Finding a good answer is different from proving it is best

An October 26 preprint compares classical optimization, tensor networks and simulated fixed-schedule quantum circuits on binary problems. Its main benchmark includes 180 instances from ten families with 14 to 48 variables.^[1]

The distinction is the output a solver must deliver: a good answer, an optimum or proof that no better answer exists. Moving known optima to different bitstrings changes some classical comparisons. The tested linear-ramp QAOA runs do not improve solution quality over their classical comparators.^[1]

The quantum-circuit results come from classical simulation, with approximate tensor-network methods in some larger cases. The findings apply to the selected problems, implementations and budgets. They establish neither a universal solver ranking nor a quantum-hardware speed comparison.^[1]

QubitWire analysis

Our assessment is that an optimization claim becomes clearer when it separates the quality of the returned answer from confidence about the best possible answer. Those objectives may require different computations. A useful comparison would state which one the user actually needs and how success is scored. Otherwise an apparently faster solver may simply be answering an easier question than the method used as its baseline.

A convincing follow-up would preserve the same instances, accuracy target and stopping rule across approaches. Readers should ask which costs arise from preparing the problem and which arise from checking or certifying the answer. The simulation can help investigate those costs within its model, but it should not become a hardware speed claim. The most useful conclusion is the one that survives an equal accounting of both the search and the evidence supplied with its output.