

Pharma Energy Transportation Finance

molecular design chemical processes navigation market models
drug delivery material design logistics optimization

13 13
Building quantum computers @ IonQ

Source: arXiv:1903.08181
14 14
The quantum processing unit

Linear computation: montage of a photo of the chip containing the trapped

ions and an image of the ions in a 1D array (Courtesy: Christopher Monroe) 15 15
 Trapped Ion Hyperfine Qubit: 171Yb+
• State-dependent fluorescence provides high fidelity detection
Single
ion
γ/2π = 20
MHz
2 2.1 GHz
P1/2

369 nm
(811.9THz)

2 |↑〉 = S. Olmschenk et al., PRA 76, 052314, Slide (2007)

S1/2 |1,0〉 From the March Meeting talk of Jungsang Kim

|↓〉 =
|0,0〉
APS March Meeting
Los Angeles, CA, USA, March 7th, 2018
 Entangling Trapped Ion Qubits
r ~5 μm
↑ ↑
↓ ↓ δ

δ ~ 10 nm
dipole-dipole coupling eδ ~ 500
Debye

for full
entanglement
Cirac and Zoller (1995)
Mølmer & Sørensen (1999)
Native Ion Trap Operation: “Ising” gate Solano, de Matos Filho, Zagury (1999)
Tgate ~ 10−100 μs Milburn, Schneider, James (2000)
From the March Meeting talk of
F ~ 98% – 99.9% Jungsang Kim

APS March Meeting

Los Angeles, CA, USA, March 7th, 2018
Assemblying large number of qubits

More ways to apply fundamental logical operation directly on hardware 18 18

Platform comparison
#2-Q Gates Maximum
1 & 2 Qubit
QC Platform Qubit Type T1 / T2 (s)
Gate Error
Run in a # Qubits
Circuit Entangled

Trapped Ions Natural > 1010 / 103 < 10-5 / 10-3 56 20

Neutral Atoms Natural < 10-4 / 10-4 < 10-3 / 0.03 1 2

Photonic Circuits Engineered Photon Loss Probabilistic - -

Superconductors Engineered < 10-3 / 10-3 < 10-3 / 0.005 15 9

Silicon Spin Engineered 10-3/10-5 > 10-3/0.1 1 2

19 19
Nature’s qubit
#2-Q Gates Maximum
1 & 2 Qubit
QC Platform Qubit Type T1 / T2 (s)
Gate Error
Run in a # Qubits
Circuit Entangled

Trapped Ions Natural > 1010 / 103 < 10-5 / 10-3 56 20

Neutral Atoms Natural < 10-4 / 10-4 < 10-3 / 0.03 1 2

Photonic Circuits Engineered Photon Loss Probabilistic - -

Superconductors Engineered < 10-3 / 10-3 < 10-3 / 0.005 15 9

Silicon Spin Engineered 10-3/10-5 > 10-3/0.1 1 2

20 20
Time available to finish computation
#2-Q Gates Maximum
1 & 2 Qubit
QC Platform Qubit Type T1 / T2 (s)
Gate Error
Run in a # Qubits
Circuit Entangled

Trapped Ions Natural > 1010 / 103 < 10-5 / 10-3 56 20

Neutral Atoms Natural < 10-4 / 10-4 < 10-3 / 0.03 1 2

Photonic Circuits Engineered Photon Loss Probabilistic - -

Superconductors Engineered < 10-3 / 10-3 < 10-3 / 0.005 15 9

Silicon Spin Engineered 10-3/10-5 > 10-3/0.1 1 2

21 21
Noise rate
#2-Q Gates Maximum
1 & 2 Qubit
QC Platform Qubit Type T1 / T2 (s)
Gate Error
Run in a # Qubits
Circuit Entangled

Trapped Ions Natural > 1010 / 103 < 10-5 / 10-3 56 20

Neutral Atoms Natural < 10-4 / 10-4 < 10-3 / 0.03 1 2

Photonic Circuits Engineered Photon Loss Probabilistic - -

Superconductors Engineered < 10-3 / 10-3 < 10-3 / 0.005 15 9

Silicon Spin Engineered 10-3/10-5 > 10-3/0.1 1 2

22 22
Longest sequence of logical operation
#2-Q Gates Maximum
1 & 2 Qubit
QC Platform Qubit Type T1 / T2 (s)
Gate Error
Run in a # Qubits
Circuit Entangled

Trapped Ions Natural > 1010 / 103 < 10-5 / 10-3 56 20

Neutral Atoms Natural < 10-4 / 10-4 < 10-3 / 0.03 1 2

Photonic Circuits Engineered Photon Loss Probabilistic - -

Superconductors Engineered < 10-3 / 10-3 < 10-3 / 0.005 15 9

Silicon Spin Engineered 10-3/10-5 > 10-3/0.1 1 2

23 23
Largest input size used so far
#2-Q Gates Maximum
1 & 2 Qubit
QC Platform Qubit Type T1 / T2 (s)
Gate Error
Run in a # Qubits
Circuit Entangled

Trapped Ions Natural > 1010 / 103 < 10-5 / 10-3 56 20

Neutral Atoms Natural < 10-4 / 10-4 < 10-3 / 0.03 1 2

Photonic Circuits Engineered Photon Loss Probabilistic - -

Superconductors Engineered < 10-3 / 10-3 < 10-3 / 0.005 15 9

Silicon Spin Engineered 10-3/10-5 > 10-3/0.1 1 2

24 24
History of improvement for gate error

25 25
 Ions vs. Superconductors
P. Murali, et al., ISCA 2019 (arXiv:1905.11349)

IBMQ5 IBMQ14 IBMQ16 Rigetti Trapped Ions

1.00

0.75
Measured
Success 0.50
Rate
0.25
F F F F F
0.00
BV4 BV10 HS2 HS4 Toffoli Fredkin Or Peres QFT Adder

Benchmark Algorithm
F : Failure

26 26
 Scaling technology

Shuttling of ion chain: Modular shuttling between Modular photonic connections

tape+head model multiple zones between chips

27 27
The promise
1 BILLION YRS

CLASSICAL COMPUTER:
Factoring Problem 30 YEARS FROM NOW
1 MILLION YRS
(MOORE’S LAW)

CLASSICAL COMPUTER:
TIME TO SOLUTION

1,000 YRS
STATE OF THE ART
100 YRS Classical computers are
10 YRS a dead end
1 YR for certain classes
1 MO
of problems.
1 DAY

1 HR
QUANTUM COMPUTER
100 SEC

1 SEC

SIZE OF PROBLEM (BITS)

28 28
The promise
1 BILLION YRS

CLASSICAL COMPUTER:
Factoring Problem 30 YEARS FROM NOW
1 MILLION YRS
(MOORE’S LAW)

CLASSICAL COMPUTER:
TIME TO SOLUTION

1,000 YRS
STATE OF THE ART
100 YRS

10 YRS So called quantum

1 YR supremacy
1 MO

1 DAY

1 HR
QUANTUM COMPUTER
100 SEC

1 SEC

SIZE OF PROBLEM (BITS)

29 29
The roadmap (near term)

Algorithm Improvements
Factoring
Computational Power
Log(Performance)

g
Quantum Machine Learnin Nitrogenase

tions
Quantum Chemistry/Simula
Simple
Organics
H2 HeH O3
H2O

2015 2017 2019 2021 2023 2025

30 30
The roadmap (near term)

Algorithm Improvements
Factoring
Computational Power
Log(Performance)

g
Quantum Machine Learnin Nitrogenase

tions
Quantum Chemistry/Simula
Simple
Organics
H2 HeH O3
H2O

2015 2017 2019 2021 2023 2025

Published works 31 31
The roadmap (near term)

Algorithm Improvements
Factoring
Computational Power
Log(Performance)

g
Quantum Machine Learnin Nitrogenase

tions
Quantum Chemistry/Simula
Simple
Organics
H2 HeH O3
H2O

2015 2017 2019 2021 2023 2025

32 32
Practical challenges of building a QC

● Computation makes qubits (unit of quantum registers)

interact with each other
● Qubits are susceptible to noise coming from the surrounding
environment
● It is hard to filter out unwanted interaction while allowing
computationally meaningful interactions

Source: Preskill, 2018

33 33
Practical challenges of building a QC

GPU errors (Tiwari et. al 2015) Quantum processing unit (QPU) errors (Wright et. al 2019)

34 34
NISQ computing

● Noisy intermediate-scale quantum (NISQ)

computers = First generation quantum computers
● Name coined by John Preskill (2018)
● Amount of computation very limited due to high
failure rate
● Upper bound for quantum advantage is quadratic
when the structure of the problem is not known
● With knowledge about the structure it may
improve

35 35
Algorithmic framework on NISQ computer

● The computational problem is encoded in a

quantum-native format
● The goal is to construct a quantum state which
encodes the solution
● This construction is done variationally using a
parametrized quantum circuit (= a simple quantum
algorithm with lots of variables)
● For many problems, the solution is the ground
state of the system
● Classical stochastic optimization routines are used
to find the ground state
36 36
Domains of applications

● Simulation of physics/chemistry
● Hard optimization problems

37 37
Standard NISQ algorithms

● Simulation of physics/chemistry
○ Potential advantage is well understood
● Hard optimization problems
○ Potential advantage is still under investigation

38 38
NISQ algorithmic schemes

● Simulation of physics/chemistry
○ Variational quantum eigensolver algorithm (McClean et. al 2015)
● Hard optimization problems
○ Quantum approximate optimization algorithm (Farhi et. al 2014)

39 39
Quantum approximate optimization algorithm

● Define a measurable quantum quantity which encodes the solution to

the problem
● Design a tunable method to explore the quantum search space (aka run
a quantum algorithm)
● Keep tuning the method using a classical routine until you get a
reasonable solution

40 40
An example NISQ problem

● Compute the MAXCUT of the following graph

1 2

4 3

41 41
An example NISQ problem

● Compute the MAXCUT of the following graph

● General MAXCUT problem is

1 2 NP-complete
● Application: Pattern recognition,
electronic circuitry design, state
problems in statistical physics,
and combinatorial optimization
4 3

42 42
An example NISQ problem

● Compute the MAXCUT of the following graph

● Best known exact algorithm

1 2 needs exponential time
● Approximation algorithms are
used in practice

4 3

43 43
An example NISQ problem

● Compute the MAXCUT of the following graph

● MAXCUT = 4
1 2 ● Two solutions

4 3

44 44
An example NISQ problem

● Let’s label the two partitions as 0 and 1.

● 0000 => all vertices belong to partition #0.
● 1010 => the first and third nodes belong to partition
#1.
● There are 24 = 16 possible partitions.
● A search problem!

45 45
An example NISQ problem

● A classical exact algorithm searches the solution in

the integer space
● A weighted version of the problem means
searching in the real space
● A quantum search (=optimization algorithm)
algorithm looks for the solution in a complex
vector space which is huuuuuuge!

46 46
An example NISQ problem

● A classical exact algorithm searches the solution in

the integer space
● A weighted version of the problem means
searching in the real space
● A quantum search (=optimization algorithm)
algorithm looks for the solution in a complex
vector space which is huuuuuuge!

Good or bad?
47 47
The Grover speed up

● Thanks to unique quantum phenomenon -

destructive interference while exploring the search
tree
● Unstructured search may speed up quadratically
● Theoretical upper limit

48 48
Back to the example

● Translate the problem into quantum-friendly

language
● For any edge <j,k> betweek the nodes j and k,
define the following: 1 2

C<j,k> = ½(-σzj σzk + 1)

4 3

49 49
An example NISQ problem

● Translate the problem into quantum-friendly

language
● For any edge <j,k> betweek the nodes j and k,
define the following:

C<j,k> = ½(-σzj σzk + 1) ● σz =

● The lowest eigenvalue occurs
when the nodes j and k are both
in the same partition
● The highest eigenvalue occurs
when the nodes j and k are in
different partitions 50 50
Tl;dr: Still a hard problem

● C = ∑j,k C<j,k> gives you the CUT

● C is to be measured on a quantum system
● Maximizing C or minimizing -C gives you the
MAXCUT
● Minimizing -C = Sending the quantum system
encoding the problem to lowest energy state
● Finding the lowest energy of a system is NP-hard
in classical physics
● It is QMA-hard in quantum physics

51 51
An example NISQ problem

● For each edge, vary the probability of the

connected nodes being in different or same
partition
● Find a combination of these probabilities which
yields the lowest value for -C

52 52
An example NISQ problem

● Varying the probability of the connected nodes =

switching the quantum system from one state to
another This angle dictates the
probability
● Hopefully one of those states will be the
(approximate?) solution

1 2

53 53
An example NISQ problem

54 54
An example NISQ problem

Creates an initial superposition of

complete search space
55 55
An example NISQ problem

Corresponds to the edge

between node #1 and #2

56 56
An example NISQ problem

57 57
An example NISQ problem

58 58
An example NISQ problem

59 59
An example NISQ problem

60 60
An example NISQ problem

Checks symmetry

61 61
An example NISQ problem

● For each term in C, measure the qubit to

determine the probability of being in 1
● Variationally update γ and β to get to the lowest
energy
● Repetition of this layer can be viewed as
constrained optimization problem which yields
higher approximation ratio
● End result will be close to -4 but never beyond
that

62 62
Improvement by past causal cones
Past causal cones?

Prof. Guifré Vidal Prof. Glen Evenbly*

Evenbly-Vidal 2013 64
A “harder” problem

Shehab et. al 2019 65 65

A “harder” problem

Shehab et. al 2019 66 66

Improvements by noise reduction?

Shehab et. al 2019 67 67

Improvements by noise reduction?

● 40% increase in accurary

● 78% fewer measurements

Shehab et. al 2019 68 68

Solving practical problems?

● Practical problems may contain millions of nodes and edges

● If one node = one qubit, we need millions of qubits
● Millions of edges ≈ 1/million noise rate in each logical operation
● Theoretical upper limit on quantum advantage is quadratic speedup
● Special knowledge about the structure of the problem might help go
beyond that
● Loading classical input onto a quantum computer is not very efficient

69 69
When a problem is quantum-worthy?

Randomly chosen instances by Zhou, 2003

70 70
Informed hybrid approach?

● Practical NP-complete problems are often easy

● Hard to get the exact solution
● Think about a search problem with gradient descent where most of the
search landscape is smooth but the basin of the global minima is rugged
● Classical computer should handle most of it and delegate to quantum
when the solution landscape is close to random

71 71
How about machine learning?

Workshop on Theory of Deep Learning:

Where next? IAS, Princeton, NJ 2019 Workshop on Quantum Machine Learning, 2018
72 72
Rays of hope

● HHL algorithm with exponential speedup for systems of linear

equation (Harrow et. al, 2009)
● Quadratic speedup for SAT problems (Dunjko et. al, 2018)
● Quantum gradient descent (Kerenidis 2019)
● Quadratic speedup for semidefinite programming solvers (Brandão
et. al, 2019)
● Error reduction with past causal cones for non-convex quantum
optimization (Shehab et. al 2019)
● Variational search space reduction with chaos theory for non-convex
quantum optimization (Shehab et. al anytime soon at your nearest
arXiv)
73 73
Quantum supremacy
How it all started?

75 75
Formalization of the experiment

76 76
The Google result

77 77
The IBM response

78 78
What is the point?

● Amazing progress in the hardware

Google IBM 79 79
What is the point?

● Nice theories

80 80
Thank you and we are hiring!

81