Unsolvability Ceiling in Theory and Practice

• Unsolvability ceiling is a boundary concept where formal systems or models fail to produce full exact solutions, marking limits across analytic, arithmetic, and computational methods.
• It spans domains—from quantum field theory’s combinatorial explosion and Diophantine arithmetic obstructions to LLM evaluation artifacts—illustrating various structural barrier mechanisms.
• These ceilings illuminate fundamental limits in solving complex problems, prompting research to explore approximate methods or conditional solutions where complete solvability is unattainable.

Searching arXiv for papers using the term and closely related formulations. “Unsolvability ceiling” denotes a boundary beyond which a formalism, model class, or evaluation regime ceases to admit a full exact solution, a stable object, a decidable procedure, or a reliable selection rule. The expression is not used uniformly across fields. In interacting bosonic quantum field theory it names a structural barrier to exact non-perturbative solution; in Erdős–Moser type Diophantine equations it names arithmetic obstructions that force nonexistence of integer solutions beyond trivial cases; in recent LLM research it names ceilings on selection accuracy, benchmark estimation, and router performance induced by model capability, correlation, and evaluation artifacts. Related notions appear in classification theory, stable roommates, planning, nested recursions, and tag systems as boundaries between solvable and structurally irreducible regimes [(Aglietti, 2018); (Baoulina, 2018); (Bay et al., 27 Jun 2026); (Garg et al., 8 May 2026); (Walter et al., 2014); (Glitzner et al., 10 May 2025); (Sarwar et al., 22 Apr 2025); (Sreedharan et al., 2019); (0906.3329); (Isgur et al., 2012)].

1. Conceptual scope

Across these literatures, the ceiling is not a single theorem but a recurrent type of claim. In some settings it concerns exact analytic solvability: one asks whether all observables can be written in closed non-perturbative form, and the ceiling is reached when the exact problem expands into an infinite-dimensional hierarchy. In others it concerns existence: whether a stable matching, a stable partition of a classification problem, or an integer solution exists at all. In a third group of cases it concerns selection and evaluation: correct outputs may be generable, yet the system cannot identify them, or the measured unsolved fraction may be inflated by the evaluation protocol itself (Aglietti, 2018, Glitzner et al., 10 May 2025, Bay et al., 27 Jun 2026, Garg et al., 8 May 2026).

A common misconception is to equate “unsolvability” with nonexistence of the underlying object. The bosonic QFT analysis is explicit that its target is not existence of the theory, but the impossibility or extreme implausibility of a full exact analytic solution for all correlators. Conversely, in Stable Roommates, unsolvability means literal absence of a stable matching, even though related objects such as stable partitions or maximum stable matchings still exist. In LLM routing, “unsolvability ceiling” can be empirical rather than intrinsic, because the unsolved fraction depends on what counts as a correct answer and how that correctness is judged (Aglietti, 2018, Glitzner et al., 10 May 2025, Garg et al., 8 May 2026).

This suggests that the phrase is best understood as a family of structural boundary claims. The relevant boundary may live in function space, number-theoretic congruence classes, abstraction lattices, matching structures, or evaluation pipelines.

2. Exact analytic ceilings in interacting theories

In the bosonic QFT setting, the ceiling is formulated for Euclidean bosonic theories such as the anharmonic oscillator and scalar $\lambda \phi^4$ on a lattice. The basic observables are lattice correlators

$$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$

and “solving the theory” means knowing all such $G(\nu)$ exactly in non-perturbative closed analytic form. On a finite lattice, the lattice Dyson–Schwinger equations reduce every correlator to a finite linear combination of primitive correlators $P(\mu)$ with $\mu_i\in\{0,1,2\}$, so the hierarchy closes algebraically and exactly. For quartic interaction, however, the number of primitives is $3^N$, and more generally scales as approximately $(m_{\mathrm{anh}}-1)^N$ for polynomial interaction of maximal power $m_{\mathrm{anh}}$; symmetries reduce this only by power-like factors. In the physically relevant weak continuum limit, the primitive basis becomes countably infinite, and any exact analytic method must solve a countable system of coupled linear PDEs for countably many primitive correlators in countably many coupling variables. The paper characterizes this as a structural obstruction short of a formal impossibility proof, and treats free theories and trivial random fields as exceptional cases where the primitive basis collapses to a finite set (Aglietti, 2018).

The nested-recursion literature reaches a different but related ceiling. It studies recursions of the form

$$R(n)=\sum_{i=1}^k R\!\left(n-s_i-\sum_{j=1}^{p_i}R(n-a_{ij})\right)+\nu,$$

and asks when the solution can be a ceiling function $\left\lceil rn/q\right\rceil$. A finite decision procedure is obtained by reducing formal satisfaction to checking $$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$0 and by quotienting parameter vectors under explicit equivalence relations. The main positive theorem is that every $$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$1 is the solution of infinitely many such recursions. The main conjecture is the converse: if $$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$2 is generated by any recursion in the class, then $$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$3. This suggests a slope ceiling: the family appears to admit ceiling solutions only at slopes $$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$4, not at $$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$5 with $$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$6 (Isgur et al., 2012).

In both cases, solvability fails not because the formalism lacks expressive power, but because exact closure produces an object of prohibitive structural size: an infinite primitive basis in one case, and a sharply restricted admissible parameter geometry in the other.

3. Arithmetic ceilings in Erdős–Moser type equations

In the arithmetic setting, the ceiling is literal nonexistence of integer solutions. The paper studies

$$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$7

as the odd-sum analogue of the Kellner–Erdős–Moser equation. The first barrier is parity: if $$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$8, then any solution must have even $$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$9. The deeper barriers come from Bernoulli numbers, irregular primes, and the “helpful pairs” machinery. If $G(\nu)$0, then $G(\nu)$1 must be an irregular prime, $G(\nu)$2 must lie in congruence classes determined by irregular pairs, and one has

$G(\nu)$3

These local conditions lead to strong global nonexistence criteria (Baoulina, 2018).

The central global statement is Theorem 1: if $G(\nu)$4 is even, or $G(\nu)$5 has a regular prime divisor, or $G(\nu)$6, then

$G(\nu)$7

has no solutions with $G(\nu)$8. For $G(\nu)$9, Theorem 2 does not prove impossibility but forces any hypothetical solution to have an enormous exponent $P(\mu)$0, divisible by

$P(\mu)$1

and requires every prime divisor of $P(\mu)$2 to exceed $P(\mu)$3. The associated conjecture is that for $P(\mu)$4 the ratio $P(\mu)$5 is never an integer. Here the ceiling is not heuristic in the QFT sense: it is a web of congruence and valuation obstructions that already proves nonexistence for broad parameter ranges (Baoulina, 2018).

A plausible implication is that arithmetic ceilings are often sharper than analytic ones. The obstruction is encoded in congruence classes, divisibility, and Bernoulli numerators rather than in infinite-dimensional functional systems.

4. Decidability ceilings and structural cores

In classification theory, the ceiling is formulated through unsolvability cores. A classification problem $P(\mu)$6 is a $P(\mu)$7-core of a family $P(\mu)$8 if every subproblem with more than one component remains $P(\mu)$9-unsolvable. The main characterization is that, for nontrivial $\mu_i\in\{0,1,2\}$0, one has

$\mu_i\in\{0,1,2\}$1

Thus cohesiveness is the structural form of the ceiling: the union of the components cannot be split into two infinite $\mu_i\in\{0,1,2\}$2-definable pieces. Unlike the promise-problem case, unsolvable classification problems with more than two components need not possess cores. The paper therefore introduces conditional classification problems and conditional cores, and relates conditional $\mu_i\in\{0,1,2\}$3-cores to proper hard cores; for WP-recursive language families with Boolean closure, conditional cores with recursive components exist (Walter et al., 2014).

Tag systems provide a classic computability-theoretic boundary. A tag system has alphabet size $\mu_i\in\{0,1,2\}$4, deletion number $\mu_i\in\{0,1,2\}$5, and appendants of lengths between $\mu_i\in\{0,1,2\}$6 and $\mu_i\in\{0,1,2\}$7. The basic frontier facts are sharp: for $\mu_i\in\{0,1,2\}$8, halting and reachability are decidable; for every $\mu_i\in\{0,1,2\}$9, there exists a tag system with an unsolvable decision problem. Wang proved decidability if $3^N$0 or $3^N$1. Universality already appears for $3^N$2, $3^N$3, and $3^N$4. At the small end, TS$3^N$5 and TS$3^N$6 are decidable, whereas TS$3^N$7 contains the Collatz tag system and TS$3^N$8 contains Post’s system, both treated as boundary cases. The paper also reports a structural distinction: TS$3^N$9 cannot produce the irregular periodic types labeled 2 and 4, whereas TS$(m_{\mathrm{anh}}-1)^N$0 and TS$(m_{\mathrm{anh}}-1)^N$1 can (0906.3329).

Taken together, these results present two versions of the same idea. One is semantic: no algorithmic solution exists beyond certain parameter thresholds. The other is structural: some unsolvable regions are cohesive or core-like, while others lack cores and require conditionalization.

5. Local obstructions in matching and planning

In Stable Roommates, solvability means existence of a stable matching. Tan’s stable-partition theory identifies the exact obstruction: an instance admits a stable matching iff no stable partition contains an odd cycle of length at least $(m_{\mathrm{anh}}-1)^N$2. The paper reviews the long-standing question, posed by Gusfield and Irving, of the asymptotic solvability probability $(m_{\mathrm{anh}}-1)^N$3 for random instances. For impartial culture, earlier work by Pittel and Mertens is extended empirically: for even $(m_{\mathrm{anh}}-1)^N$4, the paper fits

$(m_{\mathrm{anh}}-1)^N$5

while for odd $(m_{\mathrm{anh}}-1)^N$6 it reports the steeper decay

$(m_{\mathrm{anh}}-1)^N$7

At the same time, unsolvability remains localized. The number of agents lying on odd cycles is small, 3-cycles dominate, and the maximum stable-matching ratio $(m_{\mathrm{anh}}-1)^N$8 stays extremely close to $(m_{\mathrm{anh}}-1)^N$9 even when $m_{\mathrm{anh}}$0 is effectively $m_{\mathrm{anh}}$1; for example, at $m_{\mathrm{anh}}$2 the paper reports $m_{\mathrm{anh}}$3 for IC and $m_{\mathrm{anh}}$4 for Attributes. The ceiling here is therefore an existence ceiling for exact stability, but not for near-stable structure (Glitzner et al., 10 May 2025).

Planning work turns the ceiling into an explanatory object. In hybrid planning, a planning problem is unsolvable if it has no valid plan, but the explanation is sought through inevitable waypoints: sub-problems whose goals appear on every discrete path from source to goal. The paper extracts these waypoints by computing a longest common subsequence over all bounded paths in the hybrid automaton graph, then performs symbolic reachability on the resulting chain and returns the earliest unreachable waypoint as the explanation. In classical planning with plan advice, unsolvability is analyzed in an abstraction lattice: explanatory fluents move the human model to the most abstract unsolvable node, landmarks are extracted one level up where plans still exist, and the first landmark that becomes unachievable under the refined model explains why the problem cannot be solved. In both formulations, the ceiling is the first necessary subgoal that cannot be crossed (Sarwar et al., 22 Apr 2025, Sreedharan et al., 2019).

This suggests a general explanatory pattern: local obstructions can be more informative than global impossibility claims. Odd cycles, unreachable waypoints, and failed landmarks all identify where solvability breaks.

6. Modal, correlation, and routing ceilings in LLM systems

Recent LLM work uses “unsolvability ceiling” in a markedly different sense. In test-time scaling by repeated sampling, the key distinction is between coverage and selection accuracy. For a problem $m_{\mathrm{anh}}$5, one draws $m_{\mathrm{anh}}$6 and defines $m_{\mathrm{anh}}$7. Coverage is

$m_{\mathrm{anh}}$8

and increases with $m_{\mathrm{anh}}$9. A deployed system, however, must select one answer. Under self-consistency or plurality, selection accuracy converges not to coverage but to the modal ceiling

$$R(n)=\sum_{i=1}^k R\!\left(n-s_i-\sum_{j=1}^{p_i}R(n-a_{ij})\right)+\nu,$$0

where $$R(n)=\sum_{i=1}^k R\!\left(n-s_i-\sum_{j=1}^{p_i}R(n-a_{ij})\right)+\nu,$$1 is the modal answer. The paper also defines the correlation ceiling through the effective number of samples

$$R(n)=\sum_{i=1}^k R\!\left(n-s_i-\sum_{j=1}^{p_i}R(n-a_{ij})\right)+\nu,$$2

which saturates at $$R(n)=\sum_{i=1}^k R\!\left(n-s_i-\sum_{j=1}^{p_i}R(n-a_{ij})\right)+\nu,$$3 as $$R(n)=\sum_{i=1}^k R\!\left(n-s_i-\sum_{j=1}^{p_i}R(n-a_{ij})\right)+\nu,$$4. Empirically, self-consistency on MATH-500 plateaus around $$R(n)=\sum_{i=1}^k R\!\left(n-s_i-\sum_{j=1}^{p_i}R(n-a_{ij})\right)+\nu,$$5 by $$R(n)=\sum_{i=1}^k R\!\left(n-s_i-\sum_{j=1}^{p_i}R(n-a_{ij})\right)+\nu,$$6, while coverage at $$R(n)=\sum_{i=1}^k R\!\left(n-s_i-\sum_{j=1}^{p_i}R(n-a_{ij})\right)+\nu,$$7 is about $$R(n)=\sum_{i=1}^k R\!\left(n-s_i-\sum_{j=1}^{p_i}R(n-a_{ij})\right)+\nu,$$8; the gap is the identifiability gap, the set of problems the model can generate but not pick. In Brown et al.’s GSM8K and MATH logs, estimated intraclass correlations $$R(n)=\sum_{i=1}^k R\!\left(n-s_i-\sum_{j=1}^{p_i}R(n-a_{ij})\right)+\nu,$$9–$\left\lceil rn/q\right\rceil$0 imply that even $\left\lceil rn/q\right\rceil$1 samples per problem are worth only about two independent samples for estimating benchmark means (Bay et al., 27 Jun 2026).

Multi-LLM routing uses the term differently again. There the unsolvability ceiling is the fraction of queries for which no model in the pool attains score $\left\lceil rn/q\right\rceil$2 under the evaluation metric. In the Gemma-4 study, the pool comprises four tiers and $\left\lceil rn/q\right\rceil$3 queries, giving $\left\lceil rn/q\right\rceil$4 query-model pairs; the primary judge yields a global unsolvable fraction of $\left\lceil rn/q\right\rceil$5, hence an empirical ceiling of $\left\lceil rn/q\right\rceil$6. The paper argues that much of this apparent ceiling is an artifact of evaluation: judge bias, truncation under fixed output budgets, and output-format mismatch. On MMLU, judge-derived unsolvability is $\left\lceil rn/q\right\rceil$7 whereas exact-match implies about $\left\lceil rn/q\right\rceil$8; on MedQA, judge-derived unsolvability is $\left\lceil rn/q\right\rceil$9 whereas exact-match implies about $$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$00. These artifacts also corrupt router labels, and standard routers collapse to majority-class prediction, with about $$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$01 of Alpaca queries labeled E2B-optimal and a reported opportunity cost of $$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$02–$$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$03 percentage points on knowledge benchmarks (Garg et al., 8 May 2026).

The two LLM usages jointly separate three ceilings: a generation ceiling when the model never outputs a correct answer, a selection ceiling when the correct answer is not identifiable from samples, and an evaluation ceiling when the measured unsolved mass is itself distorted by the benchmark protocol.

7. Comparative interpretation and limitations

Taken together, these works suggest that “unsolvability ceiling” names several recurring structures rather than a single doctrine. One structure is combinatorial explosion: the finite lattice closure of bosonic correlators becomes countably infinite in the continuum, and exact methods inherit a countably infinite PDE system (Aglietti, 2018). A second is local obstruction: odd cycles in Stable Roommates, the first unreachable waypoint in hybrid planning, and the first unachievable landmark in advised planning are all localized barriers that certify failure of the whole task (Glitzner et al., 10 May 2025, Sarwar et al., 22 Apr 2025, Sreedharan et al., 2019). A third is metric dependence: routing ceilings can be inflated or deflated by the judge, truncation budget, or output parser, so the ceiling may be epistemic rather than intrinsic (Garg et al., 8 May 2026).

The literature is also careful about exceptions. Free Gaussian theories, random fields with $$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$04, symmetric and Euclidean roommates cultures, and certain conditional-core constructions show that strong symmetry or factorization can collapse the obstruction. Conversely, some unsolvable classification problems with more than two components possess no cores at all, so a ceiling need not always admit a minimal irreducible witness [(Aglietti, 2018); (Glitzner et al., 10 May 2025); (Walter et al., 2014)].

A final caution is that many of these ceilings are not formal impossibility proofs. The bosonic QFT argument is explicitly a structural obstruction short of proof. The nested-recursion result proves infinite families for $$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$05 but leaves the $$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$06 barrier as conjecture. The tag-systems survey places TS$$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$07 and TS$$G(\nu)=\langle \Phi^\nu\rangle =\int_{\mathbb{R}^N}D\Phi\,\Phi^\nu e^{-S[\Phi]}, \qquad \Phi^\nu=\prod_{i=1}^N \phi_i^{\nu_i},$$08 on a frontier without resolving their global status. By contrast, the Erdős–Moser-type results and several decidability theorems are genuine nonexistence or undecidability statements in defined parameter ranges [(Aglietti, 2018); (Isgur et al., 2012); (0906.3329); (Baoulina, 2018)].

In that qualified sense, the term denotes a precise but domain-dependent insight: once a problem crosses a certain structural boundary, additional effort no longer yields an exact global solution within the same formal regime. What changes from field to field is the locus of that boundary—primitive correlators, congruence classes, cohesive sets, odd cycles, waypoints, modal answers, or evaluation artifacts.