Witness Automaton Prediction

• Witness automaton prediction is a framework that investigates which classes of infinite ω-words can be mastered by different automata, including DFA, DPDA, DSA, and multihead DFA.
• The research establishes separation and characterization theorems with explicit algorithms such as DSA for purely periodic, 2-head DFA for ultimately periodic, and sensing 10-head DFA for multilinear sequences.
• It demonstrates a clear hierarchy where enhanced automata resources, like additional heads or stack access, overcome the limitations of simpler models to achieve perfect prediction.

Witness automaton prediction investigates which classes of infinite sequences ("ω-words") over a finite alphabet can be mastered by automaton-based predictors, where mastery entails that after a finite learning phase, all subsequent predictions are correct. The central problem is to characterize the predictive power of various automata types—including deterministic finite automata (DFA), deterministic pushdown automata (DPDA), stack automata (DSA), and multihead finite automata—against increasingly complex classes of ω-words, such as purely periodic, ultimately periodic, and multilinear sequences. The main results consist of separation and characterization theorems, constructive algorithms, and complexity bounds, establishing a clean hierarchy of predictive capability that depends on automaton resources and structural properties of the target sequence (Smith, 2016).

1. Formal Definitions and Classes of Automata Predictors

The prediction task for infinite words is formalized as follows. Let $\Sigma$ denote a finite alphabet. An infinite word $\alpha \colon \mathbb{N} \to \Sigma$ can be written $\alpha = \alpha[1]\alpha[2]\alpha[3] \ldots$. A predictor $M$ reads the prefix $\alpha[1 \ldots i-1]$ and produces a guess $g_i \in \Sigma$ for the next symbol $\alpha[i]$, yielding the process $M(\alpha) = g_1g_2g_3\ldots \in \Sigma^\omega$, with each guess $g_i$ made before reading $\alpha[i]$. The predictor masters $\alpha$ if there exists $N$ such that for all $i \geq N$, $g_i = \alpha[i]$.

Four deterministic automaton-based predictors are investigated:

1. DFA Predictor: A tuple $(Q,\Sigma,\delta,q_0)$ with $$\delta: Q \times (\Sigma \cup \{\triangleleft\}) \to Q \times \Sigma$$; at each step the automaton transitions, emits a guess, and moves right.
2. DPDA Predictor: $(Q,\Sigma,\Gamma,\delta,q_0,\bot)$ tracks stack configurations; transitions have the form $\delta(q,c,A) = (q',act,g)$, incorporating stack operations and output.
3. DSA Predictor: An augmentation of DPDA with an additional read-only stack head, allowing traversals of the stack without modification; transitions depend on input, stack symbol under the stack-head, and its position.
4. Multihead DFA Predictor (with Sensing): $(Q,\Sigma,k,\delta,q_0)$ with $k$ input heads; transitions function as $\delta(q,c_1,\ldots,c_k) = (q',d_1,\ldots,d_k,g)$, where $d_j$ denotes head movements. Sensing enables head-coincidence information.

2. Main Theorems: Capabilities and Separations

The theory identifies several key ω-word classes:

• $P$: Purely periodic ($x^\omega$)
• $U$: Ultimately periodic ($xy^\omega$)
• $M$: Multilinear words, specified as $$q \cdot \prod_{n \geq 0} r_1^{a_1 n + b_1} \cdots r_m^{a_m n + b_m}$$ with constraints on $r_i, a_i, b_i$.

Principal results include:

• DFA and DPDA Insufficiency: No DFA or DPDA predictor can master all purely periodic words when $|\Sigma| \geq 2$ (Theorems 2.1–2.2). Proofs apply pumping-style arguments: limited states or stack capacity guarantees repeated incorrect guesses in inputs such as $(a^{p+1}b)^\omega$.
• DSA Sufficiency and Characterization: DSAs master precisely the purely periodic words (Theorem 2.3). Any word mastered by a DSA predictor is multilinear, with strict inclusion $P \subseteq$ Mastery(DSA) $\subseteq M$.
• Multihead DFA for Ultimately Periodic: A 2-head DFA (using the Floyd cycle-finding principle) can master all ultimately periodic words (Theorem 2.4). No single-head DFA suffices, establishing the lower bound of $k=2$ (Proposition 2.5).
• Sensing Multihead DFA for Multilinear: A sensing multi-DFA with 10 heads masters all multilinear words, with no constructions known for $\leq 9$ heads (Theorem 2.6).

A visual summary:

DFA ⊂ DPDA ⊂ DSA × × ✓ (P) (U) (Multilinear)

Also, multi-DFA (2 heads) masters $U$ but not $M$; sensing multi-DFA (10 heads) masters $M$.

3. Concrete Witness Algorithms and Sample Runs

For each ω-word class admitting mastery, explicit witness algorithms are provided:

• DSA Predictor for Purely Periodic Words (Algorithm A): The stack accumulates a candidate period; verification proceeds by matching the stack buffer against the input, with mismatches triggering period extension. Once the period in the stack matches the underlying periodicity, subsequent "verify" phases guarantee correct predictions.
• 2-head DFA Predictor for Ultimately Periodic Words (Algorithm B): The two heads ("tortoise" and "hare") move, copying symbols from "tortoise" to predict the input for "hare". Mismatches cause the hare to skip extra symbols; after finite errors, the heads become offset by the period length, enabling perfect copying.
• Sensing 10-head DFA Predictor for Multilinear Words (Algorithm C): Heads are aligned and corrected to segment boundaries using incremental spacing routines; matching exploits sensing to synchronize segment predictions according to the multilinear block structure. Sample runs (e.g., for $\alpha = \prod_{n \geq 1} a^n b^n$) illustrate alignment of heads at the start of blocks and subsequent prediction via previously learned patterns.

4. Computational Complexity of Prediction Algorithms

Complexity attributes of these algorithms are as follows:

• Per-symbol time
  • DFA/2-head DFA: $O(1)$ per input symbol (constant number of head moves and transitions).
  • DPDA/DSA: Actual advances cost $O(1)$ time; DSA "verify" passes cost up to $O(p)$ for candidate period length $p$, but occur only upon period extension, yielding amortized $O(1)$ per symbol.
  • Sensing 10-head DFA: Each loop increments by $O(1)$, with correction and matching taking bounded $O(r-l)$ steps per iteration. Each input cell is visited a constant number of times per head, maintaining overall $O(1)$ time per symbol.
• Space usage
  • DFA/2-head DFA: $O(1)$ finite-state memory, $O(1)$ number of heads.
  • DPDA: Unbounded stack height proportional to learned period, i.e., $O(n)$ worst-case.
  • DSA: Unbounded stack plus read-only head movements.
  • Sensing Multi-DFA: $O(1)$ finite control, 10 heads, $O(1)$ bits for sensing.

5. Proof Sketches and Hierarchical Separations

• Memory-Limited Impossibility: For DFA, the state-space pigeonhole principle dictates repeated configurations within one period, leading to predictive failure. DPDA encounters repeating configurations when stack height increases, yielding similar errors.
• DSA Sufficiency: The DSA leverages its read-only stack-head to test candidate periods of increasing length, pushing new symbols on mismatch, and ensures that once the stack buffer matches the true period, all predictions remain accurate.
• Cycle Detection and Multihead: The 2-head DFA simulates tortoise-hare cycle detection: mismatches increment the hare, aligning the offset with the period and enabling accurate copying. 10-head sensing multi-DFA employs combinatorial argumentation to capture complex multilinear block structures; correction and matching algorithms reconfigure head alignments to accommodate polynomial segment growth, guaranteeing eventual stable perfect prediction.

A plausible implication is that the predictive power of automaton-based learning strictly scales with augmentation: stack-head access enables mastery of periodicity, additional heads (with sensing) extend prediction up to multilinear sequences.

6. Synthesis: Automaton Predictor Hierarchy

The automata-based sequence prediction model reveals a strict hierarchy:

• DFA/DPDA: Insufficient for mastering even the simplest classes of periodic ω-words.
• DSA: Exactly adequate for purely periodic sequences; any mastered word must be multilinear.
• 2-head DFA: Precisely matches the ultimately periodic class.
• Sensing 10-head DFA: Extends coverage to all multilinear sequences.

Each class is accompanied by explicit algorithms and rigorous guarantees that, after a finite learning period, prediction becomes indefinitely accurate, with time and space bounds reflecting the structural complexity of the sequence mastered (Smith, 2016).