Multihead Finite-State Dimension

• Multihead finite-state dimension extends classical automata by using multiple oblivious read heads to refine measures of information density and compression.
• The framework establishes a strict hierarchy where adding heads enhances predictive power, demonstrated through explicit separation results.
• It reveals a duality between gamblers and compressors, linking effective randomness, symbolic dynamics, and resource-bounded inference.

Multihead finite-state dimension is a generalization of the classical finite-state dimension, extending the capabilities of finite automata by equipping them with multiple obliviously moving read heads that operate in parallel on an infinite data stream. This framework refines our understanding of information density, algorithmic randomness, and resource-bounded compressibility by measuring how much additional predictive or compressive power multiple coordinated agents can provide when restricted to finite memory and strictly prescribed movement rules. The theory introduces a strict hierarchy among devices with differing numbers of heads and provides explicit separation results, stability properties, and links to compression. Its development synthesizes concepts from classical automata theory, fractal dimensions, algorithmic information theory, and effective randomness.

1. Definition and Operational Model

A multihead finite-state gambler (h-FSG) is a tuple $$(T \times Q, \Sigma, \delta, \mu, \beta, (t_0, q_0))$$, where $Q$ and $T$ are finite sets that encode computational and positional states, $\Sigma$ is a finite alphabet, $\delta$ is a transition function receiving current state and the $h$-tuple of input symbols (one from each head), $\mu$ governs the position updates for the trailing $h-1$ heads (according to an oblivious, data-independent rule), and $\beta: Q \to \Delta_\mathbb{Q}(\Sigma)$ outputs the bets (probabilities) that will be placed on the symbol to be read next by the leading head. The leading head always advances one symbol at each time step, while the trailing heads follow their own pre-determined schedules. The entire configuration thus encodes the state of the automaton and the vector of head positions in the input stream.

Given a sequence $S \in \Sigma^\omega$, the $h$-head finite-state dimension, referred to as the $h$-head finite-state predimension ${}^{(h)}(S)$, is defined as follows: for $s \geq 0$, the $s$-gale of $G$ is $d_G^{(s)}(w) = |\Sigma|^{(s-1)|w|} d_G(w)$, and $s$ is said to be achievable if there exists an $h$-FSG whose $s$-gale succeeds on $S$ (capital diverges along $S$). Then,

$${}^{(h)}(S) = \inf\{s \geq 0 : \exists\, h\text{-FSG }G \text{ s.t. } \limsup_n d_G^{(s)}(S \upharpoonright n) = \infty \}.$$

The multihead finite-state dimension $\mathrm{MH}(S)$ is the infimum over $h$:

$$\mathrm{MH}(S) = \inf_{h \in \mathbb{Z}^+} {}^{(h)}(S).$$

For $h=1$ this recovers the classical finite-state dimension as in Dai, Lathrop, Lutz, and Mayordomo.

2. Hierarchy Theorem and Separation Results

The principal structural result is a strict hierarchy: for every $h \geq 1$, there exists an explicit sequence $Y$ such that

${}^{(h+1)}(Y) < {}^{(h)}(Y),$

demonstrating that adding an extra head strictly enhances the predictive/compressive power. The construction employs a function $F_{h+1}$ that, for each positive integer $h$, intertwines the bits of the original sequence at locations indexed by the first $h+1$ primes: at positions that are multiples of $p_{h+1}$, the bit is defined as the parity of bits at corresponding positions defined by the previous $h$ primes, while at all other locations, bits are copied directly. A key insight is that an $h+1$-FSG can align its heads to access all relevant bits needed to compute the parity and thus win efficiently, while an $h$-FSG lacks enough heads to access all required positions and cannot improve its capital growth rate. Thus, for each $h$ there is a family of $F_{h+1}(S)$, constructed from a Martin-Löf random $S$, that witnesses the separation.

This establishes that the sequence of predimensions $\{{}^{(h)}(S)\}_{h \geq 1}$ is strictly decreasing for appropriate $S$, encoding a genuine complexity hierarchy sensitive to the head count.

3. Multihead Finite-State Gales and Betting Strategies

Gales generalize finite-state martingales by incorporating a real exponent $s \geq 0$. An $s$-gale is a function $d:\Sigma^* \to [0,\infty)$ satisfying

$d(w) = |\Sigma|^{-s} \sum_{b \in \Sigma} d(wb).$

For an $h$-FSG, the capital update leverages the augmented input vector $(x_{i_1},...,x_{i_h})$, where each $i_k$ encodes the position of the $k$th head in $T$. At each step, the gambler distributes its capital according to the bet vector $\beta(q)$, conditioned on the state transitions and observed symbols. Critical technicalities involve coordinating head movement so that the leading head always reads the next symbol (for one-way models) and the trailing heads follow a pre-determined "oblivious" path described by the movement function $\mu$.

Success of a gale is measured by the divergence of capital along the sequence. The minimal $s$ for which success is achievable determines the predimension; in essence, the more heads available, the more the gambler can "coordinate" to exploit correlations or long-range dependencies in the sequence.

4. Compression Duality and Exact Characterization

A core advancement is the precise connection between multihead finite-state dimension and compression. For each $h$, the infimum compression ratio achievable by $h$-head finite-state information-lossless compressors (those whose output, together with final state, enables unique recovery of the input) equals the $h$-head finite-state predimension of the sequence. Formally, the main theorem in (Lutz, 20 Oct 2025) asserts:

${}^{(h)}(S) = {}^{(h)}(S),$

where the left side is the best achievable asymptotic compression ratio and the right side the predimension. The overall multihead finite-state dimension is the infimum over all $h$, matching exactly the infimum over all multihead compressed representations. These results are proved by constructing explicit reductions between gamblers and compressors: any information-lossless multihead compressor can be simulated by a gale-based investor, and reciprocally, any winning multihead gale can be converted into a compressor that compresses at the corresponding rate.

This duality serves as a multihead analogue of the classic equivalence between finite-state dimension and optimal compression ratio, and directly links dimensions to actual feasible compression schemes in this resource-limited model.

5. Stability and Structural Properties

Multihead finite-state dimension enjoys robust stability properties. For any finite collection of sets $X_1, ..., X_k$,

$$\mathrm{MH}\left(\bigcup_i X_i\right) = \max_i \mathrm{MH}(X_i),$$

i.e., it is stable under finite unions. In contrast, for every fixed $h>1$, the $h$-head predimension ${}^{(h)}(S)$ lacks this property: there exist sets $A,B$ with ${}^{(h)}(A), {}^{(h)}(B) < 1$ but ${}^{(h)}(A \cup B) = 1$. The non-stability of fixed-$h$ predimension arises from the limited ability of a gambler with a fixed number of heads to "coordinate" different strategies that succeed separately on $A$ and on $B$.

This robustness at the MH level reinforces its candidacy as an effective fractal dimension suited for highly resource-bounded and compositional settings, and mirrors the behavior of classical Hausdorff dimension and other effective dimensions (Huang et al., 26 Sep 2025).

6. Connections with Classical, Multi-bet, and Relative Dimensions

Multihead finite-state dimension extends the classical (single-head) finite-state dimension, providing a strictly finer-grained hierarchy: only when $h \to \infty$ do all possible "coordination" or nonlocal correlations become accessible to the device. There are close connections to the multi-bet finite-state dimension concept studied in (S, 10 Feb 2025), where $k$-bet finite-state gamblers place $k$ separate bets per symbol. However, (S, 10 Feb 2025) establishes that allowing multiple independent "accounts" (multi-bet) does not increase finite-state dimension beyond the classical setting, while multihead models, via the coordinated access to spatially-separated symbols, do produce a strict hierarchy.

Relative and conditional dimensions (Nandakumar et al., 2023, Shen, 2024) provide resource-bounded analogues of mutual information where oracles or side information are available. Multihead dimension, in turn, can be interpreted as a nonadaptive, distributed version of relative dimension, where the extra heads serve as "forgetful oracles" limited by oblivious movement. Furthermore, recent Markov-chain and Weyl-criterion characterizations of finite-state dimension suggest the possibility of analogous divergence-based or Fourier-theoretic invariants for multihead dimensions (Bienvenu et al., 21 Oct 2025, Lutz et al., 2021).

7. Applications and Theoretical Implications

Multihead finite-state dimension offers a nuanced quantification of information rate, compressibility, and effective randomness under highly restrictive resource constraints. Hierarchical separation shows that augmenting resource pools, even minimally, can improve prediction and compression capabilities. Applications include:

• Data compression: Optimizing multihead finite-state compressors for embedding in lightweight or parallelizable data transmission protocols, especially where low memory and simple device architecture are required.
• Algorithmic randomness and effective fractal geometry: Refining classification of sequences and reals in terms of discoverable structure and resistance to prediction or compression by bounded agents, informing new classes of normal numbers and effective analogues of Hausdorff dimension.
• Symbolic dynamics and multihead automata theory: Fine-structure analysis of symbolic systems, especially in settings with multidimensional or nonlocal dependencies (cf. plane-walking automata (Salo et al., 2014)).
• Information theory and resource-bounded inference: Sharp characterization of the limits of learnability, prediction, and extraction in models constrained by locality and memory.

Taken together, multihead finite-state dimension bridges automata theory, randomness, and information theory, providing a comprehensive and operationally meaningful measure of effective information density in sequential and symbolic data streams.