Boustrophedon Automata in Picture Languages

• Boustrophedon automata are finite-state models that traverse picture inputs in alternating left-right directions, converting 2D data into linear serializations.
• They extend to different geometries, including hexagonal grids, and relate closely to returning and typewriter automata through formal language equivalences.
• This model bridges classical automata theory with DAG encodings, influencing pattern recognition and offering insights into specialized coverage strategies.

Searching arXiv for recent and foundational papers on boustrophedon automata and closely related models. Searching for "boustrophedon automata picture languages" on arXiv. Boustrophedon automata are scan-based automata whose defining operation is traversal in alternating directions on successive lines, “like an ox plowing a field.” In the standard picture-language setting, a boustrophedon finite automaton (BFA) recognizes rectangular symbol arrays by inducing a fixed total order on the cells of a boundary-padded picture and processing the resulting serialization with finite-state control. Recent work places this model inside a DAG-automata framework, extends it to hexagonal pictures, and relates it to returning and typewriter automata; at the same time, several adjacent formalisms exhibit boustrophedon-like dynamics without being standard BFAs in the picture-automata sense (Meeres et al., 15 Sep 2025, D et al., 11 Aug 2025, Smith, 2022).

1. Core idea and historical placement

In the picture-automata literature, a BFA is a finite-state recognizer for two-dimensional words or pictures. Its essential feature is not a local neighborhood rule, but a scanning strategy: one row is read left-to-right, the next right-to-left, then left-to-right again, and so on. This alternating rowwise traversal is the source of the term boustrophedon.

The 2022 study of typewriter automata notes two distinct historical layers. First, the name “boustrophedon automaton” was originally used by Pécuchet in the 1980s for a one-dimensional model. Second, the two-dimensional BFA relevant here was introduced by Fernau et al. in 2015 and developed further in later work. In that same literature, boustrophedon automata are compared with returning automata, three-way automata, and other picture-language devices (Smith, 2022).

A recurrent source of confusion is that “boustrophedon” names a traversal discipline, not a single universally fixed machine formalism. In the standard rectangular setting, the model is a picture automaton. In later work on hexagonal pictures, the same alternating-line idea is transplanted to a different geometry. In still other settings, the term functions only as an analogy for ordered sweeps or alternating scan orientation rather than as the name of a formal automaton.

2. Rectangular boustrophedon finite automata

A recent unifying restatement defines a deterministic or nondeterministic BFA as a 6-tuple

$A=(Q,\Sigma,\delta,q_0,F,\#),$

where transitions are over $\Sigma\cup\{\#\}$. The input picture is surrounded on all four sides by the boundary symbol $\#$, and the scanning strategy defines a total order on the cells of that boundary picture. Acceptance is then delegated to the underlying finite automaton acting on the induced string serialization (Meeres et al., 15 Sep 2025).

This presentation makes explicit two structural facts. First, a BFA is fundamentally a one-dimensional finite automaton applied to a two-dimensional input through a fixed scan order. Second, the boundary is part of the formal input representation, not merely an informal convention. The 2025 paper stresses that its definition is an alternative but equivalent recasting chosen to expose both the boundary symbols and the scanning strategy more clearly.

An older operational presentation, used in the comparison with typewriter automata, packages the model as a tuple with a rule set $R \subseteq Q \times (\Sigma \cup \{\#\}) \times Q$, an accepting state, the boundary symbol $\#$, and an erasure marker $\square$. In that formulation, the marker $\square$ records already-read cells, and acceptance requires more than entering an accepting state: an $m\times n$ picture is accepted only if some accepting computation ends with all $m\times n$ cells replaced by $\square$. This makes explicit that the automaton must read the entire input before accepting (Smith, 2022).

These two descriptions emphasize different aspects of the same model family. The serialization-oriented formulation highlights the induced string language, whereas the erasure-oriented formulation highlights the operational fact that the entire picture is consumed by the scan.

3. Expressive power and equivalence landscape

Within classical picture automata, BFAs occupy a tightly characterized position. The following relations are stated in the recent literature (Meeres et al., 15 Sep 2025, Smith, 2022).

Relation	Status	Informal significance
$\Sigma\cup\{\#\}$0	equality	Returning and boustrophedon scans define the same picture-language family
deterministic BFA = nondeterministic BFA	equality	Determinization does not reduce recognition power
1D restriction of BFA	regular languages exactly	On pictures of height $\Sigma\cup\{\#\}$1, the model collapses to ordinary regular string recognition
$\Sigma\cup\{\#\}$2	equality	Nondeterministic typewriter automata and BFAs recognize the same class
$\Sigma\cup\{\#\}$3	strict inclusion	Three-way nondeterministic 2D automata are strictly stronger

The equality $\Sigma\cup\{\#\}$4 is one of the basic classical facts recalled in the DAG-automata paper. The same paper also states that deterministic and nondeterministic RFAs, and likewise deterministic and nondeterministic BFAs, have the same expressive power. A further collapse occurs in one dimension: when these picture automata are run on $\Sigma\cup\{\#\}$5, they recognize exactly the regular string languages (Meeres et al., 15 Sep 2025).

The typewriter-automata paper refines this landscape. Its central comparison theorem states that nondeterministic typewriter automata and boustrophedon automata recognize the same language class, $\Sigma\cup\{\#\}$6. The proof runs through returning automata: a typewriter automaton can simulate a returning scan by resetting only at row ends, while a returning automaton can simulate an early typewriter reset by continuing to the right boundary and ignoring the skipped suffix before making a deferred reset. The same paper also states that the deterministic analogue fails: the language $\Sigma\cup\{\#\}$7, consisting of pictures whose first column and last row are all $\Sigma\cup\{\#\}$8s and all other entries are $\Sigma\cup\{\#\}$9, is recognized by a boustrophedon automaton but not by any deterministic typewriter automaton (Smith, 2022).

The same source reports standard closure properties for the BFA language family: closure under union, intersection, complement, row concatenation, and reversal, and nonclosure under column concatenation and rotation. Together with the strict inclusion into nondeterministic three-way automata, these facts situate BFAs as robust but deliberately limited scan-based picture automata rather than general two-dimensional local-computation devices.

4. DAG encodings and the serialization viewpoint

A major 2025 development is the embedding of BFAs into a unifying DAG-automata framework. In that approach, a picture is encoded as a directed acyclic graph, and a deterministic DAG automaton recognizes the encoded structure. For BFAs, the relevant encoding is input-agnostic and produces a string DAG: a path that snakes through the picture in alternating row directions, with vertical connectors at the border where the scan turns into the next row. The paper’s Theorem 4.3 states that RFAs and BFAs recognize exactly the same picture languages as deterministic DAG automata run on the corresponding RFA and BFA encodings (Meeres et al., 15 Sep 2025).

This characterization sharpens the standard intuition. A BFA does not derive its power from a genuinely two-dimensional local dependency relation. Its computation is instead finite-state processing over a specific serialization topology. In the DAG formulation, the crucial object is the path-shaped encoding itself. Once the picture has been turned into that path, recognition is equivalent to DFA computation on the same underlying string.

The same paper uses this framework to illuminate both strengths and limitations. BFAs and RFAs are described there as the “most limited” classical picture automata under discussion. They naturally recognize horizontal regularities aligned with their scan order; the language of horizontal stripes with adjacent rows of different colors is the main example. By contrast, the paper argues that diagonal patterns are awkward for a fixed rowwise scan unless one changes the scanning direction to suit the language. It explicitly observes that the scanning strategy may partly determine which languages are easy to recognize, to the point that changing the scan to fit the target language can amount to a kind of “cheating.” This makes the serialization bias of BFAs a central theoretical fact rather than a mere implementation detail.

The comparison with online tessellation automata reinforces the point. BFA encodings are path-like and therefore impose a total order, whereas tessellation-style models rely on genuinely two-dimensional partial orders with local dependence on two predecessors. The difference is not simply one of notation; it is a difference in computational geometry.

5. General hexagonal boustrophedon finite automata

The 2025 paper on hexagonal picture scanning automata extends the boustrophedon idea from rectangular grids to hexagonal pictures. Its basic input objects are finite hexagonal pictures over an alphabet $\#$0, with $\#$1 denoting all such pictures and $\#$2 the non-empty ones. A picture of size $\#$3 has a bordered version of size $\#$4, using the boundary symbol $\#$5 (D et al., 11 Aug 2025).

A General Hexagonal Boustrophedon Finite Automaton (GHBFA) is defined as

$\#$6

with a state partition $\#$7. Transitions on ordinary symbols preserve the phase: forward transitions stay within $\#$8, backward transitions stay within $\#$9. Border transitions have the form $R \subseteq Q \times (\Sigma \cup \{\#\}) \times Q$0 with $R \subseteq Q \times (\Sigma \cup \{\#\}) \times Q$1, thereby encoding the alternation of scan direction at line boundaries. As in the rectangular operational formulation, scanned symbols are replaced by $R \subseteq Q \times (\Sigma \cup \{\#\}) \times Q$2 (D et al., 11 Aug 2025).

The distinctive geometric ingredient is the direction set $R \subseteq Q \times (\Sigma \cup \{\#\}) \times Q$3, consisting of 12 direction modes represented by $R \subseteq Q \times (\Sigma \cup \{\#\}) \times Q$4 matrices. Each mode specifies a starting corner, an initial reading direction, and whether the shift to the next line proceeds clockwise or anticlockwise at each turn. The scan therefore alternates line direction while adapting to hexagonal geometry rather than rectangular rows. The paper emphasizes that the next-line movement is not chosen freely by the automaton; it is prescribed by the selected direction mode.

Several structural results parallel the rectangular theory. For each fixed direction mode $R \subseteq Q \times (\Sigma \cup \{\#\}) \times Q$5, determinization holds: $R \subseteq Q \times (\Sigma \cup \{\#\}) \times Q$6 The paper also introduces General Hexagonal Returning Finite Automata (GHRFA) and proves an equality of language families,

$R \subseteq Q \times (\Sigma \cup \{\#\}) \times Q$7

Finally, the full GHBFA family is shown to be invariant under the dihedral symmetries of the hexagon: rotations $R \subseteq Q \times (\Sigma \cup \{\#\}) \times Q$8 and reflections $R \subseteq Q \times (\Sigma \cup \{\#\}) \times Q$9 preserve the recognized language family. In effect, the 12 scanning modes fall into symmetry-related classes rather than constituting 12 unrelated models (D et al., 11 Aug 2025).

6. Adjacent formal models and boustrophedon-like dynamics

Several papers study models that are not standard BFAs but are closely aligned with boustrophedon computation.

One example is the reflexive-composition formalism for one-dimensional lattices of identical finite-state machines. That paper does not use the term “Boustrophedon Automata,” yet its update law is explicitly described as an ordered sweep through a lattice with sender/receiver reversal and with directionality alternated at each step. The authors call the induced spatial symmetry “folding.” Because each site both stores state and transforms a transmitted message, the system is more than synchronous nearest-neighbor cellular automata. In this setting, the elementary machine M45 reproduces the Sierpinski triangle and is equivalent, on alternating steps and with suitable boundaries, to Rule 90, while M54 and M60 generate trajectories corresponding to Rule 90 in reverse. The paper presents this as the first study to produce such inverse dynamics naturally from the formalism itself (Salzberg et al., 12 May 2025).

A different neighboring line of work concerns restricted two-way motion in string automata. The 2014 paper on finite automata with restricted two-way motion does not formalize boustrophedon sweeps, bounded reversals, or alternating full passes. Instead, it measures only the total number of left moves in accepting computations and defines the class $\#$0 by the condition $\#$1. This is best understood as a quantitative backtracking restriction rather than a direct boustrophedon model: small $\#$2 allows only limited backing up, whereas the standard BFA discipline prescribes a specific alternating scan order (Damanik, 2014).

The 2010 paper on zigzags in Turing machines is a closer analogue for one-head rewriting machines. It studies bounded cycles, bounded zigzags, and $\#$3-cycles, all defined geometrically from head trajectories on a bi-infinite tape. Its main theorem states that a machine is bounded-zigzag if and only if the associated head-trace subshift $\#$4 is recognized by some deterministic pushdown automaton. The paper interprets the stack as exactly the memory needed for one large outward-and-return excursion whose crossed cells are revisited in LIFO order. This is a strong analogue of structured alternating scanning, but it is still not a classical endmarker-based sweeping automaton (Gajardo et al., 2010).

Taken together, these models show that boustrophedon behavior has at least two mathematically distinct aspects: a fixed alternating serialization of a two-dimensional input, and a broader family of restricted left-right sweep disciplines in which traversal order itself carries computational content.

7. The term beyond automata theory

Outside formal language theory, “boustrophedon” often refers only to a coverage discipline. The 2019 paper on boustrophedon coverage path planning for MAVs is representative: it revisits classical boustrophedon decomposition of polygonal free space into sweepable cells, then formulates global sweep selection and cell ordering as an Equality Generalized Traveling Salesman Problem. The closest automaton-like object in that work is a directed graph $\#$5 whose nodes are alternative sweep patterns and whose clusters correspond to cells, but the paper does not use automata terminology and does not define a finite-state recognition model (Bähnemann et al., 2019).

This distinction matters because it separates two uses of the same word. In coverage planning, boustrophedon denotes a lawnmower-style traversal strategy and the combinatorics of connecting such sweeps. In automata theory, it denotes a computational model in which alternating scan direction is part of the formal semantics of input processing.

The strongest current technical understanding of boustrophedon automata therefore lies in the picture-language setting. There, the model is now characterized in three complementary ways: as a finite automaton over an alternating rowwise serialization, as a special case of deterministic DAG automata on a path-shaped input-agnostic encoding, and, in the hexagonal setting, as a symmetry-structured family of alternating-line scanners over bordered hexagonal pictures. Beyond that core, the term remains a productive analogy for ordered sweeps, alternating orientation, and restricted back-and-forth computation, but not every boustrophedon system is formally a boustrophedon automaton.