2
$\begingroup$

Combinatorial optimization is often mentioned as a potential application of quantum computers. One of the main paradigms here is to reduce combinatorial optimization to an Ising problem, which is mathematically known as QUBO (quadratic unconstrained binary optimization). A QUBO then can be solved by a variety of methods, like quantum annealing or variational algorithms.

There are well known obstacles to building and scaling quantum computers. But here I want to discuss a different potential problem with this approach. Assume you have a large and reliable enough quantum computer that can find approximate solutions to QUBO. Will that actually be useful for real world problems?

While this is not the technically a quantum computing question, I believe this is the right place to ask. For one, the entire business of reducing different combinatorial problems to QUBO seems to be motivated by quantum computing alone (and perhaps classical analog Ising solvers). This gives me the first red flag, and raises the question

  1. Is it ever a good idea to reformulate your combinatorial optimization as QUBO? (Except if you absolutely have to, because you work in QC?)

I have not found any refs where people with your standard digital computer would for some reason put their logistics or graph problem into a QUBO form. From my own limited experience, the way QUBO deals with constraints is super inefficient.

For instance, assume you're doing some logistics problem and need to parameterize a path between $n$ locations. One standard way to do this is by introducing boolean variables $x_{t,i}$, with $x_{t,i}=1$ meaning that you visit location number $i$ at time $t$. Multiple visits to the same location, and being at different places at the same time is not allowed. This can be enforced by constraints $\sum_i x_{t,i}=\sum_{t} x_{t,i}=1$. There is a standard way to incorporate these constraints into QUBO as penalty terms, but there is a problem.

The number of valid solutions is $n!\sim 2^{n\log n}$, while the number of possible QUBO configurations is $2^{n^2}$. So the fraction of feasible configurations in QUBO formulation is exponentially small. I believe that a typical QUBO landscape is trapping, and most local minima do not correspond to feasible solutions. Is there any evidence that this is not a dealbreaker?

Almost all benchmarks I saw were for unconstrained problems, like MaxCut. There, any solution is a feasible solution, and techniques like quantum annealing make sense to me. However, what about generic real-world problems with constraints?

  1. Are there any studies discussing the difficulty in finding feasible solutions for QUBO reformulations of large scale constrained problems?

If QUBO reformulation is indeed a poor choice for most constrained problems, perhaps there are real-world problems, which are unconstrained? Here my question is mostly about the commercial applications. Yes, QUBO can model spin glasses, but this is physics. Yes, MaxCut and other graph problems are theoretically useful, but in practice these graph problems can be solved very efficiently anyway.

  1. Are there industrially relevant problems that are QUBO-native, and that classical solvers struggle with?

Thanks!

$\endgroup$

3 Answers 3

3
$\begingroup$

There are several important classes of problems that are well-suited to QUBO, such as quadratic assignment problems, Max-Cut, and graph coloring, to name a few. For these types of problems, it is reasonable to expect that certain quantum approaches might offer an advantage over classical methods.

However, the largest and most widely studied class of problems, both in industry and in the mathematical theory of optimization, are linear problems with linear constraints. These problems have been extensively researched, and a rich and elegant theory underpins them. For instance, the Simplex algorithm, regarded as one of the most important theoretical and practical algorithms of the century, is designed for linear problems with linear constraints. This algorithm, along with its subsequent commercial refinements, is exceptionally hard to surpass in general.

Given the existence of Simplex-like algorithms, it is difficult to justify the use of heuristics, especially those that require conversion to QUBO. The only situation in which optimization specialists might opt for a heuristic is when the problem is extremely large and a feasible solution is needed quickly. And even then, this would be a custom-made heuristic that exploits the structure of the problem.

In my opinion, converting linear problems with constraints into QUBOs is a thing of the past and is mostly popular among physicists who are more focused on the quantum aspects than on the mathematical theory of optimization or practical considerations.

  1. The answer to your first question is no, unless the problem is naturally QUBO. A rigorous explanation of why this approach is problematic can be found in Lagrangian Duality in Quantum Optimization: Overcoming QUBO Limitations for Constrained Problems. If quantum computing is necessary, it is better to use Lagrangian duality theory (formulated by Von Neumann and Dantzig), which is simpler and more efficient.

  2. I’m not aware of any studies discussing the difficulties of solving problems after converting them to QUBO. Such a study likely wouldn’t make much sense in the optimization community. In the optimization community, QUBOs are not commonly used, and no one would convert a problem into QUBO unless it is naturally suited to that formalism. A partial answer to your question could also be found in the abovementioned study.

To broaden the scope of your question, it’s worth noting that quantum optimization doesn’t have to involve QUBO, Hamiltonians, or QAOA-like methods. In the distant future, we might adopt approaches similar to Oracle-based methods, such as running Simplex-like algorithms in superposition or implementing objective functions with constraints directly using quantum arithmetic circuits.

$\endgroup$
6
  • $\begingroup$ Thanks a lot for the well informed answer! A brief followup. (1) Does linear programming really dominate the industrial space? By that I mean are problems with nonlinear constraints or objectives relatively rare? (2) you mentioned a few problems that are QUBO-native, like quadratic assignment, maxcut and graph coloring. I know QC companies mentioned them all the time, but will the industry really care about these? Is there a need to solve any of these or similar problems beyond the capacity of current classical methods? $\endgroup$ Commented Sep 27 at 18:28
  • $\begingroup$ @NikitaNemkov Thanks and welcome. (1) At the expense of generalization, industrial space is dominated by Gurobi or custom solutions. Whatever Gurobi solves (linear, quadratic, integer, conti) is in demand by industries. (2) Depending on the industry, yes, and these are millions of dollar problems (e.g., global logistics for FedEx or Amazon). There is a significant need to scale up as there is a massive need for faster hardware. Just look for the demand for certain electronic circuits. It is not only AI that needs the compute. $\endgroup$
    – MonteNero
    Commented Sep 28 at 0:01
  • $\begingroup$ Heuristics are very much needed in optimization. Not all problem are linear problems. Why optimization community uses at the moment mostly easy and old methods is because that is much easier and simpler to do and explain for them that are not working with those problems. Innovativity and new approaches are very much needed, and more or less done in research. $\endgroup$ Commented Sep 29 at 16:55
  • $\begingroup$ @MonteNero No doubt that there are high-stake optimization problems in industry. My question is whether any of them are QUBO-native, i.e. are binary, quadratic, and have no or few constraints. MaxCut is more or less pure QUBO, but I couldn't really find any modern industrial applications. Quadratic assignment -- not sure where it is used, also it has quite non-trivial constraints (similar to logistics). Graph coloring -- also has constraints, and I'm not sure that current specialized graph algorithms are insufficient so that graph coloring is a bottleneck somewhere. $\endgroup$ Commented Sep 30 at 8:32
  • $\begingroup$ @NikitaNemkov, good points. Maxcut as all problems you listed has constraints too, but it just happens so that these constraints play nicely with QUBO formulation. $\endgroup$
    – MonteNero
    Commented Sep 30 at 13:45
1
$\begingroup$

I think this is a general problem we're struggling to deal with. We take a problem that exists as a MILP (like your one) which has a lot of rich structure that can be exploited by classical solvers with some success (nothing great as the problem is NP-complete).

We then throw away all the structure to turn the problem into a QUBO in hope that the potential speedups from a quantum computer out weighs the speedups we saved from exploiting the structure. Until we get big enough quantum computers that allow us to fully understand how these speed ups scales, we simply don't know.

Coincidently, some quantum algorithms have been made to address your problem of constraints being ignored and the space blowing up. I've got a slightly different one going on arxiv next week, I'll link to it as well once its up. and it deals with the milp you mentioned.

In theory you can convert your qubo instance into a max cut, and then work out how close to the ground state you have to be for a feasible solution. My guess is you're right and you will need to be within $\frac{1}{poly(n)}$ or worse $\frac{1}{exp(n)}$ within the ground state of the maxcut instance which probably is not attainable for the larger problems.

$\endgroup$
7
  • $\begingroup$ Thanks for the answer! This is exactly my intuition, too. By converting an MIP into QUBO you throw away tons of structure. While some people present this as a feature (notably, I've heard Peter Shor saying something along these lines about QAOA), I believe it's rather a bug, caused by the need to comply with the hardware restrictions. Please do share your work once it's ready. Do you think that incorporating constraints into ansatz or similar tricks will actually be good enough for real world-problems, or only for highly symmetric ones? $\endgroup$ Commented Sep 27 at 11:58
  • $\begingroup$ I think choosing a nice ansatz to remove symmetry is great for reducing the size of the problem. But then we'll only be looking at something like modest square root speed up over brute forcing approaches at best. Whether we can use the symmetry better classically is unclear. An example that really shocked me was in the qubo here ieeexplore.ieee.org/abstract/document/10582776. DWAVE outperformed Gurobi on solving the qubo, but Gurobi didn't blink when given the milp. Comparing like for like is easy and makes these algorithms look good, but we need to be comparing best for best. $\endgroup$ Commented Sep 27 at 12:47
  • $\begingroup$ Totally on the same page, here. If I give the problem in QUBO form to a classic solver, it struggles as well. but the original mip is usually solved quite efficiently. $\endgroup$ Commented Sep 27 at 14:31
  • $\begingroup$ I just realized what you proposed in the post. Showing that merely finding feasible solutions in QUBO formulation is NP-complete would be quite a result. Do you know any papers discussing this? $\endgroup$ Commented Sep 27 at 14:33
  • 1
    $\begingroup$ It's probably worth noting that QUBO is no more unnatural to MILP. The whole motivation of turning say TSP into a MILP is that it's probably doable and I have a MILP solver I can then plug it into. If you want to solve a problem with an existing solver the first step has to be convert that problem into something you can feed into the solver. The question of whether MILP is better than QUBO due to the potential speed of solvers really applies to every way to formulate the problem and every solver that could be used on the problem. $\endgroup$ Commented Sep 27 at 15:30
-2
$\begingroup$

QUBO form is a quadratic programming with binary variables. That is nothing new in field of optimization. That form is very much needed in optimization problems.

In generally it is many times good when different problem formulations are possible. That enforces more easily not to think an optimization problem as easy path to solution with always same algorithms etc., but as an optimization problem that can have many different fruitful approach to solve them effectively.

Example when a quadratic programming with binary variables is best formulation for optimization problem: portfolio optimization.

$\endgroup$
4
  • $\begingroup$ Several objections. (1) most real-world problems are not unconstrained, converting them to QUBO is typically a very artificial step (2) Yes you can say that formulating the problem as QUBO and putting in on a quantum computer does not require any sophisticated problem-specific classical heuristics. My question is, will that actually work? (3) How is portfolio optimization a binary optimization problem? I thought it is continuous? $\endgroup$ Commented Sep 27 at 11:53
  • $\begingroup$ (1) You can represent those constraints in QUBO form. Same way in quadratic programming you have to represent constraints as well as that problem at bottom of it in 0s and 1s, since computer works with those. (2) Yes. For example from nature.com/articles/s41598-023-45392-w "Evidences allow concluding which are the best ways to solve real Portfolio Optimization problems by VQE on quantum devices, and confirm the possibility to solve them with higher efficiency, with respect to existing methods, as soon as the size of quantum hardware will be sufficiently high." (3) PO can be binary. $\endgroup$ Commented Sep 27 at 12:30
  • $\begingroup$ Could you please give a reference for binary portfolio optimisation? $\endgroup$ Commented Sep 27 at 13:25
  • $\begingroup$ nature.com/articles/s41598-023-45392-w $\endgroup$ Commented Sep 27 at 13:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.