Resource-Estimation Frontiers
Resource estimation asks what a specified quantum calculation would cost on a specified fault-tolerant architecture. The frontier is the reliability of the bridge from logical circuit to delivered answer: approximation error, state preparation, routing, factories, decoding, retries and uncertain hardware parameters all matter. An estimate is a conditional scientific model, not a construction schedule.
Required background. Resource estimation supplies logical-to-physical accounting; resource-estimation tools explains implementation choices.
Begin with the scientific tolerance
Section titled “Begin with the scientific tolerance”For an energy calculation, separate representation error, algorithmic approximation, sampling or phase-estimation uncertainty and logical failure. Combining them requires stating whether each is a deterministic bound, a standard deviation or a failure probability. Do not add a confidence interval to a systematic bias as if they were the same quantity.
A rough union-bound allocation is
but operation classes may need different code distances. Factory output errors and data-memory errors should receive explicit allocations. The accepted cost is affected by retries: for independent attempts of cost and success probability , expected cost is ; dependence and failure detection can change this model.
What changing estimates teach
Section titled “What changing estimates teach”Gidney and Ekerå’s RSA-2048 analysis connected arithmetic to physical assumptions including code cycles and reaction times. Gidney’s 2025 preprint proposes a much smaller qubit count using further architectural and algorithmic choices. These are conditional resource estimates, not demonstrations of factoring that modulus. Their difference motivates sensitivity analysis rather than a forecast based on a single headline number. 2021 analysis, 2025 preprint.
Active: transparent compiler pipelines, experimentally constrained parameter distributions, and workload-specific comparisons of codes and resource factories. Conjectural: extrapolations that assume independently achieved best-case component values can coexist in one machine.
Open problems
Section titled “Open problems”Which uncertain parameter actually changes the preferred architecture? How should correlated rare faults enter long-duration forecasts? Can a lower logical gate count require more elapsed time because it increases adaptive depth? Where does the optimal error-budget allocation shift when classical preprocessing or state overlap is included?
Graduate project: reproduce and perturb one estimate
Section titled “Graduate project: reproduce and perturb one estimate”Reproduce a small version of a published arithmetic or chemistry resource estimate from its stated formulas. Make physical error, cycle time, reaction time, factory throughput and acceptance explicit inputs. Deliver a machine-readable assumptions file, a qubit–time Pareto plot and a sensitivity ranking. A laptop spreadsheet or short program is adequate; the hard part is consistent accounting.
The baseline is the published operating point, with discrepancies traced to conventions. Failure tests set factory throughput to infinity, remove routing, and tighten the target failure probability by an order of magnitude. Each should change only the appropriate terms. Graduate to research by replacing one poorly constrained assumption with a measured distribution or a better compilation method, then comparing at equal task accuracy across more than one workload.
Entry points and source horizon
Section titled “Entry points and source horizon”Begin with the 2021 worked analysis before examining the newer proposal. Literature search date: 11 October 2026. Evidence horizon: the cited 2021 publication and 2025 preprint; no claim is made that these are the latest or best estimates for every architecture.