Non-Redundancy (NRD) in Formal Systems

• Non-redundancy (NRD) is a minimality notion ensuring each component of a formal object is indispensable, typically defined via deletion invariance.
• In constraint satisfaction problems, NRD mandates a witness assignment for each constraint, refining complexity bounds like Θ(n^k) for k-SAT and other structured problems.
• NRD also informs coding theory and channel simulation by optimizing redundancy in code design and bounding excess communication, offering practical design insights.

Non-redundancy (NRD) denotes a family of minimality notions concerned with whether some component of a formal object can be removed without changing what that object means, satisfies, proves, compresses, or costs. In recent arXiv literature, the term appears most prominently as a structural parameter for constraint satisfaction problems, but it also names witness-based minimality properties for graded attribute implications, low-redundancy constructions in coding theory, and the failure of “no-excess-cost” exact channel simulation for non-singular channels. Related applied literatures use the same underlying idea to study duplicated meaning in non-autoregressive translation, symmetry-compatible redundant cells in woven fabrics, and overlap patterns in redundancy scheduling (Sharma et al., 23 Apr 2026, Vychodil, 2015, Liu et al., 2023, Flamich et al., 23 Jan 2025).

1. Core formal patterns

A recurrent formal pattern defines redundancy by deletion invariance. In CSPs, a constraint $C$ is redundant if removing it does not change the satisfying set, $\mathrm{Sat}(I\setminus\{C\})=\mathrm{Sat}(I)$; an instance is non-redundant if every constraint is essential, equivalently if every constraint has a witness assignment satisfying all others but not that one (Sharma et al., 23 Apr 2026). In graded attribute implications, a theory $\Sigma$ is redundant if it has a proper equivalent subset and non-redundant otherwise; the paper on witnessed non-redundancy further characterizes non-redundancy by requiring that for each rule $A\Rightarrow B$, the antecedent $A$ is a model of the remaining theory, so the necessity of the rule is witnessed directly by its antecedent (Vychodil, 2015).

A second recurring pattern concerns necessity of assumptions. In the mathematical-methodology literature, redundancy is analyzed both globally—some statement of a theory being derivable from the rest—and locally, either as mutual redundancy among hypotheses or as redundancy of a hypothesis relative to the theorem’s thesis. That work distinguishes ordinary non-redundancy from the stronger situation in which the thesis determines all hypotheses, which it treats as characterization rather than ordinary NRD, and stresses that formal non-redundancy is only one preliminary criterion of mathematical relevance (Catsigeras, 2017).

A third pattern replaces deletion by excess cost. In exact channel simulation, “non-redundancy” would mean achieving the information-theoretic minimum without asymptotic overhead, but the non-singular case provably forbids this: the simulation cost is lower bounded not only by mutual information but by channel simulation divergence, and the asymptotic overhead has a universal $\tfrac12\log n$ term in the paper’s normalization (Flamich et al., 23 Jan 2025). Taken together, these formulations suggest that NRD is less a single invariant than a family of indispensability criteria adapted to different semantic objects.

2. Constraint satisfaction problems as the main modern setting

A prominent recent usage treats NRD as a structural parameter of $\mathsf{CSP}(\Gamma)$. For a fixed finite-domain constraint language $\Gamma$, the quantity

$\mathsf{NRD}_n(\Gamma)$

is the maximum number of constraints in a non-redundant instance on $n$ variables. In this setting, each constraint must admit a witness assignment satisfying all other constraints while violating the distinguished one. The single-pass streaming complexity of deciding satisfiability is characterized, up to a logarithmic factor, by this parameter: deterministic streaming admits an $\mathrm{Sat}(I\setminus\{C\})=\mathrm{Sat}(I)$0-space algorithm, while randomized single-pass streaming requires $\mathrm{Sat}(I\setminus\{C\})=\mathrm{Sat}(I)$1 space (Sharma et al., 23 Apr 2026).

This characterization refines earlier lower bounds by replacing arity alone with non-redundancy. The same paper recovers $\mathrm{Sat}(I\setminus\{C\})=\mathrm{Sat}(I)$2 space for $\mathrm{Sat}(I\setminus\{C\})=\mathrm{Sat}(I)$3-SAT because $\mathrm{Sat}(I\setminus\{C\})=\mathrm{Sat}(I)$4, gives quadratic behavior for graph $\mathrm{Sat}(I\setminus\{C\})=\mathrm{Sat}(I)$5-coloring/disequality when $\mathrm{Sat}(I\setminus\{C\})=\mathrm{Sat}(I)$6, and shows that $\mathrm{Sat}(I\setminus\{C\})=\mathrm{Sat}(I)$7-LIN satisfiability is consistent with linear NRD behavior. It also proves a basic universal lower bound $\mathrm{Sat}(I\setminus\{C\})=\mathrm{Sat}(I)$8 for any non-trivial bounded-arity language by placing a proper relation on disjoint $\mathrm{Sat}(I\setminus\{C\})=\mathrm{Sat}(I)$9-tuples (Sharma et al., 23 Apr 2026).

The same literature also identifies the scope of the parameter. For general finite-domain CSPs, NRD controls exact streaming complexity up to logarithmic factors. For positive Boolean CSPs, the same is true under the non constant-satisfiable assumption. But for positive CSPs over larger alphabets, the Boolean characterization fails in general: explicit languages over $\Sigma$0 and $\Sigma$1 have $\Sigma$2 non-redundancy while remaining $\Sigma$3-space streaming tractable (Sharma et al., 23 Apr 2026). This marks NRD as a sharp invariant in the general and Boolean-positive settings, but not a universal proxy for every promise or positivity restriction.

3. Extremal growth, classifications, and reductions in CSP non-redundancy

Recent work develops NRD as an extremal growth function. For Boolean symmetric predicates of arity at most $\Sigma$4, one paper gives a near-complete classification. Its upper-bound mechanism is $\Sigma$5-balancedness, equivalent to the existence of degree-$\Sigma$6 multilinear capturing polynomials and implying $\Sigma$7; its lower-bound mechanism transfers $\Sigma$8 bounds from $\Sigma$9-ary OR when the universal $A\Rightarrow B$0-cube pattern fails to preserve the predicate. The classification resolves every symmetric predicate of arity at most $A\Rightarrow B$1 and all but two predicates of arity $A\Rightarrow B$2, leaving $A\Rightarrow B$3 and $A\Rightarrow B$4 with bounds $A\Rightarrow B$5 (Sharma et al., 13 May 2026).

The arity-$A\Rightarrow B$6 Boolean case has also been studied in full detail modulo one remaining open predicate. Among the $A\Rightarrow B$7 distinct non-trivial Boolean predicates of arity $A\Rightarrow B$8, an algorithmic pipeline based on OR-projection lower bounds and lattice/polynomial upper bounds classifies $A\Rightarrow B$9 automatically. Of the exceptional cases, $A$0 has $A$1 non-redundancy, $A$2 is the first Boolean predicate shown to have non-polynomial NRD asymptotics,

$A$3

and $A$4 remains open between $A$5 and $A$6 (Brakensiek et al., 22 Mar 2026).

The most expansive existence result shows that every rational exponent occurs. For every rational $A$7, there exists a finite predicate $A$8 such that $A$9. The same paper completely classifies conditional non-redundancy for binary predicates by connecting $\tfrac12\log n$0 to high-girth graph extremal problems, develops an algebraic theory of conditional NRD via patterns and fgppp-definitions, and revisits Mal’tsev embeddings as a source of linear NRD, including a predicate whose Mal’tsev embedding comes from the quantum Pauli group rather than an Abelian group (Brakensiek et al., 10 Jul 2025).

Reduction theory has become correspondingly precise. Hypergraph projections and their shrinking factor now provide a framework for transferring super-linear lower bounds: if projected edge sets shrink as $\tfrac12\log n$1, then the target conditional NRD grows as $\tfrac12\log n$2. This sharpens earlier gadget analyses and yields new lower bounds for predicates that were previously out of reach (Brakensiek et al., 18 May 2026). At the same time, NRD is now known to be essentially the correct upper bound for unweighted sparsification: every unweighted instance of $\tfrac12\log n$3 has a sparsifier of size at most the non-redundancy of the complement predicate up to polylogarithmic factors, while the weighted case is governed by chain length rather than NRD (Brakensiek et al., 2024).

4. Rule systems, implications, and proof search

In formal concept analysis and graded dependencies, NRD is a property of implication bases rather than clause families. A graded attribute implication has the form $\tfrac12\log n$4 with $\tfrac12\log n$5, truth degree

$\tfrac12\log n$6

and semantic entailment

$\tfrac12\log n$7

A theory is non-redundant when no proper equivalent subset exists, and its non-redundancy is witnessed by antecedents when each antecedent $\tfrac12\log n$8 is a model of the theory with its own rule removed. In the finite-residuated-lattice setting with arbitrary idempotent truth-stressing hedges, one paper gives a reduction procedure that transforms any theory into an equivalent non-redundant theory with saturated consequents whose non-redundancy is witnessed by antecedents, and from this resolves the existence of general systems of pseudo-intents in the finite case (Vychodil, 2015).

A closely related earlier result handled the globalization hedge in detail. There, every complete set of graded attribute implications can be transformed into a base with saturated consequents and witnessed non-redundancy, allowing pseudo-intents to be read off directly from the antecedents of the transformed rules. The paper emphasizes that saturation preserves equivalence but not necessarily ordinary non-redundancy, which is why the stronger witnessed form is operationally important (Vychodil, 2015).

In saturation-based theorem proving, the same minimality idea is refined into partial redundancy. “Partial Redundancy in Saturation” introduces redundancy formulas $\tfrac12\log n$9 and partial clauses $\mathsf{CSP}(\Gamma)$0, where the formula records exactly which ground instances of $\mathsf{CSP}(\Gamma)$1 are already redundant. This produces a superposition calculus, PaRC, in which redundancy can eliminate not only whole clauses but also specific inference instances. The calculus is refutationally complete, strong enough to capture some standard restrictions of superposition, and in Vampire yielded $\mathsf{CSP}(\Gamma)$2 TPTP problems not previously solved by any prover, including previous versions of Vampire (Hajdu et al., 28 May 2025).

Outside formalized rule systems, the methodology literature offers a broader interpretation. It distinguishes global logical-formal redundancy, mutual non-redundancy among hypotheses, and non-redundancy relative to a theorem’s thesis, while arguing that mathematically valuable assumptions may remain formally redundant if they are explanatory, proof-relevant, or conceptually illuminating. Examples involving Picard’s theorem, Anosov diffeomorphisms, and Pesin’s formula are used to separate necessity for a conclusion from necessity for a proof strategy (Catsigeras, 2017).

5. Coding-theoretic and information-theoretic meanings

In coding theory, non-redundancy is the search for maximal code size, equivalently minimal check overhead, under fixed length and distance. For a $\mathsf{CSP}(\Gamma)$3-ary code,

$\mathsf{CSP}(\Gamma)$4

and the central question is whether nonlinear constructions can achieve smaller redundancy than the best linear codes. “Nonlinear Codes with Low Redundancy” answers this affirmatively for small or constant distance by constructing VT-style nonlinear codes

$\mathsf{CSP}(\Gamma)$5

with

$\mathsf{CSP}(\Gamma)$6

and quasi-linear-time decoding for erasures and adversarial errors. In several regimes, including explicit $\mathsf{CSP}(\Gamma)$7 ranges and a general corollary under the Main MDS conjecture, these nonlinear codes have strictly smaller redundancy than any linear code with the same parameters (Liu et al., 2023).

In exact channel simulation, the word “non-redundancy” refers instead to the absence of excess communication beyond the mutual-information term. For a joint law $\mathsf{CSP}(\Gamma)$8, blocklength-$\mathsf{CSP}(\Gamma)$9 schemes simulate $\Gamma$0 with expected normalized rate $\Gamma$1, and the issue is the gap above $\Gamma$2. “The Redundancy of Non-Singular Channel Simulation” proves a one-shot lower bound via channel simulation divergence $\Gamma$3 and shows that for i.i.d. non-singular channels the asymptotic redundancy is lower bounded by $\Gamma$4 in the paper’s normalized second-order sense. Equivalently, exact simulation of non-singular channels necessarily incurs a $\Gamma$5-scale penalty, so NRD is impossible in this regime (Flamich et al., 23 Jan 2025).

A related information-theoretic literature studies redundancy not as excess rate but as shared information. “Quantifying multivariate redundancy with maximum entropy decompositions of mutual information” does not introduce NRD as a separate primitive, but it formalizes the complementary notions of unique information and unique redundancy. In the bivariate case, redundancy is expressed by

$\Gamma$6

and the multivariate extension uses rooted-tree decompositions consistent with the redundancy lattice (Chicharro, 2017). This suggests a broader conceptual relation: some uses of NRD minimize dispensable symbols or clauses, while others isolate the part of information not replicated elsewhere.

6. Applied and geometric reinterpretations

In non-autoregressive translation, the relevant phenomenon is not formal non-redundancy of constraints but duplication of meaning in outputs. “On the Information Redundancy in Non-Autoregressive Translation” argues that the traditional continuous repetition ratio is too narrow because modern NAT systems often avoid adjacent exact repeats while still duplicating information via synonymy or reordering. The paper distinguishes continuous repetition, discontinuous repetition, continuous synonym, and discontinuous synonym, and groups the latter three under a broader information-redundancy perspective. It proposes automatic continuous and discontinuous redundancy ratios, $\Gamma$7 and $\Gamma$8, using synonym detection from pretrained mBART embeddings, and reports that OaXE and GLAT mainly reduce continuous redundancy, whereas DAT can eliminate continuous repetition yet still exhibit substantial discontinuous redundancy (Wang et al., 2024).

In isonemal-fabric theory, the phrase is different again: “non-twilly redundancy” denotes a redundancy configuration for perfect colouring in which same-colour warp–weft crossings do not form a twill or doubled twill. The paper derives a symmetry criterion using the $\Gamma$9 groups of the fabric and the derived prefabric, proves that thin striping with $\mathsf{NRD}_n(\Gamma)$0 colours and satin redundancy works for species $\mathsf{NRD}_n(\Gamma)$1 through $\mathsf{NRD}_n(\Gamma)$2 in appropriate orders, that thick striping with doubled satin redundancy works only for species $\mathsf{NRD}_n(\Gamma)$3, and that the isonemal fabric $\mathsf{NRD}_n(\Gamma)$4 serves as a non-twilly redundancy configuration for certain $\mathsf{NRD}_n(\Gamma)$5-colour thin-striping constructions (Thomas, 2024). Here, “NRD” does not mean elimination of redundancy; it names a structured way of placing redundant cells so that colouring remains symmetry-compatible.

Queueing and distributed-systems work uses non-redundancy as a benchmark against which explicit redundancy is compared. In heterogeneous data centers, the non-redundant baseline is Bernoulli routing, where each arrival is sent to exactly one uniformly chosen compatible server; its stability threshold is

$\mathsf{NRD}_n(\Gamma)$6

Under full redundancy, stability is governed instead by a recursive least-loaded-subsystem construction with threshold $\mathsf{NRD}_n(\Gamma)$7. The paper shows that sufficiently strong heterogeneity can make redundancy strictly better than the non-redundant benchmark, with simple sufficient conditions such as $\mathsf{NRD}_n(\Gamma)$8 in redundancy-$\mathsf{NRD}_n(\Gamma)$9 systems and $n$0 under a linear capacity profile (Anton et al., 2020).

A complementary scheduling paper studies not whether to replicate, but how to avoid harmful overlap among replicas. It introduces the overlap variable $n$1 between replica sets of two jobs, the indicators

$n$2

and a block-design-based policy using $n$3-BIBDs. The resulting incidence graph has spectral gap $n$4, compared with a vanishing asymptotic gap for round-robin, and simulations report average waiting-time reductions of up to $n$5 relative to random scheduling and up to $n$6 relative to round-robin (Behrouzi-Far et al., 2022). A plausible implication is that in applied systems the practical counterpart of NRD is often not zero duplication, but disciplined control of which duplications remain indispensable and which merely waste resources.