Multiplicity Automata Equivalence

• Multiplicity automata equivalence is the problem of determining whether two weighted finite automata, defined over a field, yield identical output functions.
• The complexity classification shows that the problem is C_=L-complete, leveraging reductions to matrix rank computation and polynomial identity testing.
• Algorithmic solutions include linear-algebraic methods and randomized NC approaches, with extensions to probabilistic, pushdown, and multitape automata models.

A multiplicity automaton (also known as a weighted automaton over a field) is a finite automaton where transitions carry weights from an underlying field (typically $\mathbb{Q}$), and each word is assigned a weight computed as a weighted sum over runs. The equivalence problem for multiplicity automata asks whether two such automata define the same function from words to field elements. This problem is central in automata theory, formal power series, and links deeply to polynomial identity testing (PIT) and counting complexity. Recent advances have precisely characterized the complexity of multiplicity automata equivalence and developed efficient algorithms in various settings.

1. Formal Definitions and Structural Properties

A multiplicity (word) automaton over a field $\mathbb{K}$ and alphabet $\Sigma$ is a tuple $$A = (Q, \Sigma, \alpha, \beta, \{M_a\}_{a\in\Sigma})$$, where $Q$ is a finite set of states, $\alpha \in \mathbb{K}^Q$ is the initial (row) vector, $\beta \in \mathbb{K}^Q$ is the final (column) vector, and for each $a \in \Sigma$, $M_a \in \mathbb{K}^{Q \times Q}$ is a transition matrix. The weight assigned to a word $w = a_1 a_2 \ldots a_k$ is computed as $$\sigma_A(w) = \alpha M_{a_1} M_{a_2} \ldots M_{a_k} \beta$$.

Two automata $A$ and $B$ over the same $\Sigma$ and $\mathbb{K}$ are equivalent if $\sigma_A(w) = \sigma_B(w)$ for all $w \in \Sigma^*$. Equivalence is captured more succinctly by considering the difference automaton $A \ominus B$, whose states are the disjoint union of the states of $A$ and $B$, initial vector $\alpha_{A\ominus B} = (\alpha_A, -\alpha_B)$, final vector $\beta_{A\ominus B} = (\beta_A; \beta_B)$, and transition matrices as block-diagonals of those of $A$ and $B$. Equivalence $A \equiv B$ holds iff $A \ominus B$ is the zero automaton, i.e., it produces weight 0 on every $w$ (Kiefer, 2020).

The behavior of a multiplicity automaton defines a rational formal power series. The rank of the associated Hankel matrix determines the minimal state-space size among equivalent automata, and minimality testing reduces to rank computations (Kiefer, 2020).

2. Complexity Classification

The complexity of the multiplicity automata equivalence problem has been precisely pinned down. For word automata over $\mathbb{Q}$, deciding equivalence is complete for the class $\mathsf{C}_{=}L$ under logspace many-one reductions (Černý et al., 15 Dec 2025).

• $\mathsf{C}_{=}L$ consists of all languages $\{x \mid f(x)=0\}$ for $f$ in GapL; equivalently, it is the class of problems logspace many-one reducible to determining whether the determinant of an integer matrix is zero.
• The $\mathsf{C}_{=}L$-completeness means both that the problem is solvable in deterministic logarithmic space with access to a GapL oracle and that every $\mathsf{C}_{=}L$ problem can be reduced to multiplicity automaton equivalence.
• This classification is tight: upper bounds leverage reduction to matrix rank and sum-of-powers checks; lower bounds are inherited from trace and characteristic polynomial verification, which are canonical complete problems for $\mathsf{C}_{=}L$ (Černý et al., 15 Dec 2025).

Improvements beyond $\mathsf{C}_{=}L$ are considered unlikely, given established logspace interreducibility with classical counting algebraic problems and PIT (Černý et al., 15 Dec 2025).

3. Algorithmic Methods: Linear Algebra, Polynomial Identity Testing, and Isolation

Several complementary algorithmic approaches have been developed for the equivalence problem.

Linear-Algebraic Method

• Construct the difference automaton $A \ominus B$.
• Build the forward space $$F = \operatorname{span}\{\alpha M(w)\ |\ w \in \Sigma^*\}$$ and test whether all $v\in F$ are orthogonal to the final vector $\beta$.
• Iteratively extend a basis for $F$ by breadth-first exploration over $\Sigma$, adding $v\cdot M_a$ only if it is linearly independent.
• If any basis vector $v$ yields $v\cdot\beta \ne 0$, the label $w$ is a counterexample; otherwise, equivalence holds.
• Complexity is $O(|\Sigma| N^3)$ field operations for automata of combined size $N$ (Kiefer, 2020).

Polynomial Identity Testing (PIT) and Randomized NC

• Encode the set of words of length less than $n$ in a multivariate polynomial $P(x)$, whose monomials correspond to weighted computations of the automaton.
• Use fast parallel algorithms for matrix products and the Schwartz–Zippel lemma to probabilistically check whether $P$ is the zero polynomial.
• Isolation Lemma methods allow witness extraction: assign random weights to positions and letters, construct a univariate polynomial in $x$, and with high probability, the lowest-degree term corresponds to a unique counterexample word.
• These techniques give rise to randomized NC ($\mathsf{RNC}^2$) and deterministic NC procedures for word automata (Kiefer et al., 2013, Kiefer et al., 2011).

Complexity for Broader Models

• For partially commutative monoids, the problem's complexity depends on the clique-cover number $\kappa$ of the non-commutation graph:
  • Deterministic quasi-polynomial time for constant $\kappa$
  • Randomized polynomial time if the commutation structure is covered by a bounded number of cliques or stars (Arvind et al., 2020)
• For multitape automata, randomized polytime algorithms exist for fixed $k$, and the best deterministic complexity is quasi-polynomial for constant $k$ (Worrell, 2013, Arvind et al., 2020).

4. Extensions: Probabilistic Automata, Rewards, and Pushdown Models

Multiplicity automaton equivalence generalizes to numerous models:

• Probabilistic automata are a special case with nonnegative weights and stochastic transition matrices. Equivalence of probabilistic automata reduces directly to multiplicity automaton equivalence (Kiefer et al., 2011).
• When transitions carry cost/reward vectors, two equivalence notions arise: expectation-equivalence and distribution-equivalence. Both problems reduce to the basic $\mathbb{Q}$-weighted (multiplicity) case, preserving algorithmic tractability (randomized polynomial time for fixed reward dimension; deterministic for fixed number of counters) (Kiefer et al., 2013, Kiefer et al., 2011).
• For visibly pushdown automata (VPA), the equivalence problem is logspace-equivalent to the arithmetic circuit identity testing (ACIT) problem. Thus, efficient algorithms for VPA equivalence would imply breakthroughs in ACIT derandomization (Kiefer et al., 2013, Kiefer et al., 2011).

A representative table summarizes complexity status:

Model/Class	Deterministic Complexity	Randomized Complexity	Reference
Word automata ($\mathbb{Q}$)	$\mathsf{C}_{=}L$-complete	$\mathsf{RNC}^2$	(Černý et al., 15 Dec 2025, Kiefer et al., 2013)
$k$-tape automata (constant $k$)	Quasi-poly time ($k$ cliques: $(ns)^{O(k^2 \log ns)}$)	Polynomial time	(Arvind et al., 2020, Worrell, 2013)
Probabilistic automata w/ rewards	Polytime for fixed counters	Polytime	(Kiefer et al., 2011, Kiefer et al., 2013)
Partially commutative monoids	Quasi-poly for constant $\kappa$	Polytime if clique/star cover bounded	(Arvind et al., 2020)
Visibly pushdown automata (VPA)	? (open; in coRP)	Logspace-equivalent to ACIT	(Kiefer et al., 2013, Kiefer et al., 2011)

5. Connections to Homomorphism Indistinguishability and PIT

Recent results reveal a deep relationship between the equivalence problem and homomorphism indistinguishability for graphs: For bounded pathwidth graph classes definable in counting MSO$_1$, the problem of distinguishing graphs by homomorphism counts reduces in logspace to multiplicity automaton equivalence (Černý et al., 15 Dec 2025). This positions automata equivalence as a canonical $\mathsf{C}_{=}L$-complete problem with implications for graph isomorphism relaxations, cospectrality, and quantum isomorphism.

The tight connection to PIT is pivotal: For bounded-treewidth graph classes, homomorphism indistinguishability (and, by reduction, automata equivalence) is logspace interreducible with PIT. This connection sets a barrier for deterministic polynomial-time algorithms for automata equivalence in general settings, conditioning further progress on derandomizing PIT (Černý et al., 15 Dec 2025).

6. Open Problems and Research Directions

Prominent open problems include:

• Derandomizing algorithms for equivalence over multitape and partially commutative models with moderate commutation structure (i.e., closing the gap between randomized polytime and deterministic quasi-polytime for bounded clique-cover) (Arvind et al., 2020).
• Determining the precise complexity for automata with unbounded commutation (large $\kappa$) or with tree automata (where only randomized algorithms are currently efficient) (Černý et al., 15 Dec 2025, Arvind et al., 2020).
• Improving PIT algorithms or establishing lower bounds for automata equivalence in pushdown, tree, or infinite word models, which are currently linked to deep algebraic circuit properties (Kiefer et al., 2011, Kiefer et al., 2013, Černý et al., 15 Dec 2025).
• Exploring further connections to algebraic and combinatorial invariants (e.g., graph homomorphism counts, characteristic polynomials) and their computational boundaries.

These directions highlight the centrality of multiplicity automata equivalence in complexity theory, formal languages, and algebraic computation.