Linear Automata: Concepts and Applications

• Linear automata are computational models with linear dynamics governing state transitions across algebraic, geometric, arithmetic, and order-theoretic structures.
• They encompass varied forms including cellular automata over groups, two-head automata for linear languages, and module-valued systems in decomposition theory.
• Applications range from cryptographic sequence generation and event-driven switched systems to automata for temporal logics and linear integer arithmetic.

Linear automata is a heterogeneous term rather than a single canonical formalism. In current arXiv usage, it denotes several distinct but mathematically related traditions: automata with linear local dynamics over vector spaces or fields, two-head inward-reading automata for linear languages, module-valued automata studied by decomposition theory, and automata-theoretic constructions for linear-time logics or linear arithmetic. What these traditions share is not a common acceptance model, but the presence of a linear structure—algebraic, geometric, arithmetic, or order-theoretic—that governs transition behavior, expressiveness, or decomposition (Bartholdi, 2016, Bedregal, 2016, Plotkin et al., 2015, Smith et al., 2021).

1. Terminological scope and major meanings

In the literature represented here, “linear automata” is used in several non-equivalent senses.

Usage	Core object	Source of linearity
Linear cellular automata over groups	$\Theta:V^G\to V^G$	$\Bbb K$-linearity and matrices over $\Bbb KG$
Deterministic/nondeterministic linear automata	Two-head automata reading from both ends	Recognition of linear languages
Linear automata in decomposition theory	$\Lambda=(A,\Gamma,B)$	Module structure and linear actions
Automata for linear-time logics	NFAs/transducers over traces or linear orders	Linear-time semantics, not linear state evolution
Register or arithmetic automata with linear arithmetic	Register automata over $\mathbb Q$ or LIA derivatives	Affine-linear updates or arithmetic residuals

Two nearby notions are explicitly different. First, the linear-bounded automata question concerns whether deterministic and nondeterministic LBAs coincide, formulated in one paper as whether $\mathrm{NSPACE}[n]=\mathrm{DSPACE}[n]$; there, “linear” refers to a space bound, not to linear transitions or vector-space structure (Lin, 2021). Second, “Linear functional classes over cellular automata” studies strict real time, synchronous real time, and linear time functional computation by cellular automata; its “linear” refers to time complexity, and the paper explicitly states that there is no use of linearity in the sense of linear automata or linear state transition over semirings or vector spaces (Grandjean et al., 2012).

This terminological dispersion suggests that the phrase is best interpreted contextually. In algebraic automata theory it usually signals a vector-space or module action; in formal-language theory it often denotes the two-head model for linear languages; in temporal-logic papers it typically refers to linear time rather than linear algebra.

2. Algebraic linear automata over groups and cellular dynamics

A central algebraic meaning is the linear cellular automaton over a group. Fix a group $G$, a field $\Bbb K$, and a finite-dimensional $\Bbb K$-vector space $V=\Bbb K^n$. A linear cellular automaton is a $\Bbb K$0-equivariant, continuous linear map

$\Bbb K$1

with local rule

$\Bbb K$2

where $\Bbb K$3 is finite and the coefficients $\Bbb K$4 are $\Bbb K$5-matrices. Equivalently, $\Bbb K$6 is encoded by a matrix in $\Bbb K$7, so the dynamics is transferred to linear algebra over the group ring (Bartholdi, 2016).

A key simplification in this setting is the reduction of dynamical properties to ordinary linear algebra on finitely supported configurations $\Bbb K$8. The paper proves

$\Bbb K$9

and

$\Bbb KG$0

Using the anti-involution $\Bbb KG$1 on $\Bbb KG$2 and the adjoint matrix $\Bbb KG$3, it derives the adjointness identity

$\Bbb KG$4

and from the resulting kernel-image orthogonality obtains the duality principles

$\Bbb KG$5

$\Bbb KG$6

This yields, for every non-amenable group and every field, surjective but non-pre-injective linear cellular automata, answering Open Problem (OP-14) from Ceccherini-Silberstein and Coornaert positively. The same duality gives a direct linear proof that over sofic groups every post-surjective linear cellular automaton is pre-injective (Bartholdi, 2016).

A further development studies multiband linear cellular automata over $\Bbb KG$7. Here the alphabet is $\Bbb KG$8, the configuration space is $\Bbb KG$9, and a local rule is encoded by a Laurent polynomial matrix

$\Lambda=(A,\Gamma,B)$0

The paper establishes a correspondence between certain such automata and endomorphisms of the vector group $\Lambda=(A,\Gamma,B)$1 over $\Lambda=(A,\Gamma,B)$2, with a universal embedding $\Lambda=(A,\Gamma,B)$3 that sends Frobenius to shift. Under a confinement hypothesis, this algebraic correspondence yields fixed-point counts, a dichotomy for the Artin–Mazur dynamical zeta function, and an asymptotic formula for periodic orbits; it also applies to higher-order linear automata because those can be rewritten as multiband first-order systems (Byszewski et al., 2022).

A cryptographic strand treats one-dimensional binary 90/150 cellular automata as discrete linear models for sequence generation. Rule 90 and rule 150 are

$\Lambda=(A,\Gamma,B)$4

Using the Cattell–Muzio synthesis algorithm and repeated concatenation of a basic automaton with its reversal, the paper constructs automata with characteristic polynomials $\Lambda=(A,\Gamma,B)$5 and shows that, for suitable $\Lambda=(A,\Gamma,B)$6, these generate all solutions of the linear binary difference equation with characteristic polynomial $\Lambda=(A,\Gamma,B)$7. The intended significance is cryptographic: many pseudo-random keystream sequences used in symmetric cryptography can be represented within this linear cellular-automaton framework (Caballero-Gil et al., 2010).

3. Two-head linear automata and the theory of linear languages

A second major meaning of linear automata is the two-head model for linear languages. In this tradition, a linear automaton has two reading heads starting from the two ends of the input and moving inward until they meet. One formulation uses a 5-tuple

$\Lambda=(A,\Gamma,B)$8

with transition relation

$\Lambda=(A,\Gamma,B)$9

A configuration is $\mathbb Q$0, where $\mathbb Q$1 is the unread middle. A transition consumes either one symbol from the left or one symbol from the right, and acceptance occurs when the unread part becomes empty (Nagy, 21 Jul 2025).

A closely related formulation uses two disjoint state sets $\mathbb Q$2 and $\mathbb Q$3 indicating which head is active. The paper on subclasses of linear languages defines $\mathbb Q$4-nondeterministic linear automata, proves equivalence with linear grammars, and then proves that $\mathbb Q$5-moves do not increase power: every $\mathbb Q$6-nondeterministic linear automaton can be converted to a $\mathbb Q$7-free nondeterministic linear automaton accepting the same language. Deterministic linear automata are obtained by requiring $\mathbb Q$8 for each state-symbol pair, and even linear automata are the bipartite restriction in which transitions alternate between $\mathbb Q$9 and $\mathrm{NSPACE}[n]=\mathrm{DSPACE}[n]$0; these capture even linear languages (Bedregal, 2016).

The deterministic language class is subtle. One paper denotes by 2detLIN the class accepted by complete deterministic linear automata and states that it is a proper superclass of the regular languages and a proper subset of the linear languages. It also notes equivalence with sensing $\mathrm{NSPACE}[n]=\mathrm{DSPACE}[n]$1 Watson–Crick finite automata. Determinism here means that at any given state only one head may be active: for every state, either all left-reading transitions are absent and right-reading transitions are single-valued, or the converse (Nagy, 21 Jul 2025).

Because such automata can consume both a prefix and a suffix before reaching the middle, ordinary Myhill–Nerode equivalence on prefixes is insufficient. The relevant objects are prefix-suffix pairs $\mathrm{NSPACE}[n]=\mathrm{DSPACE}[n]$2, called presus in the paper, with equivalence defined by

$\mathrm{NSPACE}[n]=\mathrm{DSPACE}[n]$3

The resulting characterization states that a language belongs to 2detLIN iff there exists a complete border classification of finite index with no crossing pairs. Unlike the DFA case, a single equivalence class may require up to two states, one for each possible active head (Nagy, 21 Jul 2025).

The same tradition also supports a graded notion of nondeterminism. If $\mathrm{NSPACE}[n]=\mathrm{DSPACE}[n]$4 denotes the explicit nondeterminism degree of a nondeterministic linear automaton $\mathrm{NSPACE}[n]=\mathrm{DSPACE}[n]$5, then

$\mathrm{NSPACE}[n]=\mathrm{DSPACE}[n]$6

This gives an infinite hierarchy from deterministic linear languages to all linear languages (Bedregal, 2016).

4. Linear automata as module-valued automata and decomposition objects

A third sense is the algebraic one used in decomposition theory. A linear automaton is a triple

$\mathrm{NSPACE}[n]=\mathrm{DSPACE}[n]$7

where $\mathrm{NSPACE}[n]=\mathrm{DSPACE}[n]$8 and $\mathrm{NSPACE}[n]=\mathrm{DSPACE}[n]$9 are $G$0-modules and $G$1 is a semigroup acting on $G$2, from $G$3 to $G$4, and on $G$5. The operations

$G$6

satisfy

$G$7

$G$8

$G$9

The motivating representation is by block upper-triangular matrices in

$\Bbb K$0

This is the framework in which the paper develops a linear analogue of Krohn–Rhodes theory (Plotkin et al., 2015).

The basic decomposition operations are threefold: the triangular product of linear automata, the wreath product of pure automata, and the wreath product of a linear automaton with a pure automaton. The triangular product is the terminal object in the category of cascade connections of linear automata, just as the wreath product is the terminal object in the Krohn–Rhodes theory for pure automata. For faithful representations $\Bbb K$1 and $\Bbb K$2, the triangular product

$\Bbb K$3

uses triangular matrices

$\Bbb K$4

For linear automata, an analogous upper-triangular construction yields $\Bbb K$5 (Plotkin et al., 2015).

This theory defines the complexity of a linear automaton as the minimal number of decomposition operations needed to obtain indecomposable components, or atoms. If $\Bbb K$6 and $\Bbb K$7 have composition series of lengths $\Bbb K$8 and $\Bbb K$9, then $\Bbb K$0 is linearly decomposable into $\Bbb K$1 indecomposable factors, all of which are semi-automata. After compression to completely $\Bbb K$2-simple semigroups and further wreath-product reduction, the terminal atoms are irreducible representations of simple groups, together with the pure simple-group components inherited from Krohn–Rhodes theory (Plotkin et al., 2015).

The decomposition viewpoint turns “linear automaton” into a structural rather than language-theoretic notion. It studies how module actions and semigroup actions interact under divisibility, cascade connection, compression, and atomic factorization.

5. Linear time, linear orders, and arithmetic automata

Several papers use automata in settings where “linear” refers primarily to time or arithmetic structure. Automata Linear Dynamic Logic on Finite Traces (ALDL$\Bbb K$3) replaces the regular-expression path modalities of LDL$\Bbb K$4 with NFA-based path automata

$\Bbb K$5

whose labels may be propositional formulas, past-labeled formulas, or tests $\Bbb K$6. ALDL$\Bbb K$7 is expressively equivalent to LDL$\Bbb K$8, hence to monadic second-order logic over finite traces, and satisfiability remains PSPACE-complete. The key technical path is a translation from ALDL$\Bbb K$9 formulas to two-way alternating automata on finite words and then to NFAs (Smith et al., 2021).

A different construction studies automata over arbitrary linear orderings. There a word has length $V=\Bbb K^n$0 for a linear ordering $V=\Bbb K^n$1, and runs are indexed not by positions but by the set of cuts $V=\Bbb K^n$2. The automata include successor transitions and also left and right limit transitions to handle cuts without immediate predecessor or successor. This supports a translation from LTL with Until, Since, and the Stavi connectives into non-ambiguous transducers that compute truth words. The constructed automaton has exponentially many states in the formula size, and if limit transitions are counted explicitly its size can be doubly exponential; the paper describes the resulting construction as a 2ExpSpace procedure (Cristau, 2011).

The phrase also appears in proof theory for LTL. The paper on linear factors defines an LTL analogue of Antimirov’s linear factors, where a formula is decomposed into a monomial for the current letter and a temporal residual for the suffix. It proves an expansion theorem and finiteness of descendants, then shows that Vardi’s alternating $V=\Bbb K^n$3-automaton construction and Wolper’s semantic tableaux are isomorphic presentations of the same factor structure (Sulzmann et al., 2017).

On the arithmetic side, the technical report on quantified linear integer arithmetic combines automata with algebraic reasoning by making every automaton state a derivative formula. Integers are encoded in LSBF two’s-complement, and derivative states are simplified using arithmetic rewriting, disjunction pruning, quantifier instantiation, range analysis, and modulo linearization. The prototype Amaya is evaluated on 372 SMT-COMP formulas after filtering and on a 55-instance Frobenius coin benchmark; the paper reports that it is competitive with, and often superior to, several SMT solvers on quantified instances (Habermehl et al., 2024).

A related but distinct model is register automata over the rationals (RA-Q). Here “linear” means affine-linear arithmetic on stored rational values: control variables govern order and equality tests, data variables are updated by linear expressions, and outputs are linear expressions. The model generalizes classical register automata, affine programs, and arithmetic circuits while retaining decidability of the invariant problem (Chen et al., 2017).

6. Applications, adjacent models, and recurring misconceptions

Automata-theoretic linear structure also appears in application-driven models. In event-driven switched linear systems, a black-box switched system is represented by a deterministic finite automaton

$V=\Bbb K^n$4

whose node labels are subsystem matrices $V=\Bbb K^n$5. The paper combines matrix reconstruction from trajectory queries with an Angluin-style $V=\Bbb K^n$6 procedure, proving termination and correctness under full-rank and oracle assumptions; benchmark examples include systems with 1000 states, 19 events/labels, $V=\Bbb K^n$7, and 2000 states, 9 events/labels, $V=\Bbb K^n$8 (Kundu et al., 2020).

In linear automata networks, a network is linear if $V=\Bbb K^n$9 over a finite ring or field. For field-linear networks, expansivity is characterized by nonsingularity of certain trace matrices $\Bbb K$00, and a digraph admits expansive networks iff it is strong and coverable. For sufficiently large prime powers, expansive linear networks exist on every strong coverable graph, while fixed alphabet size imposes strong limitations (Bridoux et al., 2019).

Several recurrent misconceptions follow directly from the literature. Linear-bounded automata are about linear space, with deterministic LBAs corresponding to $\Bbb K$01 and nondeterministic LBAs to $\Bbb K$02; they are not a species of algebraic linear automaton (Lin, 2021). Likewise, linear-time cellular automata and linear-time logics concern temporal complexity or trace structure rather than vector-space dynamics (Grandjean et al., 2012, Smith et al., 2021). A plausible implication is that “linear automata” functions more as a family resemblance term than as a universally standardized class name.

Across these traditions, the term marks one of three recurring design patterns: inward-reading control for linear languages, algebraic linearity of states or updates, or automata constructions indexed by linear time or arithmetic residuals. The research landscape is therefore unified less by a single definition than by a stable cluster of linear methods applied to automata-theoretic problems.