Restricted Complexity

• Restricted Complexity is defined as the analysis of complexity under explicit formal constraints, distinguishing tractable local measures from broader systemic challenges.
• It encompasses applications such as algorithmic complexity, compressed sensing, and learning models, where formal restrictions enable both efficient algorithms and precise evaluations.
• The study reveals a balance between tractable frameworks and NP-hard challenges, illustrating how domain-specific restrictions shape computational and combinatorial problem solving.

Searching arXiv for the cited paper and closely related work on “restricted complexity.” Restricted complexity denotes complexity analyzed under explicit formal, structural, or representational constraints. In the conceptual reconstruction associated with “Restricted Complexity, General Complexity,” it names mathematically formal notions of complexity developed within sharply delimited frameworks, while “general complexity” concerns the broader ontological, epistemological, and methodological problematic posed by real systems with many interacting components, nonlinear dynamics, multi-scale organization, emergence, and context dependence [0610049]. Across later work, the expression recurs in more specialized senses: restricted isometry in compressed sensing, restricted architectures such as restricted Boltzmann machines, maps with restricted singularities, restricted graph cuts, restricted graph classes, restricted edges in matching, restricted place environments in Petri-net synthesis, and arithmetical theories axiomatizable at a fixed level $\Sigma_n$ (Natarajan et al., 2014, Zhang, 2015, Grant et al., 2011, Montejano et al., 2015, Cseh et al., 2014, Devillers et al., 2021, Enayat et al., 20 Aug 2025).

1. Restricted complexity as complexity under explicit constraints

In the reconstructed conceptual landscape of [0610049], restricted complexity is complexity defined within a well-delimited formal framework, subject to precise rules, limited domains, and clear evaluative criteria. It is restricted in three senses: domain restriction, because it applies to particular objects such as words, programs, algorithms, formal languages, or state spaces; conceptual restriction, because complexity is operationalized via a specific quantifiable measure; and methodological restriction, because analysis proceeds under assumptions that simplify or idealize away many features of real phenomena. This yields paradigmatic forms such as algorithmic complexity, computational complexity, and combinatorial or state-space complexity.

Algorithmic complexity is represented by Kolmogorov complexity: $K(x) = \min_{p: U(p) = x} |p|$ for a string $x$ relative to a universal Turing machine $U$. Here complexity is the length of the shortest description, and the restriction lies in the object type, the representational system, and the purely syntactic notion of description. Computational complexity similarly studies formally specified problems and algorithms through time, space, and standard machine models, with complexity classes such as $\mathbf{P}$, $\mathbf{NP}$, $\mathbf{PSPACE}$, and $\mathbf{EXP}$ [0610049]. Combinatorial complexity appears when the size of a formally defined state space grows like

$|\Omega| = k^N$

or more generally $|\Omega| \propto f(N,d)$, again within a predefined configuration space.

By contrast, the same reconstruction treats general complexity as a global, systemic problematic. It addresses biological, social, ecological, cognitive, and economic systems in which many interacting components, nonlinear and history-dependent dynamics, multi-scale organization, and context-sensitive emergence prevent complexity from being reduced to description length, runtime, or cardinality alone. A schematic contrast is: $K(x) = \min_{p: U(p) = x} |p|$0 This suggests that restricted complexity is best understood not as a single invariant, but as complexity measured after a domain, a language, and an admissible class of operations have already been fixed.

2. Historical emergence and justification of the problematic

The abstract of [0610049] asks why the problematic of complexity appeared so late and why it is justified. The reconstructed answer is historical as well as methodological. Classical science was organized around reductionism, linearity, superposition, and closed-system idealizations, so complex phenomena were often treated as complicated but in principle reducible. Under that picture, complexity appeared as a by-product of limited computational power or incomplete information rather than as a distinct object of inquiry.

The late emergence of complexity as a problematic is tied, in [0610049], to several twentieth-century developments: nonlinear dynamics and chaos theory, where determinism survives but predictability and closed-form solvability break down; statistical mechanics and systems theory, where wholes, ensembles, feedback, and openness become central; cybernetics and information theory, which introduced rigorous notions of feedback, communication, entropy, and noise; computer science and algorithmic information theory, which supplied restricted complexity measures; and the increasing confrontation, in biology, neuroscience, sociology, and economics, with self-organization, emergence, and many-agent interaction. The same reconstruction adds that the problematic also required computational tools such as simulations, agent-based models, and network analysis, together with large-scale data and enough conceptual maturity to recognize that the success of reductionism in physics did not automatically generalize to biological, cognitive, and social domains.

Its justification is correspondingly twofold. First, many phenomena are not well captured by additive decompositions of the form

$K(x) = \min_{p: U(p) = x} |p|$1

because their behavior depends on interaction structure, history, and environment: $K(x) = \min_{p: U(p) = x} |p|$2 Second, the broad complexity problematic can still be partially formalized through network models, nonlinear dynamical systems, and information-theoretic quantities, even if no single scalar captures all relevant aspects [0610049]. The conceptual role of restricted complexity is therefore double: it provides rigorous local measures, but also marks the limits of those measures when they are transferred to real open systems.

3. Sparse recovery and the restricted complexity of certification

One of the most precise technical uses of restricted complexity in the supplied material arises in compressed sensing. The restricted isometry property (RIP) is “restricted” because norm preservation is required only on sparse vectors. A matrix $K(x) = \min_{p: U(p) = x} |p|$3 is $K(x) = \min_{p: U(p) = x} |p|$4-RIP if, for all $K(x) = \min_{p: U(p) = x} |p|$5 with $K(x) = \min_{p: U(p) = x} |p|$6,

$K(x) = \min_{p: U(p) = x} |p|$7

or, in the squared form used in (Tillmann et al., 2012),

$K(x) = \min_{p: U(p) = x} |p|$8

This restriction is exactly the structural condition that underwrites sparse recovery by tractable algorithms, including $K(x) = \min_{p: U(p) = x} |p|$9-minimization, while the underlying sparsest-solution problem remains NP-hard (Tillmann et al., 2012).

The computational complexity of checking these restricted properties is, however, intractable in general. For a given matrix $x$0 and positive integer $x$1, computing the best constants for which RIP or the nullspace property hold is NP-hard; deciding whether there exists some $x$2 such that $x$3 satisfies RIP of order $x$4 is coNP-complete; computing the symmetric restricted isometry constant $x$5 is NP-hard; computing the upper asymmetric restricted isometry constant is strongly NP-hard, even for square matrices; and computing the nullspace constant $x$6 is NP-hard (Tillmann et al., 2012). The same paper emphasizes the tension: restricted properties such as RIP, NSP, and spark make recovery tractable, yet the meta-question of whether a given matrix has those restricted properties is itself computationally hard.

A stronger hardness-of-approximation result is obtained under the Small-Set Expansion hypothesis. For any $x$7 and any constant $x$8, there exists $x$9 such that, given a matrix $U$0, it is SSE-hard to distinguish between a “Highly RIP” case, where $U$1 is $U$2-RIP, and a “Far away from RIP” case, where $U$3 is not $U$4-RIP (Natarajan et al., 2014). The technical bridge is a sparse Cheeger inequality: $U$5 which ties expansion of small sets to Laplacian behavior on sparse vectors. The resulting message is not merely exact-decision hardness but hardness of coarse approximation in both sparsity and distortion parameters.

The average-case picture is comparably sharp. For Gaussian $U$6 matrices with entries $U$7, RIP holds with high probability for

$U$8

but certifying RIP in the “possible but hard” regime $U$9 appears to require subexponential time. Using the low-degree likelihood ratio method, (Ding et al., 2020) gives evidence that runtime

$\mathbf{P}$0

is required, and this essentially matches the runtime of the “lazy algorithm” of Koiran and Zouzias. This yields a fine-grained restricted complexity law: the complexity of certification scales smoothly with the restricted dimension $\mathbf{P}$1, not merely with ambient size.

4. Capacity, representation, and restricted complexity in learning and analysis

In machine learning, restricted complexity appears both as an architectural constraint and as a capacity measure. For restricted Boltzmann machines with visible units $\mathbf{P}$2, hidden units $\mathbf{P}$3, and energy

$\mathbf{P}$4

the bipartite restriction implies a factorization of the log-likelihood term over hidden units (Zhang, 2015). Under norm constraints $\mathbf{P}$5 and $\mathbf{P}$6, the empirical Rademacher complexity of the exact log-likelihood satisfies

$\mathbf{P}$7

The same paper then analyzes the practical CD-1 training procedure with mean-field approximations and shows that the effective function class becomes richer: the approximate partition term becomes a compositional function of $\mathbf{P}$8, and the resulting Rademacher complexity bound acquires an additional term involving the VC-dimension of a compositional class $\mathbf{P}$9. The paper’s stated conclusion is that the practical implementation training procedure indeed increased the Rademacher complexity of restricted Boltzmann machines (Zhang, 2015).

A different, but related, use appears in higher-type complexity for analysis. The framework of (Kawamura et al., 2013) replaces infinite-sequence names by regular string functions $\mathbf{NP}$0, equipped with a size function

$\mathbf{NP}$1

and measures time and space by second-order polynomials $\mathbf{NP}$2. This extends earlier settings that were described in the paper as more restricted. It yields analogues of $\mathbf{NP}$3, $\mathbf{NP}$4, and $\mathbf{NP}$5 for operators on represented spaces and supports completeness results at operator level: for example, the convex hull operator on closed subsets of $\mathbf{NP}$6 is $\mathbf{NP}$7-$\mathbf{NP}$8-complete, and the operator $\mathbf{NP}$9 solving a Lipschitz initial-value problem is $\mathbf{PSPACE}$0-$\mathbf{PSPACE}$1-complete (Kawamura et al., 2013). A plausible implication is that restricted complexity can also designate complexity measured relative to a representation regime: the restriction is not on the object alone, but on the admissible names, sizes, and oracle access patterns through which the object is presented.

5. Geometric and combinatorial manifestations

In differential topology, restricted complexity means restricting the admissible singular behavior of maps that realize homology classes. The paper (Grant et al., 2011) studies two regimes: immersions, where there are no singularities beyond self-intersections, and $\mathbf{PSPACE}$2-maps, where each fiber is allowed only a finite set $\mathbf{PSPACE}$3 of multi-singularity types. Its main results are negative: for every codimension $\mathbf{PSPACE}$4, there exists a closed manifold $\mathbf{PSPACE}$5 and a class $\mathbf{PSPACE}$6 that cannot be realized by an immersion, and more generally cannot be realized by a $\mathbf{PSPACE}$7-map for any fixed finite $\mathbf{PSPACE}$8. For immersions, a concrete obstruction is

$\mathbf{PSPACE}$9

for admissible Steenrod monomials $\mathbf{EXP}$0 of excess $\mathbf{EXP}$1; in particular, when $\mathbf{EXP}$2 is even, $\mathbf{EXP}$3 obstructs realization by immersion. Here restricted complexity is literally a restriction on the complexity of singularity types.

Several graph-theoretic uses have the same pattern. The $\mathbf{EXP}$4-restricted edge-connectivity $\mathbf{EXP}$5 is the minimum size of an edge set whose removal leaves two connected components each containing at least $\mathbf{EXP}$6 vertices. It is equivalent to the notion of a $\mathbf{EXP}$7-good edge separation, and its complexity sharply depends on parameterization: the existential problem EREC is NP-complete, $\mathbf{EXP}$8-REC is W[1]-hard when parameterized by $\mathbf{EXP}$9, fixed-parameter tractable for $|\Omega| = k^N$0, fixed-parameter tractable for $|\Omega| = k^N$1, and fixed-parameter tractable for $|\Omega| = k^N$2, while kernelization lower bounds rule out polynomial kernels for parameter $|\Omega| = k^N$3 unless $|\Omega| = k^N$4 (Montejano et al., 2015). Restricted edge-connectivity thus turns a polynomial-time minimum-cut notion into a parameter-sensitive landscape of hardness and FPT regimes.

For rainbow vertex connectivity, restricting the graph class often fails to restore tractability. Deciding whether a given vertex-colored graph is rainbow vertex connected or strongly rainbow vertex connected remains NP-complete on very restricted graph classes including bipartite planar graphs of maximum degree $|\Omega| = k^N$5, interval graphs, and $|\Omega| = k^N$6-regular graphs for $|\Omega| = k^N$7; both problems remain NP-complete for bounded pathwidth graphs, yet are fixed-parameter tractable when parameterized by tree-depth, and become polynomial-time solvable on block graphs, while strong rainbow vertex connectivity is tractable on cactus graphs and split graphs (Lauri, 2016). The same contrast appears in fall colouring: deciding whether a bipartite planar graph can be partitioned into three independent dominating sets is NP-complete, deciding whether a bipartite graph can be partitioned into $|\Omega| = k^N$8 independent dominating sets is NP-complete for every $|\Omega| = k^N$9, deciding whether a graph has two disjoint independent dominating sets is NP-complete even for triangle-free planar graphs, yet every maximal outerplanar graph with at least three vertices satisfies $|\Omega| \propto f(N,d)$0 (Lauri et al., 2019).

Restricted star colouring imposes its own local prohibition: a $|\Omega| \propto f(N,d)$1-rs colouring is a proper $|\Omega| \propto f(N,d)$2-colouring with no bicoloured $|\Omega| \propto f(N,d)$3 whose middle vertex has the higher colour. For $|\Omega| \propto f(N,d)$4, deciding $|\Omega| \propto f(N,d)$5-rs colourability is NP-complete on planar bipartite graphs of maximum degree $|\Omega| \propto f(N,d)$6 and arbitrarily large girth; deciding whether a 3-star colourable graph admits a 3-rs colouring is NP-complete; and the optimization problem is NP-hard to approximate within $|\Omega| \propto f(N,d)$7 on 2-degenerate bipartite graphs for every $|\Omega| \propto f(N,d)$8. On the positive side, there is a linear-time algorithm to test 3-rs colourability of trees and an $|\Omega| \propto f(N,d)$9-time algorithm for chordal graphs (A. et al., 2021). These examples show that restricted complexity often names a local structural prohibition whose global algorithmic effect is highly nonlocal.

In matching theory, restricted edges produce a similarly bifurcated picture. In stable marriage and stable roommates, one distinguishes forced edges $K(x) = \min_{p: U(p) = x} |p|$00 and forbidden edges $K(x) = \min_{p: U(p) = x} |p|$01. A stable matching satisfying all restricted-edge constraints must satisfy

$K(x) = \min_{p: U(p) = x} |p|$02

If such a stable matching does not exist, the paper (Cseh et al., 2014) studies two relaxations. In case (1), one keeps all edge restrictions and minimizes the number of blocking pairs; in the stable marriage setting this leads to NP-hardness and strong inapproximability results, while in case (2), one preserves stability and minimizes the number of violated constraints

$K(x) = \min_{p: U(p) = x} |p|$03

which is polynomial-time solvable in stable marriage through minimum-weight stable matching. In stable roommates, the corresponding stable version is NP-hard. Restricted complexity here is the complexity generated by adding edge-level admissibility constraints to a classical stability problem.

6. Synthesis problems and logical restricted complexity

In Petri-net synthesis, restricted complexity is expressed by structural restrictions on places. For weighted P/T nets, (Devillers et al., 2021) studies synthesis into nets where every place $K(x) = \min_{p: U(p) = x} |p|$04 satisfies

$K(x) = \min_{p: U(p) = x} |p|$05

For any fixed natural numbers $K(x) = \min_{p: U(p) = x} |p|$06 and $K(x) = \min_{p: U(p) = x} |p|$07, deciding whether a transition system has a realizing Petri net with these restricted place environments is polynomial-time solvable, because one can reduce separation conditions to linear systems and only polynomially many candidate environment choices arise when $K(x) = \min_{p: U(p) = x} |p|$08 are fixed. Once $K(x) = \min_{p: U(p) = x} |p|$09 and $K(x) = \min_{p: U(p) = x} |p|$10 are taken as input, however, the resulting Environment Restricted Synthesis problem is NP-complete for impure nets and also NP-complete for pure nets; in the impure case it is $K(x) = \min_{p: U(p) = x} |p|$11-hard when parameterized by $K(x) = \min_{p: U(p) = x} |p|$12 (Devillers et al., 2021). An analogous phenomenon appears for Boolean Petri nets: if $K(x) = \min_{p: U(p) = x} |p|$13-synthesis is NP-complete, then dependency $K(x) = \min_{p: U(p) = x} |p|$14-restricted $K(x) = \min_{p: U(p) = x} |p|$15-synthesis is also NP-complete; parameterized by $K(x) = \min_{p: U(p) = x} |p|$16, the problem lies in XP with running time $K(x) = \min_{p: U(p) = x} |p|$17, yet it is $K(x) = \min_{p: U(p) = x} |p|$18-hard for many Boolean types that allow unconditional set and reset (Tredup et al., 2020). In both papers, the restriction is local to places, but the induced synthesis problem becomes globally hard once the structural bound is itself variable.

The most literal use of the phrase “restricted complexity” in the supplied material appears in weak arithmetic. The paper (Enayat et al., 20 Aug 2025) defines a first-order arithmetical theory $K(x) = \min_{p: U(p) = x} |p|$19 to be of restricted complexity if there exist $K(x) = \min_{p: U(p) = x} |p|$20 and a set $K(x) = \min_{p: U(p) = x} |p|$21 of $K(x) = \min_{p: U(p) = x} |p|$22-sentences such that $K(x) = \min_{p: U(p) = x} |p|$23 can be axiomatized by $K(x) = \min_{p: U(p) = x} |p|$24. This measures the complexity of a theory by the syntactic complexity of its axioms rather than by the computability of the axiom set. Motivated by the result that no consistent arithmetical theory extending $K(x) = \min_{p: U(p) = x} |p|$25 has a consistent completion of restricted complexity, the paper constructs models of weak arithmetic whose complete theories do have restricted complexity. The strongest theorem states that $K(x) = \min_{p: U(p) = x} |p|$26 has a completion axiomatized by a sentence $K(x) = \min_{p: U(p) = x} |p|$27 together with a set of $K(x) = \min_{p: U(p) = x} |p|$28-sentences, hence of restricted complexity (Enayat et al., 20 Aug 2025). This is a direct logical counterpart to the broader pattern seen elsewhere: strong ambient theories rule out bounded-complexity completions, while weaker settings permit them through carefully constructed models and interpretations.

Taken together, these developments support a unified, though heterogeneous, interpretation. Restricted complexity is not exhausted by one formalism. It may refer to low-level syntactic axiomatizability, to behavior on sparse vectors, to bounded place environments, to finite singularity palettes, to graph classes with constrained width or degree, or to architectural constraints in learning models. What is common across these uses is that complexity is evaluated only after the admissible ambient structure has been sharply delimited. In some settings that delimitation enables polynomial-time algorithms; in others it preserves NP-completeness, yields $K(x) = \min_{p: U(p) = x} |p|$29- or $K(x) = \min_{p: U(p) = x} |p|$30-hardness, or exposes new incompleteness thresholds.