m@berryhill: ~/berryhill.dev/posts/near-optimal-is-not-a-deployment-claim.md
~ homeposts/about.md
m@berryhill in ~/posts$ cat near-optimal-is-not-a-deployment-claim.md
---
title:  Near-Optimal Is Not a Deployment Claim
date:   2026-08-14
topic:  quantum computing
read:   13 min
words:  2,805
slug:   near-optimal-is-not-a-deployment-claim
views:  live post
tags:   [quantum computing, AI strategy, technical due diligence, deployment readiness]
---
essay · long read

Near-Optimal Is Not a Deployment Claim

A six-part comparator card for reading quantum algorithm claims without turning scoped theory or simulation into deployment evidence.

table of contents
  1. Start with the actual result
  2. The Six-Part Comparator Card
  3. A claim map prevents accidental promotion
  4. Deployment begins where the evidence ends
  5. What to ask before this enters a roadmap
  6. Sources

“Near-optimal” sounds complete. It is not.

Put the adjective into an executive summary and the rest of the claim often disappears: near-optimal at what task, against which comparator, with what access, at what precision, with what success requirement, measured in which resource?

Remove those fields and a scoped mathematical result starts to sound like a verdict on the whole system. Query complexity becomes practical speed. Numerical convergence becomes hardware evidence. A preprint becomes a roadmap.

That is not a problem with mathematical optimization. It is a problem with how technical claims travel.

Near-optimal is a scoped relationship between an objective and a comparator under stated assumptions. It is not a portable deployment verdict. The adjective is not the claim; the comparator and assumptions are part of the claim, and deployment evaluation begins where that evidence ends.

A useful current example is DEFEAT, a quantum algorithm introduced by Po-Wei Huang, Bence Bakó, and Bálint Koczor in the August 12, 2026 v1 preprint “Eigenstate Preparation Through Near-Optimal Eigenprobability Filtering”. The result is interesting precisely because its scope can be stated. It is less useful when “near-optimal” is detached from that scope and asked to carry a deployment claim the paper does not make.

Here is the practical reading rule: the adjective is not the claim. The comparator and assumptions are part of the claim.

Start with the actual result

The authors introduce DEFEAT to identify and filter the eigenstate with the largest overlap with a supplied initial state, without prior knowledge of that target eigenvalue. Their setup treats the dominant state as the one with the largest initial overlap and assumes that dominant overlap is unique.

Under the paper’s stated access and gap assumptions, the authors prove that the filtering step’s query complexity is optimal up to logarithmic factors. Their discussion describes that filtering step as near-optimal within the block-encoding framework.

Every bolded qualifier matters:

  • Filtering step is not the same claim as an optimal end-to-end application.
  • Query complexity is not wall-clock runtime, physical gate count, or operating cost.
  • Under stated assumptions means the access model is part of the result, not an implementation detail to add later.
  • Up to logarithmic factors preserves the actual relationship to the lower bound.

The paper also presents asymptotic comparisons with phase-estimation variants and author-reported numerical tests on a five-qubit random-field Heisenberg model. Those simulations test approximation behavior. They are not runs on quantum hardware.

That combination—a scoped theorem, asymptotic comparisons, and numerical convergence—is meaningful evidence. It is simply not deployment evidence.

The Six-Part Comparator Card

The card below is my operator synthesis, not a framework proposed or validated by the DEFEAT authors. I use it to keep an optimization claim intact while it moves from a paper into a strategy conversation.

The Six-Part Comparator CardTwo columns of three numbered checks preserve the objective, comparator class, access model, precision, success probability, and resource measure attached to an optimization adjective.THE SIX-PART COMPARATOR CARDThe adjective is not the claim.1ObjectiveWhat exact task is optimized?2Comparator classNear which bound or alternative?3Access modelWhich inputs and oracles exist?4PrecisionWhich error guarantee is used?5Success probabilityWhat repetition cost is included?6Resource measureWhich cost is actually bounded?Remove one field and you have changed the engineering claim.
If any field disappears, the claim that reaches the roadmap is not the claim the source established.

1. Objective

What exact task is being optimized?

For DEFEAT, the relevant objective is dominant-eigenstate filtering: approximately construct a projector onto the unique eigenstate with the largest overlap with the supplied initial state, or the equivalent reflector.

That is narrower than “optimal eigenstate preparation,” and much narrower than “an optimal quantum algorithm.” The paper separately analyzes how the filter can be used for downstream state preparation. Do not merge the subproblem and the end-to-end task.

2. Comparator class

Near which lower bound or alternative, for which subproblem?

The near-optimality statement concerns the filtering step’s query complexity relative to the paper’s lower-bound argument, with a logarithmic or polylogarithmic gap in the relevant framework. The paper’s comparisons with phase-estimation variants are useful, but they are separate algorithm comparisons. They do not redefine the lower bound.

One model-sensitive caveat is especially important: the authors also derive a lower bound for purity amplification under purified quantum query access. They explicitly note that this lower bound cannot be transferred directly to DEFEAT’s fuller, more structured access model. A matching end-to-end lower bound in that full model remains open.

3. Access model

What inputs, oracles, and prior bounds are assumed?

The paper assumes a state-preparation unitary; controlled forward and inverse Hamiltonian time evolution or a block-encoding of the Hamiltonian; lower bounds on the relevant spectral and amplitude gaps; and a suitable threshold. The authors note that loose lower bounds increase runtime.

This is not a footnote. Oracle construction and state preparation can dominate practical feasibility. An algorithm does not become assumption-free when those assumptions disappear from the summary.

4. Precision

What error notion and target accuracy does the result use?

The core filtering theorem is stated in terms of an epsilon-approximate block-encoding of a projector or reflector. Downstream state preparation uses a fidelity target. Those are different guarantees.

Before repeating a scaling claim, name the guarantee it scales against. “Accurate” is too vague to preserve the result.

5. Success probability

What success, confidence, or repetition requirement is attached?

The filtering theorem is parameterized by approximation error rather than packaged as one universal success-probability promise. Downstream tasks add fidelity and success requirements. The paper notes that restoring success-probability dependence adds logarithmic cost, and end-to-end state preparation uses amplitude amplification.

A summary that keeps the filter but drops amplification or repetition overhead has changed the engineering claim.

6. Resource measure

Which cost is being bounded?

The preprint analyzes resources such as queries to Hamiltonian evolution and state preparation, evolution time, and ancillary qubits, with asymptotic dependence on gaps and error.

Those are not interchangeable with:

  • wall-clock runtime,
  • logical gate count,
  • physical-qubit count,
  • error-correction overhead,
  • energy cost,
  • operational reliability.

If the theorem bounds queries, say queries. Translating that into faster or cheaper hardware requires another layer of evidence.

A claim map prevents accidental promotion

A claim map is a compact way to stop evidence from being promoted while it moves through a company.

Claims travel with their qualifiersSix evidence rows connect each supported source claim to the assumption, scope, or limitation that prevents accidental promotion.CLAIM-TO-QUALIFIER LEDGERClaims travel with their qualifiersSource-supported claimWhat must stay attachedMethodFilter the dominant-overlap statewithout its eigenvalue in advance.Dominance means initial overlap;the main setup assumes uniqueness.Near-optimalityFiltering query complexity is optimalup to logarithmic factors.Filtering step, access model,precision, query resource,and logarithmic-factor gap.Lower boundsModel-specific arguments includepurified query access.That result does not directly matchDEFEAT’s fuller structuredaccess model.ComparisonsAsymptotic improvements overphase-estimation variants.Not measured hardware orwall-clock advantage.NumericsAuthor-reported convergence ona five-qubit simulated model.Not independent replicationor quantum-hardwareexecution.LimitationDegenerate eigenspaces are notresolved into individual states.Keep this visible when callingthe method general.A qualifier is part of the review object, not fine print.
A qualifier is not fine print when removing it changes the decision.
Claim What the source supports What must stay attached
Method The authors introduce DEFEAT to filter the unique eigenstate with the largest overlap with the supplied initial state without knowing its eigenvalue in advance. Dominance is defined by initial overlap; the main setup assumes uniqueness.
Near-optimality Under stated assumptions, the filtering step’s query complexity is optimal up to logarithmic factors. Filtering step, access model, approximation target, query resource, and logarithmic-factor gap.
Lower bounds The paper provides model-specific lower-bound arguments, including one under purified query access. The purified-query result does not directly match DEFEAT’s fuller structured access model.
Comparisons The authors report asymptotic improvements over phase-estimation comparators on named resources. Asymptotic comparison, not measured hardware or wall-clock advantage.
Numerics The authors report convergence behavior on a five-qubit simulated model, with public simulation code. Author-reported simulation, not independent replication or quantum-hardware execution.
Limitation The current construction projects onto a degenerate eigenspace rather than distinguishing individual degenerate eigenstates. Keep this visible when describing the method as general.

The discipline is simple: if the qualifier changes the decision, it belongs beside the claim—not three links away in a technical appendix.

Deployment begins where the evidence ends

I use a second operator tool after the comparator card: a six-rung deployment-evidence ladder. It is also my operator synthesis, not a maturity model from the paper.

  1. Model-scoped theorem: Does the proof establish the objective against the comparator under named assumptions?
  2. Numerical convergence: Do simulations reproduce the predicted approximation behavior on specified model instances?
  3. Concrete implementation and resource estimate: Are gates, depth, logical qubits, oracle construction, and the error budget explicit for a target platform?
  4. Hardware evidence: Has the method run on relevant hardware against a declared baseline?
  5. Reliability and repeatability: Does it hold across devices, workloads, noise conditions, runs, and independent teams?
  6. Operational readiness: Does the system satisfy wall-clock, cost, integration, observability, security, and service-level requirements in a real workflow?

Evidence on one rung does not silently substitute for evidence on a higher rung.

Deployment begins where the evidence endsSix evidence stages progress from model-scoped theorem to operational readiness. This article assesses the reviewed DEFEAT v1 source as supporting the first two stages, contributing partial theory to the third, and requiring new evidence before hardware claims.DEPLOYMENT-EVIDENCE LADDERA stronger summary cannot replace a new class of evidence.1 · SUPPORTEDModel-scopedtheoremNamed assumptions2 · SUPPORTEDNumericalconvergenceSpecified model3 · PARTIALImplementationresource estimateTheory, not targetNEW EVIDENCEREQUIRED4 · NOT SHOWNHardwareevidenceDeclared baseline5 · NOT SHOWNReliability +repeatabilityAcross conditions6 · NOT SHOWNOperationalreadinessReal workflowSource-bounded assessment of the reviewed August 12, 2026 arXiv v1 preprint—not an author-provided maturity score.
The boundary is not a verdict on the research. It is the point where the next decision needs a different class of evidence.

For DEFEAT, the reviewed v1 preprint provides author-reported theorem-level evidence at the first rung, numerical evidence at the second, and partial theoretical material relevant to the third. It includes circuit constructions and asymptotic resource analysis, but not a deployment-grade estimate for a declared target platform.

The source does not establish quantum-hardware execution, wall-clock advantage, fault-tolerant feasibility, independent replication, operational reliability, or deployment readiness. The public code improves inspectability; it does not turn author-provided simulations into independent validation.

That boundary is not a criticism of the paper. Papers should not be forced to answer every product question. But roadmaps should not pretend the unanswered questions were settled.

What to ask before this enters a roadmap

When a technical claim arrives with an adjective such as optimal, efficient, robust, or scalable, do not debate the adjective first. Reconstruct the evidence contract.

Ask:

  1. What exact objective was optimized?
  2. What comparator or lower bound gives the adjective meaning?
  3. What access does the method assume, and what would that access cost to realize?
  4. What precision guarantee is being measured?
  5. What success or repetition overhead is included?
  6. Which resource improved—and which practical costs were not measured?
  7. What is the highest evidence rung actually reached?
  8. Which next claim requires new evidence rather than a stronger summary?

This approach avoids two common failures. The first is hype: treating a formal advantage as a deployment verdict. The second is reflexive skepticism: dismissing a valuable theorem because it does not already contain a production system.

The better move is to preserve both truths. A scoped result can be important. Its scope can also stop before your decision does.

Near-optimal is not weak language. It is incomplete language unless the comparator travels with it. Keep the six fields attached, mark unresolved deployment evidence explicitly, and make the next investment decision at the boundary the source actually supports.

Sources

m@berryhill in ~/posts$
$ cd ../ · back to posts/