Undecidability in Physical Theory

Speculative If the undecidability theorems say something about matter rather than only about a family of Hamiltonians assembled to carry a Turing machine — which is precisely what is unsettled, and section 7 says what would settle it — then one research programme has a ceiling that nobody had drawn and that no amount of compute moves. The programme is the obvious one: given a local Hamiltonian, return its bulk properties. The theorem says no algorithm does that in general, and it says so about a class narrow enough to be recognisable as physics.

Established The magnitude is best stated as the size of the class the impossibility covers, because that is the quantity this subject actually produces. In 2015 the class was translationally invariant, nearest-neighbour Hamiltonians on a two-dimensional square lattice of d-level systems with d fixed and finite, specified by at most three finite-dimensional Hermitian matrices. In 2020 it was one-dimensional spin chains, which nobody had expected. In 2021 it was a continuous one-parameter family whose undecidable set has positive measure in parameter space rather than measure zero. In January 2026 it was models with four-body plaquette interactions obeying the lattice’s rotational symmetry. Frontier Each step removed one reason to believe the result was an artefact of a contrived setting, and none of them has yet removed the largest one.

Frontier The gain accrues to a small group, and it is not the group the headlines suggest. No experimentalist gets anything; no materials programme gets a faster answer. What gets something is the certification literature, which the theorem tells what shape a defensible answer has to take. A 2024 result gives algorithms returning rigorous upper and lower bounds on equilibrium expectation values in the thermodynamic limit, with convergence quasi-polynomial in 1/ε for commuting two-dimensional Hamiltonians at high temperature. A 2026 semidefinite hierarchy certifies spectral gaps for frustration-free chains, subsuming Knabe’s finite-size bound and reporting accuracy improvements of several orders of magnitude over it. Speculative Both exist because the general problem is closed: you build certificates when you cannot build a solver.

Handwave The condition, and it is doing all the work: the undecidability has to survive contact with a real system, and in one dimension, as of July 2026, it does not. A perturbation of norm O(ε), for arbitrarily small ε, added as a single one-local rank-one term, makes the spectral gap problem for the 2020 family decidable. Speculative Whether the two-dimensional constructions are equally fragile is open, and it is the question on which everything in section 12 turns.

A physical question becomes undecidable the way any question does: it stops being a question about one object and becomes a question about an infinite family, indexed by something a machine could enumerate. “Is this crystal gapped?” is not undecidable and cannot be. “Given an integer n, is the n-th member of this family gapped?” can be, and is. Everything difficult about this subject lives in the distance between those two sentences.

Established The machinery is a chain of three constructions, each of which was already standard before anyone pointed it at physics. Turing’s 1936 halting problem supplies a question no algorithm answers. Berger’s 1966 domino problem and Robinson’s 1971 aperiodic tilings supply a way to write a computation into the ground state of a two-dimensional lattice, by making the tiling pattern the tape. Circuit-to-Hamiltonian constructions in the style of Feynman and Kitaev supply a way to make a local energy penalty enforce the steps of that computation. Established Bolt them together and the ground state of a spin lattice becomes a record of a Turing machine’s history, and a low-energy property of the lattice becomes a statement about whether the machine halts.

Established The quantum version added a fourth ingredient and that is what made it a physics result rather than a coding exercise. The 2015 construction runs a quantum phase-estimation algorithm ahead of the universal Turing machine, so that the machine’s input is read off a continuous parameter of the Hamiltonian rather than written in by hand. Frontier The spectral gap then depends on the outcome of the halting problem for that input, and the dependence is sharp because the theorem comes with a promise: each Hamiltonian in the family is either gapless in the strongest sense — continuous spectrum above the ground state in the thermodynamic limit — or gapped with a gap bounded below by a constant that can be taken equal to the local interaction strength. There is no ambiguous middle to hide in.

Established Three infinities are always present, and the review the field wrote for itself says so in as many words. An undecidable problem “must necessarily conceal an infinite somewhere: infinitely many instances, infinitely many particles, infinite precision”. In these constructions all three appear: the family is infinite, the lattice is infinite because the gap is defined in the thermodynamic limit, and the matrix elements are exact algebraic numbers. Speculative Remove any one of them and the problem is decidable by enumeration, which is why no laboratory has ever met an undecidable question and no laboratory ever will.

Frontier The subject has two strands that are usually reported as one and should not be. The quantum many-body strand runs from 2015 and concerns spectral gaps, phase diagrams, thermalisation and renormalisation flows. The older classical strand runs from the 1980s — Wolfram on cellular automata, Moore on dynamical systems, Pour-El and Richards on the wave equation — and concerns trajectories, limiting behaviour and computability of solutions to differential equations. Speculative They share a technique and almost nothing else: the classical strand asks whether a trajectory does something, the quantum strand asks whether a bulk phase is one thing or another, and the second is far closer to a question a condensed-matter physicist would recognise as their own.

Frontier The framing this brief tests is that undecidability constrains what a general algorithm for material properties could ever be, and it survives — with the single, large qualification that every construction so far is, in its authors’ own word, artificial. Handwave The step from “this family is undecidable” to “physics is undecidable” is not a small extension of the theorem. It is a different claim with no proof attached, and section 15 exists to name it.

Established The anchor result is a theorem, published twice, and its scope conditions are stated by its authors rather than inferred. Cubitt, Pérez-García and Wolf proved in Nature 528, 207–211 (9 December 2015), with the full proof in Forum of Mathematics, Pi 10 (10 June 2022), that the spectral gap problem is undecidable. The construction is a family of translationally invariant, nearest-neighbour Hamiltonians on a two-dimensional square lattice of d-level quantum systems with d constant, specified by at most three finite-dimensional Hermitian matrices. Established The abstract names what the problem covers: the Haldane conjecture, the existence of gapped topological spin liquid phases, and the Yang–Mills gap conjecture are all instances of “given a Hamiltonian, is it gapped or gapless?” Established The paper also states a Goedel-style corollary: given any consistent recursive axiomatisation of mathematics, there exist particular Hamiltonians in this class for which the presence or absence of the spectral gap is not determined by those axioms.

Established The number the anchor implies and the papers do not give is the local dimension, and both versions say so explicitly. The Nature paper: “Although the dimension d in Theorem 1 is fixed and finite, providing an estimate for it would be cumbersome and certainly involve large exponentials.” The full version: “It is straightforward (if tedious) to extract an explicit value for d in Theorem 3 and Corollary 4 from the construction described in this paper” — an invitation, not a value. Frontier No source retrieved for this brief publishes a number for d in the two-dimensional construction, which means the phrase “explicit construction” is true of the method and false of the model. Established The same paper is equally direct about what the models are: “The models we construct in the proof of Theorem 1 are highly artificial. Whether the results can be extended to more natural models is yet to be determined.”

Established The result then tightened four times, in four different directions, and the sequence is the real evidence that it is not a fluke of one setting. In 2020 Bausch, Cubitt, Lucia and Pérez-García proved the same undecidability for one-dimensional translationally invariant nearest-neighbour spin chains (Phys. Rev. X 10, 031038), which had been expected to be the tractable case; the new ingredient was quantum interactions that, despite being local, generate patterns with arbitrarily long periods in 1D. Established In 2021 Bausch, Cubitt and Watson proved that the phase diagram is uncomputable, constructing a continuous one-parameter family H(φ) with φ real, translationally invariant with nearest-neighbour couplings on a 2D spin lattice, with row and column interaction norms bounded by 2 and 1 (Nature Communications 12, 452). Frontier That paper’s advance is a measure-theoretic one: earlier constructions gave undecidability only on a set of measure zero and required the Hamiltonian to be a discontinuous function of its parameters; this one gives it on a set of positive measure, for a continuous family.

Established Two further results extended the reach from static properties to dynamics and to method. Shiraishi and Matsumoto showed in 2021 that whether a given system thermalises is undecidable, for shift-invariant nearest-neighbour Hamiltonians on a one-dimensional chain with periodic boundaries, a fixed product initial state and a spatially averaged single-site observable (Nature Communications 12, 5084). Established That paper carries the only explicit local dimension in this brief: the authors estimate that d of about 120 suffices. Established Watson, Onorati and Cubitt then built a rigorous renormalisation group map for the original undecidable system and showed that every step of the map is computable and converges to the correct fixed points, while the flow itself is uncomputable (Nature Communications 13, 7618, December 2022). The flow runs between a gapped Ising-like fixed point and a gapless critical XY-like one, and which one a trajectory reaches cannot be predicted from exact initial parameters.

Frontier Then, in 2026, the field turned on its own construction, and this is the most important development for anyone asking whether the theorem is about physics. Castilla-Castellano and Lucia first closed the obvious escape route: in Phys. Rev. Research 8, 013033 (14 January 2026) they proved undecidability for models with four-body plaquette interactions on the square lattice satisfying rotational symmetry, answering the question of whether “the spectral gap problem could be decidable for Hamiltonians with stronger symmetry constraints” with, in their words, “a negative answer”. Frontier Then they opened a much larger one. In July 2026 the same two authors showed that the one-dimensional undecidable family is “extremely sensitive to perturbations”: for any ε > 0 arbitrarily small there is a one-local, rank-one interaction of norm O(ε) whose addition makes the spectral gap problem for that family decidable. Speculative The undecidability of the 1D construction is not robust, and robustness is the property that separates a model from a physical prediction.

Established The classical strand is older, thinner and much more often misquoted. Wolfram’s Phys. Rev. Lett. 54, 735–738 (25 February 1985) proposed treating physical processes as computations and used cellular automata to exhibit formally undecidable questions: whether a pattern from a given seed dies out or grows forever, whether a generalised Ising model has a phase transition in the infinite-size limit, whether solutions of certain differential equations enter a given region, whether an n-body system is ultimately stable. Frontier Moore’s Phys. Rev. Lett. 64, 2354 (1990) did the corresponding job for smooth dynamical systems via generalised shifts. Speculative Pour-El and Richards had shown in 1981 that the wave equation admits computable initial data whose unique solution is not computable. This session could not retrieve either the Moore or the Pour-El–Richards paper, and both are cited here by name; no claim in this brief rests on a number from them.

Established The fluid results are the most recent part of the classical strand and the most heavily over-reported. Cardona, Miranda, Peralta-Salas and Presas constructed a Turing complete stationary solution of the incompressible Euler equations — a Beltrami field — on a Riemannian three-sphere (PNAS 118, issue 19, 4 May 2021), so that “the question of determining whether a certain trajectory will enter a certain open set is as difficult as the halting problem for Turing machines”. Established The metric is part of the construction, not a given: the flows live on a specially built Riemannian manifold, and the corresponding statement for the Navier–Stokes equations was explicitly left open, because using these flows as Navier–Stokes initial conditions reproduces only “a (fixed) finite number of steps of any Turing machine”. Frontier That gap closed partway in 2026: Dyhr, González-Prieto, Miranda and Peralta-Salas constructed Turing complete stationary Navier–Stokes solutions on three-manifolds admitting nonvanishing harmonic one-forms (PNAS Nexus 5(5), pgag131), showing “that computational universality is not obstructed by viscosity, provided the underlying geometry satisfies a mild cohomological condition”. Speculative They also state the limitation plainly: “the resulting metric is in general different from the original one”, and whether the construction works for the round metric on the three-sphere “remains as an open problem”.

Established Against all of that sits a record of what is computable, and it has grown faster since 2024 than the impossibility results have. Fawzi, Fawzi and Scalet published certified algorithms for equilibrium states of local quantum Hamiltonians (Nature Communications 15, 7394, 27 August 2024) that output rigorous lower and upper bounds on local observables in the thermodynamic limit, at zero and positive temperature, with the explicit note that this “allows us to show that expectation values of local observables can be approximated in finite time, contrasting related undecidability results”; for commuting Hamiltonians on a 2D lattice at high temperature the program size is quasi-polynomial in 1/ε. Established Rai, Kull, Emonts, Tura, Schuch and Baccari published a semidefinite hierarchy of spectral gap certificates for frustration-free spin systems (Quantum 10, 2065, 13 April 2026) that contains Knabe’s finite-size bound as a special case and reports accuracy improvements of several orders of magnitude. Established And Pilatowsky-Cameo and Choi proved in December 2025 that quantum thermalisation must occur in any qubit system with local interactions given high effective temperature, translation invariance and no perfect spectral resonances — explicitly acknowledging the undecidability result and routing around it with a statement about typical states rather than individual ones. Frontier The pattern across all three is the same: the general problem is closed, so the field proves theorems about restricted classes and typical instances, and it is succeeding at that.

Established The field wrote itself a review in 2025 and its framing is the most useful single sentence in this brief. Perales-Eceiza, Cubitt, Gu, Pérez-García and Wolf, in Physics Reports 1138 (21 September 2025), pages 1–29, state: “It is important to clarify that undecidability is not a feature of the physical system; it is a feature of the mathematical model we use to describe that physical system.” Established They add that “None of these idealized limits are directly accessible experimentally” and, in the other direction, that “Undecidability in some idealized mathematical limit of a physical system can be reflected in surprising – sometimes even qualitatively new – (but still in principle computable) physical properties in the real, effectively finite physical system.” Frontier That last clause is the whole live question, and it is the one this brief follows.

Frontier The open questions here are unusually well posed, because in a field whose results are theorems the disagreements are about what the theorems are about rather than about what they say. They are set out as positions with holders.

Frontier Position one, and the one that has moved most: is undecidability stable? A physical prediction that is destroyed by an arbitrarily small perturbation is not a prediction about anything that could be built. Established In one dimension the answer arrived in July 2026 and it is no: a one-local rank-one term of norm O(ε) restores decidability, and the authors attribute this to the perturbation disrupting a fine-tuned energy cancellation in the marker Hamiltonian. Speculative The two-dimensional case is untouched, and there are reasons to think it may be sturdier — the 2015 result is stated to survive arbitrarily small quantum perturbations of classical Hamiltonians — but nobody has proved it either way. Handwave Anyone quoting the 1D theorem after July 2026 without the stability caveat is quoting a result whose physical reading has been withdrawn by its own community.

Speculative Position two: how simple can an undecidable model be? The only published local dimension anywhere in this literature is Shiraishi and Matsumoto’s estimate of about 120 for the thermalisation construction; the spectral-gap papers decline to give one and say the estimate would “certainly involve large exponentials”. Frontier Every tightening since 2015 has been about structure — fewer dimensions, continuous parameters, more symmetry — and none has been about size. Handwave There is no published lower bound on the local dimension of any undecidable family, which means the distance between these constructions and a spin-1 chain is not merely large but unmeasured, and a field that cannot bound a quantity cannot claim it is small.

Frontier Position three: does undecidability in the limit leave a fingerprint at finite size? The 2020 one-dimensional paper predicts chains with a constant spectral gap and a non-degenerate classical ground state up to some critical size, which then switch to gapless behaviour with a dense spectrum — a size-driven transition whose threshold can be uncomputably large. Speculative The review takes the same line from the other side: undecidability in an idealised limit can show up as “qualitatively new” but still computable behaviour in the finite system. Speculative If that is right, the observable residue of this subject is not an uncomputable measurement — there is no such thing — but a class of systems whose bulk behaviour no finite-size extrapolation predicts, and which can therefore be built at a designed critical size. Handwave Nobody has built one.

Frontier Position four: is this a Goedel result or a Turing result? The papers prove both — there is an undecidability statement and an axiomatic-independence corollary — but they are used interchangeably in circulation and they are not the same claim. Established Undecidability says no single algorithm settles every instance; independence says a particular instance is settled by no proof from a given axiom system. Speculative The second is the stronger and stranger statement, and it is also the one for which no explicit instance has been exhibited in physics: to point at a specific Hamiltonian whose gap is independent of ZFC you would need the explicit d that nobody has extracted.

Frontier Position five: what does this do to the named open problems the abstract invokes? The Yang–Mills mass gap is a Millennium Prize problem, the Haldane conjecture is a long-standing question about spin chains, and both are instances of the general spectral gap problem. Established Undecidability of the general problem says nothing whatever about either instance. Speculative A specific, fully specified Hamiltonian is not an infinite family, and a proof for it may well exist; the theorem forbids a method that works for all Hamiltonians, not a proof that works for one. Handwave Reading the 2015 result as evidence that the Yang–Mills gap may be unprovable is the single commonest technical error made about it, and this brief takes the position that the inference is simply invalid.

Speculative Position six: is the fluid result about fluids? The Euler construction is an exact stationary solution and the trajectories really are undecidable, which is a genuine theorem about a genuine partial differential equation. Frontier But the manifold and the metric are chosen to make the construction work: the 2023 universality theorem extends any non-autonomous flow on a compact manifold to a stationary Euler solution “on some Riemannian manifold”, and the 2026 viscous version needs three-manifolds admitting nonvanishing harmonic one-forms and changes the metric in the process. Handwave Whether a Turing complete Euler flow exists on the round three-sphere with its standard metric is stated as an open problem by the people who proved the theorems, which is the correct distance to keep from the phrase “fluid computers”.

Speculative Position seven: is there a Rice-type theorem for many-body systems? Rice’s 1953 theorem says every non-trivial semantic property of a program is undecidable, and the obvious conjecture is that every non-trivial bulk property of a local Hamiltonian family is too. Frontier The evidence is suggestive rather than complete: the gap, the existence of algebraically decaying ground-state correlations, the phase diagram, thermalisation and the renormalisation flow have each been shown undecidable or uncomputable separately, by separate constructions. Handwave No general theorem of that shape has been published for translation-invariant local Hamiltonians, and until one is, the field is proving a list rather than a principle.

Frontier Position eight, held mainly outside the field: the whole thing is an artefact of idealisation and means nothing. The review concedes the premise — the undecidability lives in the model, the idealised limits are not experimentally accessible, and any finite system is decidable by enumeration. Established The rebuttal that survives is narrow and real: the theorem bounds what a general-purpose algorithm can be, and general-purpose algorithms are exactly what computational materials science, density functional theory and machine-learned interatomic potentials are trying to supply. Speculative A bound on a method is a fact about the world even when the pathological instance is not.

Frontier The binding bottleneck is a number that nobody has published: the local Hilbert space dimension of any undecidable spectral-gap family. The Nature paper says estimating it “would be cumbersome and certainly involve large exponentials”; the full version says extracting it is “straightforward (if tedious)” and does not do it. Speculative Without that number there is no simulation, no realisation, no quantitative comparison with any material, and no way to say how far the construction sits from anything anybody studies. Handwave It is the one quantity that would convert this subject from a family of theorems into an engineering target, and it has been available for extraction for eleven years.

Established The second bottleneck is fine-tuning, and it now has a proof attached rather than a suspicion. The renormalisation-flow paper states of the original undecidable Hamiltonian that it has “an enormous local Hilbert space dimension and its matrix elements are highly artificially tuned”, and that “Both of these factors are unlikely to be present in naturally occurring Hamiltonians”. Established The July 2026 instability result makes the second half of that concrete for one dimension: an O(ε) one-local term destroys the undecidability by disrupting an energy cancellation. Speculative A construction that depends on exact cancellation is a construction that depends on exact numbers, and no physical system has exact numbers in it.

Frontier The third is that the interesting behaviour is uncomputably far away. The 2020 prediction of a size-driven transition is a prediction about a threshold that can itself be uncomputably large. Speculative A system that is gapped up to size L and gapless above it is a testable object only if L is a number you can write down and a lattice you can assemble; at the threshold the theorem actually produces, it is not an object at all. Handwave The engineering version of this subject requires deliberately choosing a small, computable threshold, which means giving up the undecidability and keeping only the phenomenon.

Frontier The fourth is the thermodynamic limit, which is not an approximation to a large system but a different object. The spectral gap in these theorems is defined as a property of the infinite lattice; the promise distinguishes continuous spectrum above the ground state from a constant lower bound, and neither is a measurement. Established The certification literature exists because of exactly this: the 2024 certified-algorithms paper opens by noting that finite-size scaling “often fails due to the lack of certified convergence bounds for this limit”. Speculative The bottleneck is not that the limit is hard to reach but that the quantity being computed is not defined at any finite size, so no amount of experimental scaling addresses it.

Speculative The fifth is that the positive results only work on classes chosen for tractability, and nobody can say how much of physics those classes cover. The 2026 gap certificates require frustration-freeness. The 2024 certified algorithms get their fast convergence for commuting Hamiltonians in 2D at high temperature. The December 2025 thermalisation theorem requires high effective temperature, translation invariance and no perfect resonances. Frontier Each is a real theorem about a real class. Handwave What is missing is any measure of what fraction of the Hamiltonians people actually care about lands inside those classes, and without it the field cannot say whether undecidability is a boundary condition or a curiosity.

Speculative And the sixth is sociological rather than technical: the results are produced by a handful of groups and consumed by nobody. The same five or six names appear on the 2015, 2020, 2021, 2022, 2025 and 2026 papers. Handwave A theorem that constrains computational materials science is being proved by mathematical physicists and cited almost exclusively by mathematical physicists, and the community it constrains has not changed a single practice because of it.

Established This subject depends on mathematics and lends to physics, and the traffic is almost entirely one-way. Its prerequisites are Hamiltonian complexity — the circuit-to-Hamiltonian mappings, the local Hamiltonian problem and its QMA-completeness, the analysis of history-state constructions — together with aperiodic tiling theory from the 1960s and 1970s and the computable-analysis tradition that produced the wave-equation results. Frontier None of those is being developed for this purpose, and all of them were developed before it.

Frontier One dependency is unusual and is the reason the 2015 proof is a quantum result rather than a classical one. The construction uses quantum phase estimation to read a Turing machine’s input off a continuous Hamiltonian parameter, which is what lets a single translationally invariant family cover infinitely many machine inputs without changing the interaction. Speculative Remove that and you are back to Berger and Robinson, where the undecidability is about tilings and the physics is decoration.

Speculative The dependency that nothing currently supplies is a definition. Every result in section 3 either is or is not about “physically relevant” Hamiltonians, and the phrase has no agreed content. The 2026 rotational-symmetry paper motivates itself by noting that the earlier models “are not invariant under the other symmetries of the lattice, a property which is commonly found in physically relevant cases”, which is a reasonable working criterion and one of several in use. Handwave Until somebody writes down a class and defends it, every claim in this subject about physical relevance is a claim about a set nobody has specified.

Frontier Within this map the dependencies run outward more than inward. Quantum Materials supplies the systems the general algorithm would be an algorithm for, and its own bottleneck — reproducible synthesis and unreported device yield — is an entirely different kind of obstacle, which is worth knowing when someone claims undecidability explains why materials are hard. Speculative Fault-Tolerant Quantum Computing supplies the Hamiltonian-complexity machinery and, eventually, the programmable simulator that section 7 wants. Autonomous Materials Discovery is the programme whose ceiling this brief describes, though not the ceiling it is currently hitting. Frontier Nothing on this map blocks this brief; it is mathematics, and it waits on nobody’s apparatus.

Frontier The commission that produced this page nominated a decisive experiment: build a physical system in the constructed class and show that no finite-size extrapolation predicts its bulk phase. This brief declines it, for two reasons that are themselves results. Established First, nobody can build a system in the constructed class, because the class is specified by matrix elements that are exact algebraic numbers and a local dimension whose estimate would “certainly involve large exponentials”. Established Second, as of July 2026 a system that merely approximated the 1D class would fall out of it: an O(ε) one-local perturbation makes the problem decidable. Speculative The experiment as commissioned is not hard; it is ill-posed, and saying so is the single most useful thing this page can do for the index it feeds.

Speculative Tier one, and the decisive experiment this brief nominates instead: engineer a size-driven quantum phase transition at a designed, computable critical size in a programmable simulator, and show that the bulk phase above the threshold is not predicted by extrapolation from any size below it. Frontier This is the physically realisable residue of the theorem: the 2020 one-dimensional paper predicts chains that are gapped with a non-degenerate classical ground state up to a critical size and gapless with a dense spectrum above it, and the only thing that makes the theoretical version untestable is that its threshold is uncomputably large. Speculative Choose the threshold instead of deriving it, and the object becomes buildable: a trapped-ion or neutral-atom chain of a few tens of sites with engineered multi-level local spaces, measured for a gap at every length and shown to change character at a designed L. Handwave What it would establish is not undecidability — a finite chain cannot be undecidable — but that the physics the undecidability constructions exhibit is a physics real systems can have, which is the step every strong claim in this subject currently skips.

Frontier Tier two is a theorem rather than an apparatus, and it is the result that would most change this brief’s assessment in the other direction: settle stability in two dimensions. Established The 1D family is now known to be fragile under an arbitrarily small one-local perturbation. Speculative The 2D construction is stated to survive arbitrarily small quantum perturbations of classical Hamiltonians, which is a different and weaker robustness claim, and nobody has tested it against the class of perturbation that killed the 1D case. Handwave If 2D is fragile too, every physical reading of this subject collapses to a statement about exactly tuned mathematical objects, and this brief would have to be rewritten around that.

Speculative Tier three: extract or bound the local dimension. The Forum of Mathematics, Pi version says the extraction is “straightforward (if tedious)”; doing it would produce the first number anyone could compare to a real spin system. Frontier The complementary question is a lower bound: below what local dimension is the spectral gap problem for translation-invariant nearest-neighbour chains decidable? Handwave A published pair of bounds would tell the condensed-matter community, for the first time, whether these results sit one order of magnitude or twenty from the models it studies.

Frontier Tier four: push the certification hierarchies until they break, and record where they break. The 2026 semidefinite hierarchy certifies gaps for frustration-free spin systems and subsumes Knabe’s bound with accuracy improvements of several orders of magnitude; the 2024 certified algorithms bound equilibrium observables in the thermodynamic limit with quasi-polynomial cost for commuting 2D Hamiltonians at high temperature. Speculative Running these against families engineered to be near-undecidable would map the boundary between the certifiable and the closed from the inside. Speculative This is the cheapest test in the brief, it needs no hardware, and it is the one most likely to produce a usable answer within five years.

Speculative Tier five belongs to the fluid strand and its authors have already written it down: construct a Turing complete stationary Euler flow on the round three-sphere with its standard metric. Frontier The 2021 construction builds its own Riemannian metric; the 2026 viscous extension needs manifolds with nonvanishing harmonic one-forms and states that “the resulting metric is in general different from the original one”, leaving the round-metric case open. Handwave Until that is settled, the distance between “Euler flows are Turing complete” and “fluids compute” is a chosen metric, and the phrase “fluid computer” is doing work the theorem does not support.

Speculative Tier six: prove or refute a Rice-type theorem for translation-invariant local Hamiltonians. Frontier Gap, correlation decay, phase diagram, thermalisation and renormalisation flow have each been closed by a separate construction; a single theorem covering every non-trivial bulk property would replace a list with a principle, and a refutation would identify the properties that stay computable. Handwave Either outcome is worth more than another instance, and nobody has published either.

Speculative Tier seven is the negative control this field has never run: finite-precision matrix elements. Frontier Every construction specifies exact algebraic entries. What happens to the undecidability when the entries are specified to k bits, as a function of k, is not answered in any source retrieved here, and the review’s own framing — that undecidability needs “infinite precision” somewhere — predicts that it should vanish. Speculative A quantitative version of that prediction would convert a philosophical objection into a theorem.

Established And the experiments that are not tests of this and are regularly reported as though they were. Programmable quantum simulators measuring gaps and phase boundaries in Hubbard and spin models, tensor-network calculations of ground-state phase diagrams, and machine-learned interatomic potentials are all measurements or computations on specific, fully specified systems. Speculative None of them is an instance of an undecidable problem, none of them could be, and their success is not evidence against the theorems any more than their failure would be evidence for them.

Frontier There is no engineering programme here and the page will not invent one; what follows is the specification for the one buildable object in section 7. Everything else in this subject is a theorem, and the correct engineering statement about a theorem is that it has no requirements.

Speculative The tier-one demonstration has a specification, and it is unusual in being modest. A one-dimensional chain of a few tens of sites; a local Hilbert space of at least several levels per site rather than a qubit, because the phenomenon is driven by a marker structure that a two-level system cannot carry; nearest-neighbour couplings held translationally invariant to the precision at which the designed threshold is stable; a spectroscopic or quench-response gap measurement repeated at every chain length from short to past the designed L; and a control family with no threshold, measured identically. Frontier Trapped-ion and neutral-atom platforms already reach those site counts and already engineer multi-level local spaces. Handwave What nobody has published is the design: which family, at which L, with which couplings, and how precisely those couplings must be held. That calculation is the gating item and it is a theory calculation, not a fabrication problem.

Established The other half of the engineering in this subject is computational and it is already deployed. The 2026 gap-certificate hierarchy is a sequence of semidefinite programs whose levels trade cost against tightness and which recovers Knabe’s finite-size bound at low level; the 2024 certified-algorithm hierarchy returns two-sided bounds on equilibrium observables with quasi-polynomial program size in 1/ε for commuting 2D Hamiltonians at high temperature. Frontier These are the working engineering artefacts of this subject: not machines, but solvers that return an interval and a proof rather than a number and a hope. Speculative The neighbouring pages that would consume them are Quantum Materials, where the prediction targets live, and Autonomous Materials Discovery, whose loop currently has no certified step in it at all.

Handwave One engineering claim in circulation deserves to be shot here rather than in section 15: that fluid Turing-completeness is a route to a computing substrate. Frontier The Euler construction needs a purpose-built Riemannian three-manifold and the viscous version changes the metric it starts from. Speculative There is no proposal anywhere in this literature for a device, no estimate of a clock rate, and no claim by any of the authors that one is possible; the phrase “3D fluid computer” is a headline written over a theorem about geodesics.

Established This commission arrived with no nearest existing brief, and after a search of the corpus that is close to correct: no page here owns computability as a constraint on physical theory. Four neighbours share a boundary with it, and they were chosen because each one contains a programme this brief puts a ceiling on or a method it borrows.

Frontier The closest is Quantum Materials, and the relationship is a correction as much as an adjacency. That brief’s binding constraint is reproducible synthesis, device yield and the fact that almost every flagship effect dies above a few kelvin — obstacles with nothing computational about them. Speculative This page owns the different and much narrower claim that even with a perfect Hamiltonian, perfect compute and no fabrication problem at all, there is no general procedure from Hamiltonian to bulk phase. Handwave The two must not be blended: undecidability is not why materials are hard, and anyone using it that way is borrowing a theorem to explain a fabrication yield.

Frontier The second is Fault-Tolerant Quantum Computing, which supplies both the mathematics and, eventually, the apparatus. The circuit-to-Hamiltonian constructions and the local Hamiltonian problem that underpin every proof in section 3 come from that field’s complexity theory, and the programmable simulator the tier-one experiment needs will come from its hardware. Speculative The inheritance runs back the other way too: a family whose bulk phase is not extrapolable from any accessible size is a family a quantum simulator cannot validate against a classical calculation, which is a constraint on the verification story quantum simulation tells about itself.

Frontier The third is Thermodynamics of Computation, and the pairing is exact. That page asks what physics costs a computation; this one asks what computation costs physics. Speculative Both are Category XI, both take a bound from theoretical computer science and ask what it does to a physical programme, and both find that the bound is real and the programme it constrains has not noticed. Handwave The symmetry is worth stating because the two bounds are the only two of their kind in this corpus.

Frontier The fourth is Deep Learning Theory, and it is a methodological neighbour rather than a physical one. That page documents a practice that works alongside a theory that cannot explain it, with bounds that are vacuous where the practice succeeds. Speculative This page documents the mirror image: a theory that proves a hard impossibility alongside a practice that is unaffected by it. Speculative In both cases the interesting quantity is the gap between the worst case the theorem describes and the typical case the practice meets, and in neither case has anyone measured it.

Speculative Three lighter adjacencies, named because readers will look for them. Autonomous Materials Discovery is the programme whose asymptotic ceiling this brief describes, though its present ceiling is characterisation yield rather than computability. Glass Transition Theory is the corpus’s other example of a physical question that has resisted a general theory for decades, and the contrast is instructive: there the obstruction is unknown, here it is proved. Frontier Grand Challenges of Humanity carries the Yang–Mills mass gap and Navier–Stokes as Millennium Prize problems, and both are named in this subject’s primary literature — the first as an instance of the spectral gap problem, the second as the target of the fluid Turing-completeness programme. Handwave Neither has been shown to be undecidable, and the seam between those pages is precisely the place that error gets made.

Established Outside this map the supplying fields are mathematical logic, Hamiltonian complexity, aperiodic tiling theory, computable analysis and contact and cosymplectic geometry. Frontier Not one of them is a physics field, and the one physics community whose practice the results constrain — computational condensed matter — is the community least represented on the author lists.

Established The institutional fact that explains this field is its size: it is roughly six people and their students. Cubitt, Pérez-García and Wolf proved the 2015 result; Cubitt and Pérez-García are on the 2020 one-dimensional paper with Bausch and Lucia; Cubitt is on the 2021 phase-diagram paper with Bausch and Watson and on the 2022 renormalisation paper with Watson and Onorati; all three of the original authors plus Gu are on the 2025 review; and Lucia is on both 2026 papers with Castilla-Castellano. Frontier Eleven years of results in a subject that constrains computational materials science come from two universities and one institute.

Frontier That concentration is efficient and it is a single point of failure. The work is done in mathematics and computer-science departments, funded as mathematical physics, and published in venues — Nature, Forum of Mathematics, Pi, Physics Reports, Quantum — that reach theorists. Speculative There is no programme anywhere whose stated purpose is to determine whether undecidability constrains real materials prediction, and therefore no budget line that would survive the retirement of the people who started it.

Speculative The institutional requirement this subject most obviously lacks is a definition it can be held to. Every advance since 2015 has been argued as a step toward physical relevance — fewer dimensions, continuous parameters, positive measure, lattice symmetry — and the target has never been written down. Frontier The 2026 rotational-symmetry paper comes closest by naming lattice symmetry as “commonly found in physically relevant cases”. Handwave A published, pre-committed class — bounded local dimension, bounded matrix-element precision, stability under one-local perturbations — together with the statement that undecidability inside that class would settle the question, would cost nothing and would convert a sequence of improvements into a programme with a finish line.

Speculative The second requirement is a consumer. The certification results of 2024 and 2026 are usable today by anyone computing ground-state properties, and there is no shared repository of certified bounds, no benchmark suite, and no convention in the condensed-matter literature for reporting whether a computed gap is certified or estimated. Frontier The cheapest institutional intervention available in this subject is a norm: report the certificate, or report that there is none.

Frontier The hazards here are entirely epistemic, and the largest is that this is the most quotable result in mathematical physics and one of the least quotable-accurately. “Physics is undecidable” and “a Goedel theorem for matter” are three words long and wrong; the true statement needs a lattice dimension, a symmetry class, a promise, a thermodynamic limit and the word “family”. Speculative The gap between those two lengths is where the damage happens, and it is not the authors’ doing: the 2015 paper says its models are “highly artificial” in its own text.

Speculative The second hazard is the use of these theorems as an argument against computational science funding, which the theorems do not support. Undecidability of a general problem is compatible with every specific case of interest being easy, and the 2024 and 2026 certification results are direct evidence that useful, rigorous computation continues inside the closed region. Frontier A field that cannot solve every instance is the normal condition of computation, not a reason to stop. Handwave Anyone deploying “it is undecidable” as a budget argument is using a theorem about an infinite family to make a claim about a finite programme.

Speculative The third runs the opposite way and is more common in physics than outside it: using undecidability to license mysticism about consciousness, free will or the limits of science. The Goedel-style corollary is real and it is narrow: it says that for some Hamiltonian in a constructed class, the gap is independent of a given axiom system. Frontier It does not say that nature is uncomputable, that minds exceed machines, or that there are physical facts beyond investigation. Handwave The corpus’s own Orch OR page documents where that style of argument leads when nobody checks the step, and this subject is structurally exposed to the same failure because its result genuinely does involve Goedel.

Established Fourth, and to the field’s credit, the correction has come from inside. The 2022 renormalisation paper says the original Hamiltonian’s features are “unlikely to be present in naturally occurring Hamiltonians”. The 2025 review says undecidability “is not a feature of the physical system”. The July 2026 instability paper, by an author of the 2026 strengthening, shows the one-dimensional result is fragile. Frontier A community that publishes the limits of its own headline theorem, in the same journals and by the same people, is behaving correctly, and it should be said plainly because the public record of this subject looks nothing like that.

Speculative The concrete consequence, if the physical reading survives, is a permanent bound on one very large ambition: a general procedure that takes a description of matter and returns its behaviour. That ambition is the implicit business case for computational materials science, for much of quantum simulation, and for the newer claim that machine learning will close the loop from composition to property. Frontier The theorems say the general version of that procedure does not exist, and they say it about a class of models that looks like physics rather than like logic. Handwave They do not say how much of the useful space is affected, and nobody has estimated it.

Speculative The second consequence is a change in what a scientific answer is allowed to look like, and it is already underway. When a general solver is impossible, the defensible output is a certificate: a two-sided bound with a proof of convergence on a stated class, rather than a number with an error bar from a method that has worked before. Established Both major positive results of the last two years are of exactly that shape. Frontier A discipline that migrates from estimates to certificates is a discipline that has absorbed an impossibility theorem, and that migration is visible in this literature and nowhere else in computational physics.

Speculative The third is about what a final theory could mean, and it needs to be stated carefully because the careless version is everywhere. A complete set of fundamental laws could be short, exact and known, and questions about the behaviour of systems governed by those laws could still have no general algorithm. Frontier Completeness of a theory and computability of its consequences are independent properties, and the undecidability results are the cleanest demonstration anyone has that they come apart. Handwave The claim that goes further — that the theory of everything may itself be uncomputable — is a different assertion about the laws rather than about their consequences, no result in this literature supports it, and this page does not make it.

Speculative And the outcome this brief judges most likely: the stability results keep arriving, two dimensions goes the way one dimension went in July 2026, and undecidability settles into its correct place as a statement about idealised limits with a real but narrow corollary for algorithm design. Frontier On that path the lasting value of eleven years of theorems is the certification literature they provoked, which is not nothing and is much less than the headlines. Handwave The alternative — that a stable, low-dimensional, physically realisable undecidable family is found — would be one of the more interesting facts about the universe, and there is currently no evidence for it beyond the absence of a proof against.

These horizons track theorems and certification tooling, because there is no experimental programme in this subject to track and section 7 explains why:

• 10 yr: Frontier Expect the stability question to be settled in two dimensions, in one direction or the other, and expect it to be the result that fixes how the subject is described for a generation. Speculative Expect the local dimension of at least one undecidable family to be extracted and published, because the 2022 full version says it is straightforward and the number is now the most-wanted quantity in the field. Speculative Expect the certification hierarchies to be extended past frustration-freeness and past the commuting case, and expect the first condensed-matter papers to report certified rather than estimated gaps. Handwave Expect no experiment.
• 25 yr: Speculative This is the window in which the tier-one demonstration plausibly runs, if anyone designs it: a chain with a deliberately chosen, computable critical size, built on a platform developed for Fault-Tolerant Quantum Computing. Speculative A Rice-type theorem for translation-invariant local Hamiltonians, or a proof that no such theorem holds, is a 25-year object rather than a 10-year one. Handwave If the subject is going to reach Quantum Materials practice at all, it reaches it here, through certified solvers rather than through impossibility.
• 50 yr: Handwave Beyond the reach of anything in section 3. The defensible statement is conditional: if a stable, low-dimensional undecidable family exists, it will have been found by then and computational physics will have a named boundary; if it does not, the subject will be a chapter in the history of Hamiltonian complexity and a paragraph in every textbook’s introduction. Speculative The fluid strand resolves earlier or not at all, because the open problem its own authors state — the round metric on the three-sphere — is a concrete question with a small number of people working on it.
• 100 / 250+ yr: Handwave Not forecastable, and the honest form of the statement is about shape rather than date: a theorem does not decay, so whatever is proved in this subject is still true in 2276, and the only thing that changes is whether anyone found a physical system it applies to. Handwave Any claim about what an undecidable material would be used for is an assertion, and this page does not make one.

• Depends on Nothing on this map blocks this brief, and the absence is a fact about the subject rather than an omission: it is mathematics and it waits on no apparatus. Fault-Tolerant Quantum Computing is the nearest thing to a prerequisite, supplying the Hamiltonian-complexity constructions every proof in section 3 uses and, eventually, the programmable simulator the tier-one demonstration needs; but that field would build both whether or not anyone cared about undecidability. Quantum Materials supplies the systems a general algorithm would be an algorithm for, and is not blocking either.
• Requires (not on this map) Six conditions, in the order of the tokens above. The dimension. Eleven years after the theorem, no published number for d in any spectral-gap construction; the Nature paper says an estimate would “certainly involve large exponentials” and the full version calls the extraction straightforward but tedious, so the quantity that would let anyone compare these models to a real spin system does not exist. The stability theorem. In one dimension an O(ε) one-local perturbation was shown in July 2026 to restore decidability; the same question in two dimensions is open, and the answer determines whether this subject is about physics or about exactly tuned mathematics. The Rice-type theorem. Gap, correlation decay, phase diagram, thermalisation and renormalisation flow have each been closed separately; a single theorem over all non-trivial bulk properties would replace a list with a principle. The round metric. The fluid results build their own Riemannian manifold; whether a Turing complete Euler flow exists on the standard round three-sphere is stated as open by the authors, and it is the difference between a theorem about geodesics and a claim about fluids. The chain. A programmable one-dimensional simulator of a few tens of sites with several levels per site and translationally invariant couplings held to a stated precision — within reach of trapped-ion and neutral-atom platforms, with no published design to build. The definition. A pre-committed, published class of “physically relevant” Hamiltonians, with the statement that undecidability inside it would settle the question; the cheapest item on this list and the one whose absence lets every result be argued about indefinitely.
• Enables Nothing is enabled on a schedule, and a page that claimed otherwise would be lying about a theorem. What the results have already produced is the certification literature: two-sided, proof-carrying bounds on equilibrium observables and spectral gaps, which Quantum Materials and Autonomous Materials Discovery could consume today and do not. A settled two-dimensional stability result would enable the rest of this page to be written with confidence in one direction or the other.
• Adjacent Within this map: Thermodynamics of Computation for the mirrored question of what physics costs a computation, Deep Learning Theory for the same worst-case-versus-typical-case gap in a different subject, Glass Transition Theory for a physical question whose obstruction is unknown rather than proved, and Grand Challenges of Humanity for the Millennium Prize problems this subject’s literature names and does not settle. Outside it: mathematical logic, Hamiltonian complexity, aperiodic tiling theory, computable analysis, and contact and cosymplectic geometry.

Established “Physics is undecidable.” The theorems are about families of Hamiltonians constructed for the proof. Established The 2015 paper describes its own models as “highly artificial” and says whether the results extend to more natural models “is yet to be determined”; the 2022 renormalisation paper says the enormous local dimension and the artificially tuned matrix elements are “unlikely to be present in naturally occurring Hamiltonians”; the 2025 review says undecidability “is not a feature of the physical system; it is a feature of the mathematical model”. Handwave Undecidability of a constructed family is a theorem; undecidability of physics is not. Every strong claim in circulation jumps that gap, and the jump is the entire content of this section.

Frontier “But it does not matter, then.” This is the opposite error and it is made by the same people about a week later. Established The results constrain what a general algorithm from Hamiltonian to bulk property could ever be, and that programme is real, funded and currently promising the general version. Speculative The honest reading is narrow and load-bearing at once: no universal solver, so certificates on stated classes — which is exactly what the 2024 certified-algorithm and 2026 gap-certificate papers deliver, and exactly why they cite the impossibility results in their opening paragraphs.

Frontier “The Yang–Mills mass gap has been shown to be unprovable.” No. Established The Nature abstract names the Yang–Mills gap conjecture as an instance of the general spectral gap problem, and the general problem is undecidable. Established A single fully specified instance is not an infinite family, and undecidability of a problem places no restriction at all on the provability of any one instance of it. Speculative The same correction applies to the Haldane conjecture and to every other named conjecture the abstract lists.

Frontier “The theory of everything may be uncomputable.” Speculative Two different claims are welded together here. That the consequences of a known, exact theory may have no general algorithm is what these results show. That the laws themselves might not be computably specifiable is a separate assertion about fundamental physics, with no support anywhere in this literature. Handwave The first is a theorem about a family of spin lattices; the second is a speculation about the universe, and they are routinely printed in the same sentence.

Frontier “It is just chaos with better branding.” Established The 2022 renormalisation paper addresses this directly: the flow it constructs has computable steps and correct fixed points and is nonetheless uncomputable, which the authors describe as going “beyond the chaotic behaviour seen previously”. Speculative Chaos is sensitivity to initial conditions; here the initial conditions are exact and the trajectory is still unpredictable, with two trajectories staying arbitrarily close for an uncomputable number of steps before diverging to fixed points in different phases. Frontier They are different failures of prediction and conflating them loses the only thing that is new.

Speculative “Fluids can compute, so water is a computer.” Established The 2021 result is a stationary Euler flow on a Riemannian three-sphere whose metric is part of the construction; the 2026 viscous extension needs a three-manifold with nonvanishing harmonic one-forms and states that “the resulting metric is in general different from the original one”. Frontier Whether the construction survives on the round three-sphere is an open problem its own authors state. Handwave “Undecidable particle paths” is accurate; “3D fluid computers” is a headline; and no author in this literature has proposed a device.

Frontier “This kills density functional theory / machine-learned potentials / materials AI.” Speculative It does not, and the evidence that it does not is that the certification results postdate the impossibility results and are stronger for them. Established Local observables in equilibrium states can be approximated in finite time with rigorous two-sided bounds, and that was proved in 2024 explicitly as a contrast to the undecidability literature. Handwave The correct statement is that a method which works on the instances people care about is not refuted by a theorem about instances nobody has ever met, and the burden it does carry is to say which class it is claiming.

Frontier “A finite system can be undecidable.” Established It cannot: any finite problem is decidable by enumeration, and the review says so in as many words, adding that an undecidable problem “must necessarily conceal an infinite somewhere: infinitely many instances, infinitely many particles, infinite precision”. Speculative What a finite system can have is behaviour that no accessible finite-size extrapolation predicts, which is the observable residue and the thing section 7 asks somebody to build.

Handwave “Goedel has been proved for the physical world.” Established The axiomatic-independence corollary is genuine: given any consistent recursive axiomatisation, there exist Hamiltonians in the constructed class whose gap is not determined by those axioms. Frontier But no explicit such Hamiltonian has been exhibited, because exhibiting one requires the local dimension nobody has extracted, and the corollary is about a mathematical model of a lattice, not about the world. Speculative Extensions of it to consciousness, free will or the limits of science are unsupported by anything in section 3.

Frontier And the correction this page makes to its own commission. Established The anchor named for this brief was “the proof that the spectral gap of a translationally invariant 2D lattice model is undecidable, with explicit constructions”. The proof is correctly identified and the scope is not: it is a family of Hamiltonians rather than a model, the undecidability is a property of the decision problem over that family rather than of any member, and it holds in the thermodynamic limit under a promise that each member is gapped or gapless in the strongest sense. Established “With explicit constructions” is the part that is wrong: the construction is explicit as a method and the model is not explicit as a number, and both papers say so. Frontier The anchor is also five results out of date — one dimension in 2020, positive measure in 2021, rotational symmetry in January 2026 — and, most importantly, it predates the July 2026 instability result, which is the first serious evidence that the physical reading of the 1D theorem does not survive. Speculative The commissioned decisive experiment inherits the same defect, which is why section 7 replaces it rather than costing it.