Computational Open Problems: Choose a Reliable Calculation
A computational open problem is a question that existing calculations cannot answer reliably at the needed scale or precision. It is not synonymous with a slow program. A useful project identifies whether the obstacle is representation, sampling, optimization, conditioning, verification or an uncontrolled physical approximation, then proposes a measurable improvement.
Required background. Convergence studies and error budgets supply the distinction between a running code and a controlled calculation.
Route by the failure mechanism
Section titled “Route by the failure mechanism”| Observed failure | Topic owner | Useful first target |
|---|---|---|
| Entanglement growth truncates dynamics | Tensor-network algorithms | Longer reliable time at fixed observable error |
| Signed sampling weights cancel | Monte Carlo sign problems | An unbiased cost improvement or a new sign-free subclass |
| Methods disagree on a physical conclusion | Many-body numerics | A converged cross-method discrepancy map |
| A quantum algorithm’s cost is opaque | Resource estimation | A reproducible, uncertainty-aware resource ledger |
| A learning model fails outside its training set | Machine learning matter | A physical diagnostic of distribution shift |
| Results cannot be exchanged or reproduced | Research tooling | A semantic round trip and independent observable check |
Established: algorithms have regime-dependent strengths and limitations. Active: error control in regimes where no exact large-system reference exists. Conjectural: an accurate phase diagram inferred from one unconverged ansatz or a generic cure inferred from a few small cases. The Hubbard comparison and MPS time-evolution review illustrate how a controlled comparison exposes productive questions. Benchmark, methods review.
Define the scientific output and stopping rule
Section titled “Define the scientific output and stopping rule”Specify one or two observables, their required tolerance and the largest affordable reference problem. A useful target is “reduce cost at fixed error in a dynamical correlation function,” not “make tensor networks faster.” Separate preprocessing, memory, parallel resources and repeated-sample costs.
Write an explicit stop condition: unresolved reweighting denominator, no stable extrapolation, failed independent conservation check or sensitivity larger than the proposed effect. Honest failure maps are scientifically useful and prevent optimizer or hardware gains from masking incorrect physics.
Entry project: map the boundary of trust
Section titled “Entry project: map the boundary of trust”Select a small model with an exact solver and one scalable approximation. Sweep a parameter that challenges the approximation, such as frustration, evolution time or temperature. Converge solver tolerance and representation size separately. Deliver an observable-error map, runtime and memory ledger, and an explanation of the first failure. A workstation and a bounded compute allocation suffice.
The baseline is the exact accessible instance, not a second run of the same approximation. Failure tests change initial conditions, symmetry sectors, boundary conditions and fit windows. Hold out a regime when tuning a proposed improvement. Research begins when a new estimator, representation or algorithm extends the trusted region and the mechanism predicts where the gain disappears.
Novelty and entry references
Section titled “Novelty and entry references”Search for the exact observable, model family and error criterion; similar algorithm names can hide different guarantees. Reproduce a published baseline before claiming improvement. Reviewed 11 October 2026; selected sources are 2015–2019 methodological anchors. The topic assessments own newer evidence and specific unresolved questions.
References
Section titled “References”- J. P. F. LeBlanc et al., “Solutions of the Two-Dimensional Hubbard Model: Benchmarks and Results from a Wide Range of Numerical Algorithms,” Physical Review X 5, 041041 (2015). Preprint.
- S. Paeckel et al., “Time-evolution methods for matrix-product states,” Annals of Physics 411, 167998 (2019). Preprint.