| name | method-polyopt |
| description | Use when a noncommutative polynomial optimization reproduction needs method-level route and tool selection — certified lower bounds on ground-state energy via the moment-SOS / SOHS (NPA-style) hierarchy solved as a semidefinite program, Bell-inequality maximum quantum violation, or state-polynomial / tracial optimization. Triggers include polynomial optimization, SOS / SOHS relaxation, moment-SOS hierarchy, NPA hierarchy, certified energy lower bound, Bell inequality, semidefinite relaxation, NCTSSoS. |
Method PolyOpt
Overview
PolyOpt turns a hard optimization over quantum operators into a semidefinite program (SDP) whose optimum is a provable bound on the true answer.
- What it does. Minimizing a polynomial in noncommuting operator variables — a Hamiltonian H over all states, a Bell expression, a polynomial in expectation values — is intractable directly. Relax it: replace the operators by their moments (expectation values ⟨w⟩ of operator words w) and require the moment matrix (the matrix of those moments) to be positive semidefinite (PSD) — a consistency condition every genuine quantum state's moments satisfy. That relaxation is an SDP, and its optimum is a one-sided certificate: a lower bound on a minimum (an upper on a maximum). This is the moment-SOS / sum-of-Hermitian-squares (SOHS) hierarchy, also called the NPA hierarchy.
- Target. A certified number: a lower bound E_lb ≤ E₀ on a ground-state energy, the largest value quantum mechanics allows for a Bell expression (its Tsirelson bound), or two-sided bounds on a ground-state observable.
- What's approximated — two layers. (1) The relaxation: with a finite basis of operator words the bound is loose; enlarging the basis (longer words, higher order) tightens it monotonically. (2) The numerics: the solver returns the bound to finite precision — the digits you may claim are capped by its residuals (how exactly the returned solution satisfies the SDP), not by the printed decimals; a claim beyond that needs exact rational post-processing (arXiv:2512.17713: floating-point "bounds" can — and often do — exceed the true optimum).
- Its place in the harness. Every other method returns an estimate — a variational upper bound, a stochastic mean, a finite-size value. PolyOpt returns a rigorous lower bound: the rigorous half of a bracket E_lb ≤ E₀ ≤ E_var, a partner to DMRG / QMC / VMC rather than a competitor. For observable certification it even consumes a variational upper bound as input (step 1).
Four problem types — operator optimization, Bell inequality, state-polynomial optimization, observable certification — and which one it is fixes the whole formulation. Step 1 classifies; step 3 formulates.
When this card is invoked, before any choice, orient the user with this table, filling the right column with their actual problem — objective, operators, target. If those aren't fixed yet, use the table to elicit them.
| Ingredient | What it is | Your setup |
|---|
| Objective | what to bound — a Hamiltonian to minimize, a Bell expression to maximize, a polynomial in expectations, an observable to certify | (user's objective) |
| Operators | the operator variables and their algebra (Pauli, fermionic, bosonic, dichotomic = ±1-valued, projector, free) | (operators + algebra) |
| Constraints | the algebra's relations + any extra equality / inequality constraints (for observable certification: the energy window) | (constraints, if any) |
| Basis & order | which operator words enter the moment matrix — the accuracy/cost dial (order d = max word length) | (basis plan / order series) |
| Target | a certified bound, a Bell maximum, or concrete operators rebuilt from the solution (GNS reconstruction) | (which, and whether GNS is needed) |
| What's approximated | finite basis/order + solver precision (digits = residuals) | (tightening plan) |
Interaction principles — all user-facing surfacing in this card. Plain language, no jargon: define every term and symbol before first use. No walls of words — a few sentences or one compact table per turn. One decision at a time, recommendation-first with one-line pros/cons. Precise and concise; let the user feel each choice, never a silent default.
Sources
- Methodology reference (reproduction-grade algorithm, parameters, validation, gap analysis):
references/polyopt-methodology.md
- Tool skills (step-2 targets):
/using-qmbcertify — the structured certifier for 1D/2D (J1-J2) Heisenberg models (Mosek); /using-nctssos — the general NC-polyopt engine for any algebra / Bell / state-polynomial (Clarabel/Mosek).
- Primary literature (rendered in
.knowledge/literature/polynomial-optimization/; each carries its source URL in the frontmatter):
- 2604.01555 — Wang, Jansen, Frérot, Renou, Magron, Acín (2026) — structured NPA certification scaling to 16×16 lattices; the current state of the art.
- 2310.05844 — Wang et al., PRX 14, 031006 (2024) — the predecessor that introduced the energy-window observable certification and the structure-exploiting relaxations.
- 10.1007/978-3-319-33338-0 — Burgdorf, Klep, Povh, Optimization of Polynomials in Non-Commuting Variables (Springer, 2016) — the foundational monograph (eigenvalue/trace hierarchy, GNS, flatness, rational certificates).
- Online (pull with
/download-ref if needed): NPA, New J. Phys. 10, 073013 (2008), arXiv:0803.4290 — the original hierarchy; Naceur, Wang, Magron, Acín, arXiv:2512.17713 — exact rational certification of solver bounds.
- The modeling craft in this card (problem-type classification, algebra selection, formulation, sparsity, GNS) is distilled from the
polyopt-guide skill (exAClior/easy-nctssos, authored by the NCTSSoS maintainers) — absorbed here rather than referenced, so method and software stay decoupled.
Select method — step 1
Suited for
- A certified lower bound on a ground-state energy, the maximum quantum violation of a Bell inequality, or two-sided bounds on a ground-state observable — wherever rigor is the point.
- Complementary certification, not a replacement: pair it with a variational/stochastic method and the pair brackets the truth. No sign problem — frustrated spins and fermions are equally admissible.
Worked examples — demonstrated reach and its limits
Anchor the user's problem to the nearest row; quote the scale and the lesson. Capability anchors from the literature (Sources), not reproduction mandates.
| Ref | Problem | Scale | Certified result | Lesson |
|---|
| 2604.01555 Tab.3 | Heisenberg chain energy | N up to 100 | E/spin within ~2×10⁻⁵ relative of DMRG | near-exact in 1D unfrustrated |
| 2604.01555 Tab.8 | square-lattice Heisenberg energy | up to 16×16 (256 spins) | 0.7% gap vs QMC | the structured-SDP scale record (32 cores / 1 TB) |
| 2604.01555 Tab.4 | J₁-J₂ chain energy | N=40, J₂ up to 2 | exact at the Majumdar-Ghosh point J₂=0.5; worst 0.7% at J₂=1 | frustration costs accuracy; the MG point is a free end-to-end check |
| 2604.01555 Tab.5-7 | chain correlations, structure factor | N=40, two-sided | 0.01% at short range; up to ~100% deep in frustration | a Hamiltonian-local basis bounds long-range observables loosely |
| 2310.05844 Tab.X; 2604 Tab.9 | 2D long-range order C(L/2,L/2) | L = 4…16 | LRO certified (lower bound > 0) only for L ≤ 8; sign lost by L=16 | finite-size bounds do not extrapolate to the thermodynamic limit |
| 2310.05844 Tabs.XIII-XV | 2D J₁-J₂ energy + correlations | L=10 (N=100), frustrated | energy 1-7% above best variational; correlation sign changes certified | certification where no exact method exists — at few-percent looseness |
| NPA / BKP | CHSH; I3322 Bell violation | few operators | CHSH exact (2√2) at low level; I3322 exact value still open | small Bell problems converge fast — but not all close |
Route elsewhere — when PolyOpt isn't the right tool
| Target | Better tool | Why |
|---|
| The ground state itself (wavefunction, entanglement, fidelities) | DMRG /method-mps, VMC /method-vmc, QMC /method-qmc | PolyOpt returns bounds + moment data, not a state (GNS rebuilds a realizing model, not the lattice ground state) |
| Best-possible energies of sign-free models | QMC /method-qmc, DMRG /method-mps | certified accuracy beyond 1D is ~10⁻³…10⁻² relative — orders looser than those estimates; certify alongside them, don't replace them |
| Thermodynamic limit | infinite-size methods (/method-mps VUMPS, /method-peps, QMC) | finite-size SDP bounds grow looser with size and cannot be extrapolated |
| Full spectrum, dynamics | ED /method-ed, MPS /method-mps | the hierarchy targets the extremal eigenvalue; no dynamics formulation exists |
| Finite temperature | LTRG /method-ltrg, QMC /method-qmc | certified finite-T relaxations exist (Fawzi-Fawzi-Scalet 2024) but have no large-scale demonstrations yet |
When the goal falls outside PolyOpt: recognize it before any setup; explain what fits better and why in a short what/why table; stay warm — guide, don't dismiss.
Options & trade-offs — the four problem types
| Type | Objective looks like | Formulation note |
|---|
| Operator optimization | minimize H = Σ couplings × operator words | the common case; SDP gives a lower bound on the minimum eigenvalue |
| Bell inequality | maximize a Bell expression B | encode as minimize −B; key choice: party-wise commuting groups (standard) vs tracial |
| State-polynomial | products of expectations — ⟨A⟩⟨B⟩, variances | needs the state-polynomial formulation (wrappers in /using-nctssos); the trickiest setup |
| Observable certification | bound ⟨O⟩ at the ground state | min and max ℓ(O) under an energy window E_lb ≤ ℓ(H) ≤ E_ub — E_ub is a variational input (DMRG/QMC); the certificate inherits its gap |
Surface the classification one question per turn — most problems classify at question 1; stop as soon as the type is fixed, and confirm it before any setup.
| Ask | Answer → type |
|---|
| 1. What are you bounding? | a Hamiltonian's minimum → operator · a Bell expression's maximum → Bell · a product of expectations → state-polynomial · an observable at the ground state → observable certification |
| 2. (only if still unclear) Are the variables physical operators or abstract ±1 outcomes? | physical → operator / certification · abstract outcomes → Bell |
| 3. (only if still unclear) Is the objective linear in the state? | linear → operator / Bell · products of expectations → state-polynomial |
Two types carry a modeling sub-choice worth surfacing (the using-card expresses it; this card decides it):
- Bell — operator vs tracial. Operator (recommended, the standard quantum-mechanics choice): each party's measurements in a separate group, so different parties' operators commute. Tracial (tracial = scored by a trace, ⟨·⟩ = tr(·), rather than by a state): one group, transpose trick — it converges to a different mathematical value (the von Neumann-algebra optimum); use only to study the tracial relaxation itself.
- State-polynomial — the wrappers. When the objective multiplies expectations (⟨A⟩⟨B⟩), wrap operator expressions as
tr(·) or s(·) (expectation in an arbitrary state), then assemble the state polynomial. /using-nctssos carries the API.
Certification role — the bracket
PolyOpt's output is one half of a two-sided certificate; plan the other half:
- A variational energy E_var is an upper bound; the SDP gives the lower bound: E_lb ≤ E₀ ≤ E_var. A small gap certifies both. Compose with
/cross-method-check.
- Observable certification needs the variational value as an input (the energy window) — pull it from the paper or plan the variational run first.
- Report bound-vs-order/basis, not a single number.
Select software — step 2
Routing rule: a 1D/2D (J1-J2) Heisenberg model where you want maximum scale → /using-qmbcertify; everything else → /using-nctssos. The choice turns on whether the structured certifier already specializes for the model.
| /using-qmbcertify | /using-nctssos |
|---|
| Use when | the problem is a 1D/2D (J1-J2) Heisenberg model and you want the tightest certified bound at large size | any other NC-polyopt problem — other algebras, custom Hamiltonians, Bell, state-/trace-polynomials |
| How it scales | hard-codes the model's symmetries plus RDM and state-optimality constraints → block-diagonalizes the SDP by many orders of magnitude (reaches 16×16) | generic correlative + term sparsity, or symmetry (Wedderburn) reduction for group-invariant problems — one lever or the other per run; scales to local Hamiltonians of moderate size |
| Solver | Mosek only (free academic license) | Clarabel (open-source) or Mosek |
| Returns | a certified numeric bound + Gram-matrix export (exact rational certification is a separate post-step — packaged for 1D chains only) | a numeric SDP bound + moment data + GNS |
Surface the software choice — four short lines, not a wall:
- What they are: QMBCertify.jl (the structured certifier) and NCTSSoS.jl (the general engine), both from the NCTSSoS author J. Wang's group.
- The deciding fact: whether the baked-in Heisenberg symmetries apply — they are what reach 16×16 where the general engine cannot.
- The one consequence: QMBCertify hard-requires Mosek (free academic license); NCTSSoS defaults to open-source Clarabel.
- When reproducing the structured-certification paper, the reassuring fact: QMBCertify is the paper's own published code.
Let the user feel the choice even when one engine is the obvious fit.
Handoff. Once the engine is fixed, invoke /using-qmbcertify or /using-nctssos — it owns install/run, the package's run knobs (step 3 software side), and the cost estimate. This card owns the modeling (below): algebra, objective, basis, relaxation strategy.
Method setup — step 3
The modeling decisions this card owns. Software-side values (API names, solver settings) live in the using-card. Two kinds of rows: confirm = determined by the problem — state it and let the user ratify; decide = a genuine choice — recommend, then ask.
| Knob | Kind | Default / how to set | Effect |
|---|
| Problem type | confirm | from step 1's classification | fixes the whole formulation |
| Algebra | confirm | the most specific algebra the operators obey (table below) | richer relations → smaller basis and tighter bound — a free win |
| Encoding | confirm | minimize H directly; maximize f as minimize −f; complex coefficients for Pauli; party-wise commuting groups for Bell | a sign or coefficient-type error silently bounds the wrong thing |
| Monomial basis + range r | decide — the expert knob | words on contiguous sites up to length d, plus two-body words σᵢᵃσᵢ₊ⱼᵇ out to range r as memory allows; re-tailor the words to a non-local target observable | where tightness per cost is won; a Hamiltonian-local basis leaves long-range observables loose |
| Relaxation order d | decide | the lowest order containing the objective (usually d=2); climb while the bound still moves and the budget allows | monotone tightening; SDP size grows ~×n per step |
| Sparsity (CS/TS) | decide | on for any local Hamiltonian (correlative = variable cliques; term = monomial blocks) | big SDP shrink; TS stabilization ≠ exactness — the bound can stay below the dense one |
| Symmetry | confirm | exploit every symmetry the model has | exact for symmetric ground states — but restricts to the symmetric sector (beware at degeneracies / critical points) |
| Strengthenings | decide | RDM positivity on (k ≈ 8 is the cost/benefit sweet spot); linear state-optimality on; PSD state-optimality off — documented solver failures in frustrated regimes | tighten at fixed order, for extra SDP size |
| Side to solve | confirm | the dual (SOHS) side — fewer constraints, same optimum | constant-factor speed |
| Solver & tolerances | decide | per using-card; trustworthy digits = solver residuals | beyond-residual claims need exact rational post-processing |
| GNS reconstruction | decide | off unless explicit operators are needed | needs a flat moment matrix and a higher order |
Algebra — pick by the operators' physics (the relations it auto-enforces give the tighter bound):
| Operators in the problem | Algebra | Relations enforced | Note |
|---|
| Pauli sx, sy, sz on spin-½ sites | Pauli | s²=I + product rules sx·sy=i·sz | tightest for spins; the product rules beat bare Unipotent at the same order |
| fermionic creation/annihilation | Fermionic (CAR) | {aᵢ,aⱼ†}=δᵢⱼ, {aᵢ,aⱼ}=0 | Hubbard, t-J, free fermions |
| bosonic creation/annihilation | Bosonic (CCR) | [aᵢ,aⱼ†]=δᵢⱼ | ∞-dim Hilbert space; GNS gives finite approximations |
| ±1 measurement observables | Unipotent | U²=I only | abstract Bell observables — not physical spins (Pauli there would wrongly fix the dimension) |
| projective measurements | Projector | P²=P | I3322, measurement compatibility |
| no special relations | free NonCommutative | none | add custom constraints by hand |
Confirm the setup with the user before running — one knob per turn, never batched (interaction principles above).
- Orient once. One plain-language hook: "We rewrite your minimization as a positivity problem whose answer is a guaranteed lower bound on the true energy. The words we put in the moment matrix are the dial: more and longer words = a tighter guarantee, at steeply growing cost."
- Confirm, then decide. Run the confirm rows first as statements to ratify — type, algebra, encoding, symmetry — even when they look obvious (a silent sign convention is exactly what the user catches at a glance). Then loop the decide rows one per turn, leading with the two that set the result: basis + range r and order d on the general engine; on the structured route (basis family package-built) order/range + which strengthenings.
- Recommended option first (labeled when there's a technical reason — e.g. the paper's value when reproducing), one-line why, 1-2 alternatives with one-line pro/con. Ask one question, then STOP and wait.
- Record each agreed choice in the invoking workflow's plan (e.g.
/reproduce-paper's parameter rows) before the run, so the proposal-first report is faithful. Then hand env + execution to the chosen using-card.
Cost & resource estimate — feeds step 4
Cost is one measured rate × a firm work count — and for an SDP the firm count is the post-reduction block inventory, known before solving: assemble the SDP without solving and read the block statistics (the structured route even has closed-form block sizes).
| Axis | Scaling |
|---|
| Compute | interior-point iterations (~tens) × a per-iteration factorization cost set by the largest PSD blocks and the constraint count. The reductions are the whole game: 1D Heisenberg N=100, d=4 — max block 8.1×10⁹ naive → 3.2×10⁸ after Pauli equalities → 12,001 after sparsity → 31 after symmetry (2604.01555 Tab.2) |
| Memory | the usual first wall — grows ~quadratically in the constraint count; 128 GB carried N=100 (1D) and L=10 (2D); 16×16 took 1 TB |
| Wall time | anchors (Mosek, single core, 2310.05844): chain N=40 ≈ 3.3 h, N=100 ≈ 12 h; 2D L=6/8/10 ≈ 1.8 / 9.7 / 21 h — per parameter point |
Probe protocol: assemble without solving, read the block statistics, and sanity-check them (a structured Pauli chain at d=4 must not show a ~10⁶ dense block — if it does, the reductions didn't fire; abort and fix). Then one low-order solve measures the throughput. Then decide local vs /using-slurm.
Surface the cost to the user before any scale choice (reproduce-paper step 4). Plain language: the basis/order and the system size drive the SDP blocks up, sparsity and symmetry drive them back down, and memory — not time — is usually what runs out first. Show the per-size reality (chains in hours on one core; large structured lattices on a fat workstation) and let the user feel that picking basis, order, and size is picking the cost.
Details
Generic methodology; paper/model facts live in /reproduce-paper and .knowledge/models/. Math is unicode/plain.
The idea
A minimization min_state ⟨H⟩ is hard because the set of valid quantum states is hard to describe. Replace the state by a linear functional ℓ on operator words (the moments ⟨w⟩) that need only satisfy necessary conditions: ℓ(1)=1, ℓ is positive on Hermitian squares (the moment matrix M with Mᵤᵥ = ℓ(u⋆v) is PSD), and ℓ respects the algebra's relations. Minimizing ℓ(H) over all such ℓ is an SDP, and because every true state gives a valid ℓ, its optimum lower-bounds the true minimum. Restricting words to a finite basis gives the truncated relaxation; enlarging the basis adds constraints and tightens the bound monotonically. Dually, a feasible point is a sum-of-Hermitian-squares decomposition certifying H − E_lb ⪰ 0 — and it is this dual (SOHS) side one actually solves (fewer constraints, same optimum, no duality gap).
Structured reductions (why it scales)
The general hierarchy blows up; exploiting the problem's structure is what makes large systems feasible:
- Basis locality. Words supported on contiguous sites (1D) or compact clusters (2D), plus long-range two-body words out to range r — the hand-crafted core of the published large-N results. The basis is chosen for the Hamiltonian; bounding a non-local observable well needs words tailored to it.
- Sparsity. Correlative sparsity groups variables into cliques (each clique → a smaller moment matrix); term sparsity finds block structure from the monomials that actually appear. Caveat: stabilization ≠ exactness — the term-sparsity iteration can stop growing while the bound stays strictly below the dense-basis bound at the same order.
- Symmetries (Heisenberg-specific, automated in
/using-qmbcertify). Sign, conjugate, permutation, dihedral, and translation symmetry block-diagonalize the SDP (translation via circulant/DFT structure; 2D adds a second round). Together with the algebraic equalities and the sparse basis they take the 1D max block from ~10⁹ to ~10¹ (Tab.2 cascade above). Exact for symmetric ground states — with the symmetric-sector caveat (step 3).
- Strengthenings. RDM positivity (PSD constraints on k-body reduced density matrices, block-diagonal by U(1) magnetization; k ≈ 8 sweet spot) and state-optimality conditions (ℓ([H,u]) = 0, plus a PSD variant known to destabilize solvers in frustrated regimes) tighten the bound without enlarging the basis.
Notation
- Moment ℓ(w) = ⟨w⟩ — expectation of operator word w; the SDP variables.
- Relaxation order d — the max word length in the basis; the ladder the bound climbs monotonically.
- Moment-SOS / SOHS (NPA) hierarchy — the moment relaxation and its sum-of-Hermitian-squares dual.
- CS / TS — correlative / term sparsity. GNS — rebuild concrete operators/state realizing the optimal moments.
- Flatness — the moment matrix's rank stops growing as the order rises; certifies the bound is (numerically) exact and GNS is reliable.
- Residuals — how exactly the returned solution satisfies the SDP's constraints; they cap the digits a bound claim may carry.
Pitfalls
- A solver's "lower bound" can exceed the true optimum when the returned point is slightly infeasible — even with an OPTIMAL status. Quote digits at the residual level; certify beyond only via rational post-processing.
- It is a bound, not the state. Don't read GNS moments as lattice ground-state correlations; GNS gives a realizing representation.
- Observable certificates inherit the variational input. The energy window's upper side comes from DMRG/QMC; a loose E_ub directly loosens the certified interval.
- Basis-vs-observable mismatch. A Hamiltonian-local basis certifies energies tightly while leaving long-range correlators uselessly wide — re-tailor the words per observable.
- min vs max. The SDP minimizes; maximize f by minimizing −f and negating.
- Pauli needs complex coefficients. sx·sy = i·sz generates imaginary terms; build the objective over ℂ.
- Too-loose algebra. Unipotent where Pauli applies wastes tightness; Pauli where Bell's dimension-freeness is the point is wrong, not just loose.
- TS stabilization ≠ exactness; no thermodynamic-limit extrapolation from a finite-size bound series; symmetric-sector restriction at degeneracies; PSD state-optimality can break the solver in frustrated regimes — drop it and re-run, don't paper over a stalled solve.
Verification
The run's own output already carries the correctness signals — read them; they're free. Anything that costs extra compute is opt-in: only when a result looks wrong, debugging hasn't resolved it, and the user explicitly asks to check. Then propose 2-3 checks from the menu below — what each confirms, its rough cost — and run only what's confirmed (the same opt-in rule the invoking workflow enforces, e.g. /reproduce-paper's implementation stage).
Intermediate output — free, read on every run
- Before the solve: the assembled block statistics vs expectation — if the structure machinery didn't fire (one huge dense block where many small blocks are expected), abort and fix; never pay for a solve that's already wrong.
- At the solve: solver status — only a clean optimal status is a bound; stalled / slow-progress / iteration-limited = no certificate. Residuals / duality gap — the digits the bound may claim.
- After: the sandwich, E_lb ≤ reference value. And if an order/basis series is already in the plan: the bound must rise monotonically — a decrease on enlarging the basis is mathematically impossible and means a modeling bug, not physics.
Opt-in checks — the user-triggered menu
| Check | Confirms | Rough cost |
|---|
| Anchor point — Majumdar-Ghosh J₂=0.5 (exact), CHSH 2√2, small-N vs ED | the whole pipeline, end to end | one small solve — cheap |
| Dense vs sparse at low order | the sparse basis isn't hiding looseness | one extra solve |
| Order/basis escalation | how far from converged the bound is | one larger solve — the cost driver |
| Flatness test + GNS round-trip | bound (numerically) exact; extraction meaningful | cheap, on existing moment data |
| Rational certification | the claim survives floating-point doubt | post-processing; frontier claims only |
Cross-method bracket (/cross-method-check) | the bound against an independent upper bound | one independent run |
Criticize (expert probes for a challenged result): trusting a stalled/infeasible solve; quoting printed decimals beyond the residuals; one order with no series; a sign error in min-vs-max; Unipotent where Pauli applies (or Pauli where Bell's dimension-freeness is the point); TS stabilization read as convergence; GNS moments read as the lattice ground state; a long-range observable certified with a Hamiltonian-local basis; "exact certification" claimed without a verified rational rounding.
Citations
Rendered under .knowledge/literature/polynomial-optimization/ (see Sources):
2604.01555_…md — Wang et al. (2026) — structured NPA certification to 16×16 lattices; the source for the basis/symmetry reductions, the block-size cascade (Tab.2), and the strengthening caveats (Remarks 3.1, 6.1).
2310.05844_…md — Wang et al., PRX 14, 031006 (2024) — energy-window observable certification; single-core wall-time anchors; RDM k=8 sweet spot.
10-1007-978-3-319-33338-0.md — Burgdorf, Klep, Povh (2016) — the moment-SOHS hierarchy, GNS, flatness, rational certificates.
- NPA hierarchy: Navascués, Pironio, Acín, New J. Phys. 10, 073013 (2008), arXiv:0803.4290. Exact rational certification: Naceur, Wang, Magron, Acín, arXiv:2512.17713.
- Modeling craft distilled from the
polyopt-guide skill (exAClior/easy-nctssos).
- Software:
/using-qmbcertify (QMBCertify.jl + Mosek) and /using-nctssos (NCTSSoS.jl + Clarabel/Mosek).