1-Limited Automata and Regular Languages

• 1-Limited automata are one-tape, two-way finite-control devices that rewrite non-endmarker cells only upon their first visit, precisely characterizing regular languages.
• They enable exponential and doubly exponential state compression compared to standard finite automata, highlighting significant descriptional complexity.
• Multiple variants—including deterministic, nondeterministic, unambiguous, and probabilistic models—offer diverse algorithmic properties and simulation techniques bridging finite automata and context-free mechanisms.

Searching arXiv for recent and foundational papers on 1-limited automata and related variants. A 1-limited automaton (1-LA) is the $d=1$ or $k=1$ specialization of Hibbard’s limited-automaton model: a single-tape, two-way finite-control device in which each non-endmarker tape cell may be rewritten only on its first visit and is frozen thereafter (Pighizzini, 2015, Yamakami, 2021, Pighizzini et al., 2023). In language-theoretic terms, 1-LAs characterize exactly the regular languages, but their operational behavior and descriptional complexity are considerably richer than that slogan suggests: the model sits at the exact threshold between two-way finite-state behavior and the context-free power already obtained for larger rewrite budgets, and it supports succinct encodings that can be exponentially or doubly exponentially smaller than equivalent ordinary finite automata (Pighizzini, 2015, Pighizzini et al., 2023).

1. Formal definition and operational semantics

A standard formulation presents a 1-LA as a tuple

$\mathcal A=(Q,\Sigma,\Gamma,\delta,q_I,F),$

where $Q$ is a finite state set, $\Sigma$ is the input alphabet, $\Gamma$ is a work alphabet containing $\Sigma$ and the end-markers, $\delta$ moves the head one cell left or right while possibly rewriting the scanned symbol, $q_I$ is the initial state, and $F$ is the set of final states. On input $k=1$0, the tape initially contains the input between left and right end-markers, the head starts on the first input cell, and acceptance is by a successful computation that reaches a final state and then passes the right end-marker. The distinctive restriction is that each non-endmarker cell may be modified only during its first visit; all later visits are read-only on that cell (Pighizzini et al., 2023, Pighizzini et al., 2023).

Two equivalent technical presentations are common. In layer-based definitions, the tape alphabet is partitioned into $k=1$1 and $k=1$2, with end-markers in $k=1$3; symbols in $k=1$4 are immutable. In rank-based deterministic formulations, input symbols start with rank $k=1$5, end-markers have rank $k=1$6, and a first visit to a rank-$k=1$7 symbol must rewrite it to rank $k=1$8, after which it cannot be changed again. In the probabilistic $k=1$9-limited formalism specialized to $\mathcal A=(Q,\Sigma,\Gamma,\delta,q_I,F),$0, turns count as double visits, and a fresh symbol in $\mathcal A=(Q,\Sigma,\Gamma,\delta,q_I,F),$1 can be legally rewritten only when the head leaves to the right; a turn at a fresh cell immediately exhausts the rewriting budget for that cell (Yamakami, 2021, Rubtsov, 2023).

2. Position in the limited-automata hierarchy

The central language-theoretic fact is that the class of languages accepted by $\mathcal A=(Q,\Sigma,\Gamma,\delta,q_I,F),$2-limited automata coincides with the class of regular languages. This places 1-LAs at the bottom nontrivial level of the limited-automata hierarchy. In the same hierarchy, limited automata with rewrite budget $\mathcal A=(Q,\Sigma,\Gamma,\delta,q_I,F),$3 already characterize the context-free languages, while the case $\mathcal A=(Q,\Sigma,\Gamma,\delta,q_I,F),$4 collapses to ordinary two-way finite automata (Pighizzini, 2015, Pighizzini et al., 2023).

Regime	Characterization
$\mathcal A=(Q,\Sigma,\Gamma,\delta,q_I,F),$5 limited automata	two-way finite automata
$\mathcal A=(Q,\Sigma,\Gamma,\delta,q_I,F),$6	regular languages
$\mathcal A=(Q,\Sigma,\Gamma,\delta,q_I,F),$7	context-free languages
deterministic $\mathcal A=(Q,\Sigma,\Gamma,\delta,q_I,F),$8	deterministic context-free languages

Historically, this threshold goes back to Hibbard’s 1967 scan limited automata. Later work, including formulations closer to Pighizzini and Pisoni, uses explicit end-markers and a bounded accessible tape segment rather than Hibbard’s original semi-infinite blank-filled tape. The same line of work recalls a sharper collapse specific to $\mathcal A=(Q,\Sigma,\Gamma,\delta,q_I,F),$9: nondeterministic 1-limited automata are equivalent in power to $Q$0-way deterministic finite automata. This is consistent with the regular-language characterization and emphasizes that, at $Q$1, neither nondeterminism nor one-shot rewriting yields pushdown-like memory. By contrast, deterministic $Q$2-limited automata form a strict hierarchy, with deterministic $Q$3-limited automata already characterizing deterministic context-free languages (Yamakami, 2021, Rubtsov, 2023).

3. Probabilistic, deterministic, nondeterministic, and unambiguous variants

The modern theory treats 1-LAs as one point in a family of machine types obtained by imposing different restrictions on the same underlying one-tape mechanism. Besides deterministic 1-LAs and nondeterministic 1-LAs, the literature defines unambiguous 1-LAs and probabilistic 1-limited automata (1-lpa’s). For a probabilistic machine $Q$4, the standard quantities are $Q$5 and $Q$6, with unbounded-error acceptance criterion

$Q$7

and expected polynomial running time. One-sided error means that for every input $Q$8, either $Q$9 or $\Sigma$0; bounded error means that there exists $\Sigma$1 such that every input is accepted or rejected with probability at least $\Sigma$2, according to truth. The key theorem for $\Sigma$3 identifies bounded-error, one-sided-error, and unbounded-error probabilistic 1-LAs exactly with the corresponding classes of $\Sigma$4-way probabilistic finite automata, while the introduction separately recalls that $\Sigma$5-lna’s are equivalent in power to $\Sigma$6dfa’s. The proof idea is bidirectional: a 1-lpa can simulate a $\Sigma$7pfa by rewriting each fresh symbol to a marked copy, and a $\Sigma$8pfa can simulate a normalized 1-lpa because the unique correspondence between original and overwritten symbols can be folded into finite control. The same paper proves that, for every $\Sigma$9, one-sided-error limited-automaton families are included in bounded-error families, hence also at $\Gamma$0. It also notes a methodological limitation: the blank-skipping normal form and reversal machinery developed for $\Gamma$1 are not established as general robustness theorems for the $\Gamma$2 case (Yamakami, 2021).

4. Regular-language power versus descriptional power

A recurrent misconception is that “1-LAs recognize only regular languages” implies that they are merely a cumbersome presentation of finite automata. The modern descriptional-complexity results show the opposite. General 1-LAs can be doubly exponentially more succinct than one-way deterministic finite automata, even though their languages remain regular. A standard simulation still converts every $\Gamma$3-state 1-LA into a one-way NFA with at most

$\Gamma$4

states, so regularity is not in question; what changes is the cost of removing two-way motion, rewriting, and nondeterministically chosen first-visit annotations (Pighizzini et al., 2023, Guillon et al., 15 Jul 2025).

The source of this succinctness is the interaction between nondeterminism and first-visit rewriting. On its first visit to a cell, a 1-LA may nondeterministically choose what to write there; later transitions may depend on that frozen choice. Earlier work already emphasized that the finite-automaton simulation proceeds by transition-table techniques related to standard conversions of two-way automata into one-way automata, rather than by pumping-lemma arguments. In this sense, 1-LAs are best viewed not as a stronger LLM than finite automata, but as a regular-LLM with unusually strong compression mechanisms (Pighizzini, 2015, Pighizzini et al., 2023).

5. Restricted 1-LA models and the structure of succinctness

Several papers isolate which aspects of first-visit rewriting are responsible for the extreme succinctness of unrestricted 1-LAs. In once-marking 1-LAs, each computation marks exactly one input cell and changes no other cell. In always-marking 1-LAs, every first visit must replace $\Gamma$5 by its marked copy $\Gamma$6. Both models still recognize only regular languages, but their succinctness profiles differ sharply. Once-marking 1-LAs preserve the full double-exponential gap to one-way deterministic finite automata; indeed, the language family $\Gamma$7 is accepted by an $\Gamma$8-1-LA with $\Gamma$9 states, and the general conversion costs are exponential to 2NFA and 1NFA, but double exponential to 1DFA. By contrast, always-marking 1-LAs lose the branch-dependent rewriting mechanism that causes the second exponent: each $\Sigma$0-state always-marking machine can be simulated by a 1NFA with at most $\Sigma$1 states and by a 1DFA with at most $\Sigma$2 states, while deterministic always-marking machines admit simulations by 2DFA and 1DFA with no more than $\Sigma$3 states. Deterministic once-marking is even closer to ordinary two-way determinism: every $\Sigma$4-state deterministic once-marking 1-LA has an equivalent 2DFA with $\Sigma$5 states (Pighizzini et al., 2023).

A different restriction is the forgetting 1-limited automaton, in which every first-visited input symbol is overwritten by a single fixed symbol $\Sigma$6. The work alphabet is then

$\Sigma$7

and if $\Sigma$8 with $\Sigma$9, then necessarily $\delta$0. This model still characterizes the regular languages, but its simulation costs are substantially smaller than those of unrestricted 1-LAs. The key quantitative parameter is Landau’s function

$\delta$1

with asymptotic estimate $\delta$2. An $\delta$3-state forgetting 1-LA can be simulated by a 1NFA with at most $\delta$4 states and by a complete 1DFA with at most $\delta$5 states; for deterministic forgetting 1-LAs, the 1DFA upper bound improves to $\delta$6. The same work also proves an exponential separation from two-way deterministic automata: the language $\delta$7 is accepted by a 2DFA with $\delta$8 states, but every forgetting 1-LA for $\delta$9 requires at least $q_I$0 states (Pighizzini et al., 2023).

6. Algorithms, complementation, common-guess models, and recent unary results

Although 1-LAs are usually discussed as language-recognition devices, they also support nontrivial algorithmic simulation results. For deterministic $q_I$1-limited automata, membership can be solved in time

$q_I$2

where $q_I$3 is the length of the machine description and $q_I$4 is the input length, provided $q_I$5 is computable in $q_I$6 time. Specializing to $q_I$7, deterministic 1-LAs therefore admit an $q_I$8 RAM simulation, and an $q_I$9 algorithm for each fixed machine. The simulation compresses maximal contiguous blocks of frozen cells into finite traversal summaries, in the style of Shepherdson’s simulation of two-way finite automata (Rubtsov, 2023).

Complementation has recently been analyzed inside the 1-LA model itself. A central construction starts from an arbitrary one-way NFA and produces an equivalent self-verifying $F$0NFA with common guess having $F$1 states and $F$2 annotation symbols. The common-guess mechanism can be viewed as a restricted 1-LA that writes an annotation once and then performs a read-only two-way verification. Combining this with the known $F$3-state simulation of 1-LAs by 1NFAs yields a general complementation theorem: for each $F$4-state 1-LA recognizing $F$5, there exists a 1-LA with a single exponential number of states in $F$6 and $F$7 work symbols recognizing the complement of $F$8. The same paper states that this single-exponential bound is necessary and sufficient (Guillon et al., 15 Jul 2025).

Recent unary lower bounds sharpen the role of nondeterministic annotation even further. Two-way finite automata with common guess, denoted $F$9nfas+cg and $k=1$00dfas+cg, perform an initial annotation phase and then a read-only two-way computation; their deterministic counterparts still have a nondeterministic annotation phase and can be considered a restriction of 1-LAs. For unary languages, there are exponential lower bounds for simulating $k=1$01dfas+cg by deterministic 1-LAs and by 2NFAs, derived from a doubly exponential lower bound for simulating $k=1$02dfas+cg by 1DFAs. These unary lower bounds close the question, left open in earlier work by Pighizzini and Prigioniero, of whether a double exponential gap between 1-LAs and 1DFAs exists in the unary case. The same abstract also states an exponential lower bound for complementing unary $k=1$03dfas+cg and thus unary 1-LAs (Guillon et al., 11 Apr 2025).