A binary choice can be recorded as zero or one, or represented by a spin taking values minus one and plus one. In an Ising model, the variables take values of −1 or +1 instead. The terminology comes from different mathematical and physical traditions, but an unconstrained quadratic model in one form can be translated into the other without losing information about its solutions.
That equivalence is useful when comparing optimization software and quantum-computing workflows. It does not mean every real-world optimization problem is naturally a QUBO, or that rewriting a problem magically makes it easier. The representation is a choice about how to express an objective and which computational tools can consume it.
A choice written two ways
Let x and y represent two decisions, each either 0 or 1. Consider the QUBO objective
E(x, y) = −3x − 2y + 4xy. The linear terms reward selecting either decision, while the positive interaction term discourages selecting both together. Evaluating the four assignments gives 0 when neither is selected, −3 for x alone, −2 for y alone and −1 when both are selected. The minimum is therefore x = 1, y = 0.
Now introduce spins s and t, each either −1 or +1, with the substitution
x = (1 + s)/2, y = (1 + t)/2. The constant offset changes the displayed coefficients but not which assignment minimizes the energy.
Substituting and collecting terms gives
E(s, t) = −1.5 − 0.5s + 0t + 1st. The constant offset changes the displayed coefficients but not which assignment minimizes the energy. The mapping x = 1, y = 0 corresponds to s = +1, t = −1. Both forms represent the same four candidate decisions and the same optimum.
The spin convention is not universal: some systems use x = (1−s)/2 instead. That is equally valid when applied consistently. Reading a solver's conventions matters more than memorizing one sign choice.
Why the Ising language appears in quantum computing
Ising models emerged from physics as simplified descriptions of interacting spins. Their quadratic energy functions also make them a natural representation of some discrete optimization problems. Quantum annealing systems and other quantum-optimization techniques often describe targets using Ising-style Hamiltonians, while software libraries may offer a QUBO interface for users who think in binary variables.
Classical optimization tools can solve both representations too. The mathematics of the model is independent of the hardware used to search for a low-energy assignment. A strong classical reference is therefore an essential part of evaluating quantum approaches.
Compare feasible solutions in the original problem, not only the numerical energies printed by different APIs.
Constraints are the difficult part
Real problems often include hard rules that are not naturally present in an unconstrained quadratic objective. For example, “choose exactly one facility” can be encouraged by adding a squared penalty such as P(x + y − 1)². Because binary variables satisfy x² = x, this particular expression reduces to a quadratic function.
But assigning a penalty does not automatically guarantee that invalid answers disappear. If the penalty is too weak, violating a rule may be rewarded overall. If coefficients have extreme scales, some algorithms and devices may struggle with representation or numerical behavior. And certain constraints require auxiliary variables or substantial reformulation to fit a quadratic binary form.
That modeling cost is often more consequential than the cosmetic conversion between QUBO and Ising. Before comparing solvers, verify that the two formulations preserve the intended objective and identify feasible solutions consistently.
What a fair comparison looks like
If a quantum system consumes an Ising model and a classical system consumes QUBO, translate the exact same objective and account for constant energy offsets. Compare feasible solutions in the original problem, not only the numerical energies printed by different APIs. Record coefficient precision, runtime, sampling budget and any preprocessing or penalty tuning.
Both representations are useful because they make discrete optimization accessible to different algorithms and devices. Their value lies in expressing the problem faithfully, not in implying a particular technology will automatically outperform another.