Estimation versus 100% proof in quantum computation
Almost every number read from a present-day quantum processor is a statistical estimate. That is not a confession of weakness — it is the correct language of experimental quantum physics. Confusion begins when marketing upgrades estimates into absolute certainty.
Thesis. “100% proof” and “hardware estimate” answer different questions. Proof belongs to mathematics, verified classical reproduction, or fault-tolerant logical certainty. Hardware outputs under Born-rule sampling belong to estimation theory — with stronger or weaker trust protocols stacked on top. Calling an estimate a proof is a category error.
1. Two different questions
When someone asks whether a quantum result is “real,” they usually mix two demands:
- Certainty of the claim type — Is this a theorem? A classically reproduced number? A logical qubit that cannot fail under the stated error model?
- Quality of an empirical measurement — Given noisy gates and finite shots, how well do we know an observable, and under what assumptions?
The first question can, in principle, admit proof. The second question, on today’s machines, admits estimators, confidence intervals, hypothesis tests, and validation protocols — never metaphysical 100% certainty about a single shot string.
2. What quantum hardware actually returns
A gate-model device prepares a state and measures. Each shot yields a bitstring drawn from a probability distribution induced by the circuit and the device noise. From \(N\) shots one constructs empirical frequencies \(\hat{p}(x)\). Almost every scientific figure of merit is then a function of those frequencies: an expectation value \(\langle O \rangle\), a correlator, a fidelity proxy, a gap between two arms.
In statistics this object is an estimator: a random variable that concentrates around a target value as \(N\) grows (under regularity conditions). Shot noise alone already forbids “the number is exact.” Device noise makes the target itself a noisy channel’s output unless one corrects or mitigates.
So the blunt fact: raw counts are already estimates of ideal (or of physical) probabilities. Mitigation does not leave the category of estimation; it changes which target the estimator aims at and how expensive the shots become.
3. What “100% proof” would have to mean
The phrase is popular and almost never defined. In a scientific register, candidates include:
3.1 Mathematical proof
An algorithm is proven correct relative to an ideal circuit model (e.g. Shor factors under standard number-theoretic assumptions about the quantum circuit semantics). That proves the algorithm, not that a particular noisy chip executed it flawlessly.
3.2 Classical verification of the instance
For small enough circuits, a classical computer can compute the exact ideal distribution or observable and compare. Agreement within error bars is strong empirical confirmation for that instance — still statistical if the quantum side is sampled, but the reference is exact.
3.3 Fault-tolerant logical certainty
With quantum error correction below threshold and enough concatenation or distance, one can bound the probability that a logical outcome is wrong to be astronomically small. That is the closest experimental physics comes to “for practical purposes, proven.” Present utility-scale demos are not there yet for large algorithms.
3.4 Complexity-theoretic hardness + certificates
Some advantage claims argue that producing the observed sample (or passing a certificate) is intractable for known classical methods, then attach a process or fidelity bound. This can be a rigorous argument under stated assumptions. It is still not a direct classical check of the “answer,” because the answer is defined to be hard to check.
If none of (3.1)–(3.4) applies in full, honesty requires speaking of estimates under a stated protocol, not of absolute proof.
4. The trust ladder for hardware claims
Claims differ less by whether they use the word “estimate” and more by where they sit on a trust ladder:
Moving up the ladder improves warrant. It does not transmute an estimator into a Euclidean theorem about the physical universe.
5. Why “estimate” sounds weak — and why that instinct is half right
In ordinary language, “estimate” suggests vagueness: a guess. In measurement science, an estimator is a precise object: a formula, a bias analysis, a variance, sometimes a theorem of asymptotic unbiasedness.
The half-right instinct is this: you should refuse claims that hide the estimator. If a vendor says “error-free” while meaning “mitigated expectation value within user tolerance,” the noun has been upgraded without new ontology. If a researcher says “estimate” and shows the protocol, that is stronger science, not weaker branding.
Heuristic mitigation (many forms of zero-noise extrapolation) can be accurate in practice and still carry unknown bias. Unbiased methods aim at zero systematic bias under a noise model; they remain vulnerable to model misspecification and always carry statistical error. Neither yields 100% proof that “the noiseless universe said X” on a single finite experiment.
6. Who accepts an estimate — and on what grounds?
Acceptance is social and methodological, not mystical:
- Peer methods — Is the estimator standard or published with proofs?
- Calibration honesty — Was the noise model characterized on the same device epoch?
- Controls — Do null experiments behave as predicted?
- Reproducibility artifacts — Circuits, job IDs, seeds, code, frozen inputs.
- Scope discipline — Does the paper claim what the protocol can support?
A community can rationally accept a result at L3–L5 for publication and engineering decisions while still denying that it is “100% proven” in the fault-tolerant or mathematical sense. Those are compatible attitudes.
7. Estimation is the native language of NISQ; proof is the destination of FTQC
Noisy intermediate-scale quantum (NISQ) computation is an experimental science of channels and samples. Fault-tolerant quantum computation aims to push logical error rates so low that, for practical purposes, a logical bitstring can be treated like a classical computer’s output bit — with a failure probability smaller than, say, cosmic-ray upsets on a classical DRAM.
Until that regime is routine for the problem class at hand, intellectual hygiene requires: state the estimator, state the assumptions, state the controls, state what remains unproven.
8. Closing definition
Estimation answers: given this device, this circuit family, this finite sample, and these assumptions, what value do we infer for an observable, and how uncertain are we?
100% proof answers: under an explicit formal model (mathematics, verified classical reproduction, or fault-tolerant error bounds), the claimed statement cannot fail except by denying the model.
Conflating the two produces either false comfort or false cynicism. Distinguishing them is how the field matures.
A. How Quantum Polycontextural Computing reports results without pretending to absolute proof
A.1 QPC outputs are estimators — and we say so
QPC architecture metrics extracted from IBM (or other) hardware — junction correlations, raw junction gaps, ICC-style separability scores, retention ratios, ranking statistics from shot histograms — are finite-shot estimators. They live on the same epistemological shelf as any other NISQ observable. Anyone who demanded “100% proven wavefunction truth” from a 8192-shot Sampler run would be asking for a category that the experiment does not contain.
A.2 What QPC proves in the strong sense
QPC’s strongest non-estimate commitments are typically classical or definitional:
- Frozen inputs — a published conflict table or desk schema that anyone can re-score classically.
- Predeclared geometry and controls — coupled vs
no_crossvs separable / merge protocols locked before the decisive run. - Pass/fail scorecards — numerical bars fixed in a preregistration so the experiment can fail.
- Audit trail — public job IDs and result bundles tying claims to specific Runtime executions.
Those elements are closer to proof of experimental integrity than to proof that an ideal noiseless circuit produced a unique mathematical constant.
A.3 What QPC estimates — and why that is still a strong architecture claim
On the ADMET Lead Desk Stage A→C ladder, for example, the Marrakesh same-day closure reports raw gap, retention, and separable discrimination as estimated metrics from hardware counts. The scientific force of the claim is architectural discrimination under controls:
- Classical methods disagree on the same frozen table (checkable without quantum hardware).
- The coupled polycontextural job realizes joint structure relative to controls that remove cross-context coupling.
- Primary architecture numbers are computed on raw counts; resilience / mitigation, when used, is labeled secondary and does not silently replace the discriminator.
That package sits roughly at L1–L3 on the trust ladder above, with a deliberate refusal to climb to L4 by swapping mitigated numbers into the headline architecture claim. The point of the refusal is honesty: architecture evidence should not be laundered through a mitigation story.
A.4 How this differs from “error-free mitigated estimates”
Error-mitigation stacks (including commercial suites such as unbiased quasi-probabilistic methods) aim to estimate noiseless circuit observables. That is a legitimate L4 research program. QPC’s public architecture WOW aims to estimate whether a polycontextural coupling geometry leaves a measurable joint-structure signature versus declared controls.
Different target. Different success criterion. Same underlying truth: both are estimates. QPC’s discipline is to keep the target aligned with architecture, keep the classical parts classically checkable, and keep the language from floating up to “100% proven physics” when the object is a scorecard on shots.
A.5 A sentence we endorse
QPC results on hardware are high-quality, controlled estimates of architecture metrics, backed by frozen classical conflict evidence and public job IDs — not a claim of fault-tolerant certainty, and not a claim that mitigation has erased noise. That is the standard we ask others to use on us, and the standard we use when reading advantage headlines.
A.6 Practical reading guide for QPC pages
- If a number comes from Sampler / measurement histograms → treat as estimator; look for shots, backend, job ID, raw vs mitigated label.
- If a statement is about a frozen table, classical ranking disagreement, or preregistered bar → treat as checkable classical fact or procedural fact.
- If a statement says “WOW closed” → it means the preregistered scorecard passed, not that quantum mechanics was re-derived from scratch.
Selected further reading
- Temme, Bravyi, Gambetta — error mitigation by probabilistic error cancellation (foundational unbiased EM).
- Qedma et al., Reliable high-accuracy error mitigation for utility-scale quantum circuits — arXiv:2508.10997 (QESEM).
- IBM Quantum Advantage Tracker and companion “trusted quantum advantage” papers (certificates when classical check fails).
- QPC ADMET WOW article — architecture scorecard with public job IDs.
- QPC resilience layer docs — raw primary metrics; mitigation secondary.