Do cognitive constraints drive mathematical abstraction through bottleneck search?

Limits become pressure to restructure knowledge

The paper's answer to its title question is yes, under a specific condition: a cognitive limit matters when it makes continued work inside the current representation so costly that searching for a new representation becomes the better use of computation. The author calls this a resource-rational search for bottlenecks. Resource rationality here means treating limited working memory, fragile memory retrieval, reward discounting, boredom and frustration as adaptations to finite metabolic resources and limited biological processing speed, rather than as pure defects.

The proposed pattern has two phases. Between methodological transitions, mathematicians explore new compositions of the existing toolkit until they hit a bottleneck. A bottleneck is either an ill-posed problem or a well-posed problem that human cognitive limits make hard to resolve. In the second case, the paper argues, mathematicians respond by generating formalisms that make the problem tractable for humans, while unresolved problems remain open for decades.

This is a theory paper, not an experiment. Its evidence is a set of episodes from the history of mathematics plus a formal account of when representational change is worth its cost. That matters for reading every result below: the cases are used to illustrate a mechanism, not to estimate how often the mechanism operates.

Working memory limits favor hierarchical compression

The clearest quantitative argument concerns working memory. Suppose a problem requires searching among N primitive concepts. With a flat representation, where each concept must be considered individually, search complexity scales linearly as O(N). If working memory can hold only C concepts at a time, the concepts can instead be grouped into a hierarchy with d levels, where d is at least log base C of N. Searching across those levels and then among at most C concepts at the final stage yields complexity O(C log base C N).

When C is finite and small, the paper notes that O(C log base C N) is contained in O(N), so the hierarchy can greatly reduce computational cost. The author uses this to explain why natural cognitive limits create pressure to organize concepts into computationally efficient hierarchies. The same section places this against the brain's roughly 20 W power budget, which the paper says is far from the resources of large-scale computing.

The step from the complexity comparison to mathematical history is the author's interpretation. The formal result shows that imposing a capacity constraint and encouraging hierarchical abstraction can reduce search complexity; it does not show that working memory limits caused any particular historical formalism.

Tedium, frustration and planning failure signal low returns

The paper formalizes three bottleneck types. Tedium arises when long sequences of repetitive operations accumulate execution costs for little additional progress. Frustration arises when repeated unsuccessful attempts accumulate search costs without the expected progress. Planning failure occurs when no solution is found within the available planning horizon, indicating that further search requires committing more computational resources.

These are unified through a local return on computation. If step t costs c_t and produces progress ΔV_t toward the goal, the return is η_t = ΔV_t / c_t. A bottleneck is a sustained decline in this return: computation consumes resources without commensurate progress. The three signals are described as different sources of evidence about one latent quantity, the expected cost of continuing under the current representation, rather than as independent determinants of behavior.

The paper pairs each signal with historical material. It quotes Napier's preface on the tedium of lengthy multiplications, divisions and root extraction as motivation for logarithms, and George Temple's description of mathematical progress as frustration at insolubility followed by the invention of more general concepts and methods. These quotations document the author's proposed link; they are not measurements of the emotions or of their causal effect.

Switching representations is a cost comparison

The decision to change representation is written as an inequality. An agent can keep searching under its current repertoire R, with expected remaining cost E[C_solve | X, R], or spend resources discovering a new abstraction r that expands the repertoire to R ∪ {r}. Representational change is worthwhile when the expected discovery cost plus the subsequent solving cost under the expanded repertoire is less than the expected cost of continuing under the current one: E[C_discover(r) + C*(X; R ∪ {r})] < E[C_solve | X, R].

The paper is explicit that bottlenecks do not themselves generate abstractions. They provide evidence that continued computation under the current representation has low expected value, which raises the relative value of searching for a new representation. A successful abstraction changes the computational structure of the motivating problem, so C*(X; R ∪ {r}) < C*(X; R).

Broader value depends on reuse. If the same abstraction lowers solving costs across a distribution of problems drawn from p(X), then the expected reduction across that distribution is positive. This is the paper's account of why abstractions can repay their discovery cost beyond the problem that prompted them, and why abstractions with broad compositional utility are especially valuable to a resource-limited agent.

Compression, new spaces and relations in practice

The paper groups historical breakthroughs into three forms. The first is compression in symbolism: repeated addition such as seven copies of 2 can be written as 2 × 7, and a system of m linear equations in n unknowns can be compressed into the matrix equation Ax = b. The matrix form treats the whole system as one relation among A, x and b, enabling higher-level operations such as inversion, factorization and transformations of the coefficient matrix.

The second is lifting a problem into a new representational space. Fourier's work on the heat equation expressed a periodic function as a linear superposition of sinusoids, so that cases like sinusoids could be solved simply and the heat equation could then be solved for a broader class of periodic functions. Eigenvectors do something similar for linear transformations: directions satisfying Av_i = λ_i v_i reduce the transformation to scalar multiplication, and when A has a complete eigenbasis, A = PDP⁻¹ turns a coupled multidimensional transformation into independent scalar transformations, with A^k = PD^kP⁻¹.

The third is relational abstraction. The search for a solution to the general quintic, which eluded mathematicians after formulas existed for quadratic, cubic and quartic polynomials, contributed to permutation groups: elements describe permutations of a polynomial's roots, allowing study of which permutations preserve algebraic relations among roots and how those structures relate to expressing solutions in radicals. The paper traces this relational emphasis onward to Lie groups, symmetry groups and category theory.

Human-like constraints as a design choice for AI

The paper turns to machine mathematical discovery. Large language models have been trained on centuries of mathematics shaped by human cognitive bottlenecks, so they have inherited its representations, symbolism and logic and can compose novel solutions. Whether they can produce novel abstractions beyond novel compositions of existing ones is described as still open. The paper cites a test in which a GPT-2 model pretrained on language could not create an abstraction of zero but could quickly learn it from a few examples.

The concern is cost. The paper says LLMs are known to generate redundant programs, show a propensity to overthink and allocate excessive computation to simple questions, and can indiscriminately expand the search space. Because Earth's energy resources remain constrained, the author argues that inefficient and indiscriminate search may not scale. The proposed remedy is to evoke signals analogous to boredom and frustration in LLMs, prompting the meta-decision to expand the representational space instead of continuing tedious computation.

The paper qualifies this proposal directly: the ability to discover novel abstractions might require architectural advancements beyond existing methods and perhaps cannot emerge through scale alone. It also states that it is conspicuously silent about the process by which novel abstractions are discovered, which it leaves as an active topic for future research.

The historical cases illustrate rather than prove causation

The paper's own limitations section is specific about the weakest link. Cognitive science has proposed that boredom is an information-seeking state, and several abstractions such as logarithms can be linked to pressure to avoid tedious computation by restructuring representations. The paper states that while historical narratives in specific instances seem to support this hypothesis, a causal link between the emotion of boredom and the pressure to restructure is yet to be established. It also notes opportunity costs in choosing between multiple goals as another possible driver of the pressure to compress.

The formal model likewise does not specify how a new abstraction is found. It compares the expected cost of continuing with the expected cost of discovering and then using a new abstraction, but the discovery process itself is left open. The historical episodes are therefore best read as demonstrations that the proposed pattern is consistent with notable transitions in mathematics, not as tests that cognitive constraints caused those transitions.

Why accessible, reusable abstractions become the priority

The paper's closing argument follows from its cost comparison rather than from the historical cases alone. If abstractions repay their discovery cost through reuse across a distribution of problems, then the abstractions worth pursuing are those that are compressed, efficient and composable. The author pushes this further: the idea of a single unique solution should be dispensed with, and effort should go toward compressed solutions accessible to humans, ideally ones an inspired high school student could understand.

This is where the paper's account of cognitive constraints becomes a claim about what mathematics should value. The same reasoning explains its warning about machines: systems unburdened by human bottlenecks can expand search indiscriminately, but the paper argues that evoking bottleneck-like signals could push them toward restructuring instead. The most consequential judgment for readers is therefore conditional: cognitive limits drive abstraction when the cost of continuing under an existing representation exceeds the expected cost of discovering and reusing a new one, and the paper treats that condition as a design target for both human mathematics and machine discovery rather than as an established causal law.

Sources used in this answer