Choosing good problems for quantum annealing

TECHNICAL REPORT

2020-09-06

Overview CONTACT

Corporate Headquarters
Many complex optimization problems arise in research and industry.
3033 Beta Ave
Solving these problems helps researchers make new scientific discov- Burnaby, BC V5G 4M9
eries, build better materials, and synthesize new medicine; and helps Canada
businesses make better decisions, reduce costs, increase production, Tel. 604-630-1428
and design better products. Quantum annealing is a heuristic method US Office
for solving many of the optimization problems that appear in a range of 2650 E Bayshore Rd
disciplines. This document provides the background to help you choose Palo Alto, CA 94303
the right problems for quantum annealing.
Email: [email protected]

www.dwavesys.com

14-1046A-A
D-Wave Technical Report Series
Choosing good problems for quantum annealing i

Notice and Disclaimer

D-Wave Systems Inc. (“D-Wave”) reserves its intellectual property rights in and to this doc-
ument, any documents referenced herein, and its proprietary technology, including copyright,
trademark rights, industrial design rights, and patent rights. D-Wave trademarks used herein
include D-WAVE®, Leap™ quantum cloud service, Ocean™, Advantage™ quantum system,
D-Wave 2000Q™, D-Wave 2X™, and the D-Wave logo (the “D-Wave Marks”). Other marks used in
this document are the property of their respective owners. D-Wave does not grant any license, assign-
ment, or other grant of interest in or to the copyright of this document or any referenced documents,
the D-Wave Marks, any other marks used in this document, or any other intellectual property rights
used or referred to herein, except as D-Wave may expressly provide in a written agreement.

Copyright © D-Wave Systems Inc.

Choosing good problems for quantum annealing ii

Summary
Many complex optimization problems arise in research and industry. Solving these prob-
lems helps researchers make new scientific discoveries, build better materials, and syn-
thesize new medicine; and helps businesses make better decisions, reduce costs, increase
production, and design better products. Quantum annealing is a heuristic method for solv-
ing many of the optimization problems that appear in a range of disciplines. This document
provides the background to help you choose the right problems for quantum annealing.

Copyright © D-Wave Systems Inc.

Choosing good problems for quantum annealing iii

Contents
1 Introduction 1

2 Complexity 1

3 Optimization 2
3.1 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
3.2 Constrained problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
3.3 Unconstrained problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
3.4 Variable types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

4 Choosing the right problem 4

5 Future 5

Copyright © D-Wave Systems Inc.

Choosing good problems for quantum annealing 1

1 Introduction
Many complex optimization problems arise in research and industry. Solving these prob-
lems helps researchers make new scientific discoveries, build better materials, and synthe-
size new medicine; it helps businesses make better decisions, reduce costs, increase pro-
duction, and design better products. Quantum annealing is a heuristic method for solving
many of the optimization problems that appear in a range of disciplines.
As you begin to consider how quantum annealing may help solve the complex problems
that you face in your research efforts or in your business, it is essential that you first under-
stand the different types of complexity and optimization problems.
Every successful application requires a careful understanding of the requirements and pa-
rameters. The chain of questions that you consider for your application will guide you
towards the mathematical formulation of your problem. These may be questions such as
“what is the business problem that the application is trying to solve?” Or, “what is the
research question?” Even at this stage, there are still many questions to address.
Does an application require a data processing pipeline or predictive analysis? Or does it
need solving an optimization problem?
In this document, we will learn that understanding complexity, optimization problems,
and variable and constraint types can help you choose the right problems for quantum
annealing.

2 Complexity
Intuitively, complexity describes how difficult a problem is and what might make it chal-
lenging to solve.
There are different types of complexity. For example, suppose that you want to make a
change to your company’s website. Since users interact with your website regularly, you
must consider how they will react to the change. If there are many users involved, they can
be treated as a statistical population. The complexity of this problem is statistical inference
and careful experimental design. In a different scenario, the complexity may be the fact
that there is a human in the loop.
Alternatively, suppose your website contains a large number of records that must be searched
efficiently. Or, you might want to build a predictive model using this large amount of data.
In both examples, the problem is complicated because it involves big data.
Each example deals with different types of complexity. Some problems are complex be-
cause finding a good model is difficult, predicting the behavior of individuals is hard,
sorting through a large amount of data is costly, or defining good experimental design
is challenging.
As we begin to explore candidate problems for quantum annealing, we look for problems
that consist of a hard optimization problem. Returning to an earlier example, after you
sort through big data, and picked a predictive model, you may still need to optimize an
objective. For example, you sort through billions of records, and you understand the inter-

Copyright © D-Wave Systems Inc.

Choosing good problems for quantum annealing 2

action of each item with the entire organization. You have learned how each item impacts
the overall costs for the organization. Now you need to optimize those costs subject to
some constraints. For example, you may need to minimize costs while maintaining sales
revenue.
Many optimization problems are NP-hard, which means no polynomial algorithm exist for
solving them. As you add variables to these problems, the solve time scales exponentially.
These types of problems are good candidates for quantum annealing. Let’s dig deeper.

3 Optimization
Optimization problems arise in many areas of science and industry. For many companies
(like retailers and distributors), improving business is equivalent to optimizing processes,
cost, productivity, and quality.
Many optimization problems are not considered hard, in a mathematical sense, but only
extensive, which means the optimization problem contains a huge number of entities and
variables — even billions and trillions! On the other hand, some optimization problems
are much more time-consuming to solve even though they involve a few hundred to a few
thousand variables. These small but challenging optimization problems, such as personnel
scheduling and the famous traveling salesperson problem, fall into the class of NP-hard
problems.
Optimization problems that have discrete or binary variables are often better suited for
quantum annealing than those with continuous variables (because conversion to discrete
is required). When first learning about quantum annealing, it may be easier, to begin with,
the former type of optimization problem. However, it is worth noting that there are also
examples of continuous-valued problems that have seen success with quantum annealing.
For example, QBoost is a boosting algorithm that assigns binary weights to a subset of
weak classifiers [6].

3.1 Objectives
Sometimes an optimization problem is written as a mathematical expression (objective)
that needs to be minimized (or maximized). This is an excellent starting point to enable
using many optimization packages. For example, to run an optimization problem on D-
Wave’s quantum annealing hardware, we can take that mathematical expression and map
it to an Ising Hamiltonian.
Objective functions can have different forms depending on the context and the type of
problems.
Even simple linear problems with discrete variables can quickly become complicated if we
have many competing constraints or interactions between many variables. For example,
a constrained integer linear problem like many scheduling problems has a simple linear
objective; however, the addition of constraints — such as contiguous shift, limited shift
duration, and so on — make it a complex optimization problem. More on the constraints
below.

Copyright © D-Wave Systems Inc.

Choosing good problems for quantum annealing 3

The most suitable objective function for quantum annealing consists of a quadratic objec-
tive or one that includes only pairwise interactions of binary variables (QUBO) because
that is the most similar objective to the Ising Hamiltonian of a quantum annealing proces-
sor. An example of a natural QUBO problem is the maximum-cut problem. This problem
occurs in many applications involving social networks. For example, the classification of
GitHub users into two categories, machine learning developers or web developers can be
efficiently formulated as a QUBO problem.

3.2 Constrained problems

Many optimization problems are complex due to constraints. Imagine you want to mini-
mize the operation cost of an organization. One method is to express the total cost as the
sum of individual costs:

C= ∑ xi ci (1)
i

Where xi is the number of units (integer), and ci is the cost per unit of the item or process i.
Easy, right? The objective above is a linear function where all variables and unit costs are
positive. The minimum is where all xi = 0 and the total cost is zero. But wait! That’s not
a practical solution. To formulate an actual problem, we need to have constraints. Usually,
these constraints add complexity.
There are many types of constraints that appear in optimization problems. For example,
we might use a binary variable to express yes/no to the existence of a feature. We might
also have many different possibilities, out of which we only want to choose one. In such
a case, not only do we want the variable to behave as binary, but we also want a group of
variables to respect the constraint that only one option is chosen. If xi is a binary variable
for feature i, then the following constraint ensures that the multiple variables in the set C
are only equal to one, one at a time. For example, in graph coloring, we have a set of binary
variables that correspond to different colors and we want each node/variable to choose
one color only.

∑ xi = 1 (2)
i ∈C

This example demonstrates a linear constraint, which can be built into the optimization
model using techniques shown in the ”Learn to formulate problems” guide [1].
There are many different types of constraints that can be modeled as a quadratic optimiza-
tion problem so that they are suitable for quantum annealing.

3.3 Unconstrained problems

There are other problems that, despite having no constraints, are complex.
In the following optimization problem, for example,

Copyright © D-Wave Systems Inc.

Choosing good problems for quantum annealing 4

f = ∑ ai xi + ∑ wi,j xi x j (3)
i i,j

the goal is to find a set of values ( x1 , x2 , ..., xn ) such that the function f is minimized (or
maximized). If the variables xi are integer or binary, this becomes a hard optimization prob-
lem.
While we can encode constraints as a set of linear and quadratic terms in the objective
above, this is not always the most efficient approach.
One example of an unconstrained quadratic problem is the weighted maximum cut for
the classification of nodes into two groups. For example, suppose that each variable xi
represents a user or an organization. An edge is defined between two nodes if there is a
relationship between the two users or organizations. The weight of the edge represents the
strength and type of relationship. Solving the optimization problem above will determine
whether variables take the value of 1 (belonging to one group) or the value of -1 (belonging
to the opposite group). One real-world example of this is the detection of the structural
imbalance of the terrorist’s network. Using QUBO formulation, we can assign the values
of 0 and 1 to two groups of terrorists. For more information please see [5].

3.4 Variable types

Optimization and combinatorial problems with discrete and binary variables are hard to
solve. These problems map to quadratic unconstrained binary optimization (QUBO) or
other discrete optimization problems that are well suited for quantum annealing and hy-
brid solvers.
Integer variables require a logarithmic number of binary variables, and any integer prob-
lem can be represented efficiently as a binary problem.
Another binary representation of integer variables, which is called one-hot encoding, uses
as many binary variables as there are integer values. For example, if ten categories are rep-
resenting different digits, we can use ten binary variables xi , where the value of xi is equal
to one, if the category i is chosen. This will require a constraint on the value of the variables.
Despite the use of the constraint, problems with categorical variables are considered good
candidates for discrete optimization with quantum annealing because they are NP-hard.
Optimization problems involving only continuous variables may be best handled by linear
and quadratic programming packages that are available freely, or by other commercial
optimization software.

4 Choosing the right problem

Putting all we learnt together, the best problems for getting started with quantum anneal-
ing are the ones in which:

• There is a complex optimization involved

Copyright © D-Wave Systems Inc.

Choosing good problems for quantum annealing 5

• The variables are discrete

• The optimization is quadratic

– The objective function is quadratic and / or

– Constraints can be efficiently represented as a quadratic objective

5 Future
Despite the rules set out in this document, other types of optimization problems may also
be suitable for quantum annealing. For example, quantum annealing can be used for ma-
chine learning applications [4], and material simulation [2, 3].
Furthermore, there may be efficient hybrid algorithms for specific classes of problems that
may not, at first glance, appear to be the best choice for quantum annealing. For such prob-
lems, a hybrid quantum-classical approach may be effective. For example, a class of con-
straint binary optimization problems can be efficiently solved by using a hybrid solver that
works with the original constraints without explicitly relaxing them as a penalty objective.

References
[1] Learn to formulate problems, https://docs.dwavesys.com/docs/latest/doc_bqm_
formulation.html (2020).

[2] Harris, R., et al. ”Phase transitions in a programmable quantum spin glass simulator.”
Science 361.6398 (2018): 162-165.

[3] King, Andrew D., et al. ”Observation of topological phenomena in a programmable

lattice of 1,800 qubits.” Nature 560.7719 (2018): 456-460.

[4] Amin, Mohammad H., et al. ”Quantum boltzmann machine.” Physical Review X 8.2
(2018): 021050.

[5] Mapping Militant Organizations, Stanford University, last modified February 28, 2016,
http://web.stanford.edu/group/mappingmilitants/cgi-bin/.

[6] Neven, H., Denchev, V.S., Rose, G. and Macready, W.G., 2012. QBoost: Large Scale
Classifier Training with Adiabatic Quantum Optimization. In ACML (pp. 333-348).

Copyright © D-Wave Systems Inc.