Complexity Inequality: Theory & Implications

• Complexity inequality is a multifaceted topic that compares complexity measures with structural invariants across dynamical systems, Gaussian analysis, and statistical learning.
• In algorithmic and information theory, it involves bounds on Kolmogorov complexity and Shannon entropy, extending to quantum complexity and coding applications.
• In socioeconomic studies, empirical evidence links economic complexity with income inequality, revealing contrasting relationships at national and regional levels.

Searching arXiv for recent and foundational papers on “complexity inequality” across the domains represented in the source material. I’m going to gather the directly relevant arXiv records, prioritizing papers explicitly centered on a “complexity inequality” or on inequalities where “complexity” is the key controlled quantity. Complexity inequality is not a single canonical theorem. In the literature represented here, the phrase appears in several domain-specific senses: as an inequality relating two formal complexity invariants; as a finite-sample or functional bound in empirical process theory and analysis; as a complexity-theoretic statement about deciding whether an inequality or an equality case holds; and, in economic applications, as the relation between economic complexity and income inequality. The common feature is that a complexity quantity is compared with a second structural quantity—matching, entropy, deviation, nonlocality, welfare, or distributive disparity—rather than studied in isolation (Schneider, 2015, Eldan et al., 2019, Hartmann et al., 2015).

1. Meanings and scope

A first usage is structural and mathematical. In dynamical systems, the asymptotic uniform complexity $\omega(X,G)$ is compared with the mean topological matching number $\mu(X,G)$ through

$\omega(X,G)\ge 2-\mu(X,G),$

with equality

$\omega(X,G)=2-\mu(X,G)$

for perfect Hausdorff uniform spaces (Schneider, 2015). In Gaussian analysis, Fisher information is bounded above by entropy and a Gaussian-width complexity term through

$I(\nu)\le 2D_{\mathrm{KL}}(\nu\|\gamma)+M+D(\nu),$

a reverse logarithmic Sobolev inequality for low-complexity functions (Eldan et al., 2019). In learning theory, relative empirical-process deviations are controlled by VC or VC-subgraph growth functions (Portier, 13 Jul 2026), and Lipschitz composition is controlled by vector-valued Rademacher contraction inequalities (Maurer, 2016).

A second usage is algorithmic and information-theoretic. Here the relevant object is often Kolmogorov complexity, ordinary or resource-bounded. One paper extends the Romashchenko–Zimand conditional inequality from partitions into combinatorial rectangles to overlapping rectangle covers, introducing a penalty $\log \rho$ for covering thickness (Romashchenko et al., 2019). Another proves that every linear inequality valid for ordinary Kolmogorov complexity and Shannon entropy also holds for space-bounded Kolmogorov complexity with a polynomial space overhead (Bauwens et al., 2020). A further paper formulates an extended coding theorem of the form

$$\min_{a\in \mathrm{Dom}(f)}\bigl(K(a)+f(a)\bigr)\lesssim -\log \sum_a m(a)2^{-f(a)} + (\langle f\rangle;\mathcal H),$$

with applications to quantum complexity measures (Epstein, 2015).

A third usage is socioeconomic. There, “inequality” denotes income dispersion rather than an inequality sign. The claim is that countries with more complex productive structures tend to have lower income inequality (Hartmann et al., 2015), while the association reverses at regional scales, where more complex regions tend to be more unequal (Hartmann et al., 2022). This suggests that the phrase has become polysemous: it can denote either inequalities about complexity measures or inequality outcomes associated with complex systems.

2. Information-theoretic and algorithmic formulations

In Kolmogorov-complexity theory, a central problem is whether conditioning on structured side information can increase dependence. For partitions of $\{0,1\}^n\times\{0,1\}^n$ into combinatorial rectangles, Romashchenko and Zimand established

$I(x:y)\ge I(x:y\mid t(x,y)) - O(\log n).$

The covering extension replaces uniqueness of the transcript rectangle by bounded overlap, giving

$$I(x:y\mid \Pi)\ge I(x:y\mid t(x,y),\Pi)-\log \rho(x,y)-O(\log n),$$

equivalently

$\mu(X,G)$0

Here $\mu(X,G)$1 is the covering thickness, and the new term quantifies the ambiguity introduced by overlap (Romashchenko et al., 2019). This inequality is used for nondeterministic, randomized, Arthur–Merlin, and information-complexity lower bounds.

Space-bounded Kolmogorov complexity introduces a different type of complexity inequality. The key technical result is an improved space-bounded Kolmogorov–Levin formula, with a tighter space bound than Longpré’s earlier version: $\mu(X,G)$2 From this, the paper derives space-bounded forms of the basic submodularity-type inequalities and then proves a general transfer principle: every linear inequality that is true for ordinary Kolmogorov complexity, and hence for Shannon entropy, is also true for space-bounded Kolmogorov complexity with polynomial space overhead (Bauwens et al., 2020). In this setting, “complexity inequality” refers not to a single formula but to the persistence of the entire linear-information-inequality calculus under space restriction.

The extended coding theorem sharpens the variational side of algorithmic information theory. For an elementary map $\mu(X,G)$3, the inequality

$\mu(X,G)$4

generalizes both the coding theorem $\mu(X,G)$5 and earlier finite-set inequalities (Epstein, 2015). The extra term $\mu(X,G)$6 measures the information that $\mu(X,G)$7 has about the halting sequence, so the obstruction to a clean coding statement is explicitly quantified by halting information. The same paper uses this inequality to compare quantum complexity measures, obtaining

$\mu(X,G)$8

3. Statistical, geometric, and learning-theoretic inequalities

In empirical-process theory, a complexity inequality often takes the form of a uniform tail bound controlled by a combinatorial growth function. For a countable class of measurable sets $\mu(X,G)$9,

$\omega(X,G)\ge 2-\mu(X,G),$0

and

$\omega(X,G)\ge 2-\mu(X,G),$1

For bounded function classes $\omega(X,G)\ge 2-\mu(X,G),$2, the analogous bounds are expressed through the shatter coefficient of the subgraph class, $\omega(X,G)\ge 2-\mu(X,G),$3, with an additive slack $\omega(X,G)\ge 2-\mu(X,G),$4 (Portier, 13 Jul 2026). These results are explicitly non-asymptotic, use relative normalization by the probability level, and improve the constants in the classical Anthony–Shawe-Taylor relative VC inequality.

The same paper emphasizes that the proof is elementary: a new convexity-based symmetrization principle is combined with a maximal sub-Gaussian inequality, and neither concentration inequalities nor entropy-integral arguments are required (Portier, 13 Jul 2026). This places “complexity” in the sense of VC and VC-subgraph combinatorics, where the shatter coefficient is the multiplicative factor governing deviation.

A different learning-theoretic notion appears in the vector-contraction inequality for Rademacher averages. If $\omega(X,G)\ge 2-\mu(X,G),$5 are $\omega(X,G)\ge 2-\mu(X,G),$6-Lipschitz and $\omega(X,G)\ge 2-\mu(X,G),$7 is a class of functions $\omega(X,G)\ge 2-\mu(X,G),$8, then

$\omega(X,G)\ge 2-\mu(X,G),$9

The paper also proves a more general version with arbitrary iid symmetric sub-Gaussian variables on the right-hand side (Maurer, 2016). The result extends the classical scalar contraction principle to genuinely vector-valued outputs and is applied to multiclass learning, $\omega(X,G)=2-\mu(X,G)$0-means clustering, and learning-to-learn.

In Gaussian analysis, low-complexity functions are measured by the Gaussian width of the gradient image,

$\omega(X,G)=2-\mu(X,G)$1

For densities $\omega(X,G)=2-\mu(X,G)$2, the reverse logarithmic Sobolev bound

$\omega(X,G)=2-\mu(X,G)$3

is dimension-free in the sense emphasized by the paper: the bound has an additive, tensorization-friendly structure (Eldan et al., 2019). This reverses the usual Gaussian log-Sobolev direction under a low-complexity hypothesis. The same work also proves the refined form

$\omega(X,G)=2-\mu(X,G)$4

and connects the resulting inequalities to Gaussian-mixture structure.

4. Dynamical systems, decoding, and robust inequality testing

For general dynamical systems $\omega(X,G)=2-\mu(X,G)$5, the asymptotic uniform complexity

$\omega(X,G)=2-\mu(X,G)$6

is compared to the mean topological matching number $\omega(X,G)=2-\mu(X,G)$7 (Schneider, 2015). The main inequality

$\omega(X,G)=2-\mu(X,G)$8

is upgraded to the exact formula

$\omega(X,G)=2-\mu(X,G)$9

for perfect Hausdorff uniform spaces (Schneider, 2015). The paper uses this to derive amenability criteria. In particular, $I(\nu)\le 2D_{\mathrm{KL}}(\nu\|\gamma)+M+D(\nu),$0 implies amenability in general, and for perfect Hausdorff systems the converse also holds. This is one of the clearest cases where “complexity inequality” denotes a precise relation between two dynamical invariants.

The same work also introduces a topologically free analogue involving

$I(\nu)\le 2D_{\mathrm{KL}}(\nu\|\gamma)+M+D(\nu),$1

and proves that vanishing topological entropy implies $I(\nu)\le 2D_{\mathrm{KL}}(\nu\|\gamma)+M+D(\nu),$2 (Schneider, 2015). A plausible implication is that the paper treats complexity growth under translates as a dynamical proxy for approximate invariance.

Coding theory furnishes a different non-asymptotic example. Sequential decoding complexity can be estimated by rewriting a path-expansion event as a probability inequality for a sum of random variables and then using the Berry–Esseen theorem rather than a purely asymptotic central limit theorem argument [0701026]. The new upper bound is valid for any blocklength. For the simplified generalized Dijkstra algorithm, the theoretical upper bound almost matches simulation as the signal-to-noise ratio per information bit $I(\nu)\le 2D_{\mathrm{KL}}(\nu\|\gamma)+M+D(\nu),$3 is greater than or equal to $I(\nu)\le 2D_{\mathrm{KL}}(\nu\|\gamma)+M+D(\nu),$4 dB, while for the maximum-likelihood sequential decoding algorithm the bound remains close to simulation results at both high and low SNR and differs by at most $I(\nu)\le 2D_{\mathrm{KL}}(\nu\|\gamma)+M+D(\nu),$5 on a $I(\nu)\le 2D_{\mathrm{KL}}(\nu\|\gamma)+M+D(\nu),$6 scale even for moderate SNR [0701026]. Here the “inequality” is a finite-blocklength probability bound that induces an expected-complexity estimate.

Robust eventual inequality testing for C-finite functions uses the term in yet another sense. Given two solutions of homogeneous linear differential equations with constant coefficients, the question is whether

$I(\nu)\le 2D_{\mathrm{KL}}(\nu\|\gamma)+M+D(\nu),$7

Because the inputs are arbitrary real parameters in the bit-model, the paper develops maximal partial decidability and then proves that eventual inequality is polynomial-time decidable in that robust sense (Neumann, 2023). The running time is polynomial in the input size and in the logarithm of the inverse distance to the boundary of the decision set. In this context, the relevant complexity inequality is a statement about the computational complexity of deciding a robust order relation on continuous-time linear dynamical objects.

5. Equality cases, nonlocality, and computational hardness

Some papers study the complexity of recognizing when an inequality is tight. For the Stanley–Yan log-concave matroid inequality

$I(\nu)\le 2D_{\mathrm{KL}}(\nu\|\gamma)+M+D(\nu),$8

the equality problem behaves very differently depending on whether $I(\nu)\le 2D_{\mathrm{KL}}(\nu\|\gamma)+M+D(\nu),$9 or $\log \rho$0 (Chan et al., 2024). For $\log \rho$1, the paper gives a complete structural characterization of equality in terms of parallel classes in contractions, proves nonvanishing conditions, and shows that the recognition problem $\log \rho$2 lies in $\log \rho$3 (Chan et al., 2024). For $\log \rho$4, the situation becomes complexity-theoretically intractable: $\log \rho$5 already for binary matroids and even for $\log \rho$6 and $\log \rho$7 (Chan et al., 2024). In the paper’s formulation, the equality cases “cannot be described” by any finite-level polynomial-hierarchy characterization unless the polynomial hierarchy collapses.

Quantum communication complexity supplies a different complexity-to-inequality principle. If a function $\log \rho$8 admits a sufficiently strong quantum advantage, then one can construct measurement statistics that violate a Bell inequality (Buhrman et al., 2015). Quantitatively, if $\log \rho$9 and $$\min_{a\in \mathrm{Dom}(f)}\bigl(K(a)+f(a)\bigr)\lesssim -\log \sum_a m(a)2^{-f(a)} + (\langle f\rangle;\mathcal H),$$0, the Bell-value ratio satisfies

$$\min_{a\in \mathrm{Dom}(f)}\bigl(K(a)+f(a)\bigr)\lesssim -\log \sum_a m(a)2^{-f(a)} + (\langle f\rangle;\mathcal H),$$1

Hence a separation

$$\min_{a\in \mathrm{Dom}(f)}\bigl(K(a)+f(a)\bigr)\lesssim -\log \sum_a m(a)2^{-f(a)} + (\langle f\rangle;\mathcal H),$$2

yields an unbounded Bell violation (Buhrman et al., 2015). In this setting, the relevant inequality is the Bell inequality itself, but its existence is derived from a communication-complexity inequality.

Fair-division theory uses “inequality” in the distributive sense but still couples it to computational complexity. For additive, nonnegative utilities over indivisible goods, the paper studies minimization of the Gini index, the subjective Gini index, and the envy index (Aleksandrov et al., 2018). All three optimization problems are NP-hard in general, but when there are $$\min_{a\in \mathrm{Dom}(f)}\bigl(K(a)+f(a)\bigr)\lesssim -\log \sum_a m(a)2^{-f(a)} + (\langle f\rangle;\mathcal H),$$3 agents and $$\min_{a\in \mathrm{Dom}(f)}\bigl(K(a)+f(a)\bigr)\lesssim -\log \sum_a m(a)2^{-f(a)} + (\langle f\rangle;\mathcal H),$$4 items each can be solved in $$\min_{a\in \mathrm{Dom}(f)}\bigl(K(a)+f(a)\bigr)\lesssim -\log \sum_a m(a)2^{-f(a)} + (\langle f\rangle;\mathcal H),$$5 time (Aleksandrov et al., 2018). The same work proves that mechanisms minimizing these indices are not strategy-proof, that minimizing Gini or subjective Gini can conflict with envy-freeness, and that the utilitarian and egalitarian prices of the Gini and subjective Gini indices are unbounded (Aleksandrov et al., 2018). This suggests that, in algorithmic allocation, minimizing inequality is computationally hard and normatively non-equivalent to classical fairness conditions.

6. Quantum subsystems and socioeconomic uses of inequality

Subsystem complexity introduces a recent quantum meaning of the term. For reduced density matrices of three disjoint regions, the paper defines the tripartite complexity

$$\min_{a\in \mathrm{Dom}(f)}\bigl(K(a)+f(a)\bigr)\lesssim -\log \sum_a m(a)2^{-f(a)} + (\langle f\rangle;\mathcal H),$$6

and the complexity gap

$$\min_{a\in \mathrm{Dom}(f)}\bigl(K(a)+f(a)\bigr)\lesssim -\log \sum_a m(a)2^{-f(a)} + (\langle f\rangle;\mathcal H),$$7

(Caputa et al., 18 Jun 2026). Across three frameworks—holographic complexity=volume, Fisher–Rao subsystem complexity for Gaussian states, and a Krylov-space-inspired effective subsystem complexity—the paper finds that $$\min_{a\in \mathrm{Dom}(f)}\bigl(K(a)+f(a)\bigr)\lesssim -\log \sum_a m(a)2^{-f(a)} + (\langle f\rangle;\mathcal H),$$8 is not sign-definite in general (Caputa et al., 18 Jun 2026). By contrast, the complexity gap has a definite sign in every example analyzed, although the sign depends on the underlying definition: in static AdS$$\min_{a\in \mathrm{Dom}(f)}\bigl(K(a)+f(a)\bigr)\lesssim -\log \sum_a m(a)2^{-f(a)} + (\langle f\rangle;\mathcal H),$$9 complexity=volume examples,

$\{0,1\}^n\times\{0,1\}^n$0

whereas in the Fisher–Rao and Krylov analyses the observed sign is

$\{0,1\}^n\times\{0,1\}^n$1

(Caputa et al., 18 Jun 2026). The paper therefore proposes the complexity gap, rather than the direct tripartite-information analogue, as a more plausible building block for future subsystem-complexity inequalities.

Economic-complexity research uses “inequality” in the sense of income distribution. One paper shows that countries exporting complex products, as measured by the Economic Complexity Index, have lower levels of income inequality than countries exporting simpler products (Hartmann et al., 2015). In pooled OLS specifications, the coefficient on ECI is negative and statistically significant, with representative estimates around $\{0,1\}^n\times\{0,1\}^n$2, and in fixed-effects panels a one-standard-deviation increase in ECI is associated with a reduction in Gini of $\{0,1\}^n\times\{0,1\}^n$3 (Hartmann et al., 2015). The paper also introduces the Product Gini Index,

$\{0,1\}^n\times\{0,1\}^n$4

which assigns to each product the weighted average inequality of its exporters (Hartmann et al., 2015).

A later review argues that the relation reverses at subnational scale: increasing ECI tends to be associated with lower income inequality at the country level, but more complex regions tend to be more unequal (Hartmann et al., 2022). The chapter explicitly describes this as a Simpson’s paradox and attributes it to institutional co-evolution at the country level, outsourcing of less desirable activities by sophisticated economies, and agglomeration and polarization within regions and large cities (Hartmann et al., 2022). In this literature, “complexity inequality” no longer denotes a formal bound on a complexity functional, but the empirical relation between a productive-structure measure and a distributive outcome.

Taken together, these uses show that “complexity inequality” is an umbrella expression whose precise meaning depends on the field. In some domains it denotes a theorem comparing invariants; in others it denotes a deviation bound or a hardness result about recognizing equality cases; and in economic applications it denotes a substantive relationship between complex systems and unequal outcomes. The phrase therefore names a family of research programs rather than a single mathematical object.