Stack Theory in Automata and Combinatorics

• Stack theory is a mathematical framework that studies stack data structures, incorporating pushdown automata and combinatorial models to analyze operations like push, pop, and replacement.
• Lang’s algorithm compresses exponentially many stack configurations into a cubic-time dynamic programming solution, enabling efficient simulation and integration with gradient-based learning.
• The theory underpins combinatorial methods, including valid hook configurations and fertility formulas, to bridge stack-sorting processes with probability and free cumulant analysis.

Stack theory encompasses a range of mathematical frameworks and computational models that analyze the properties, manipulations, and applications of stack data structures, both in formal language theory and combinatorics. Central to this field are nondeterministic (and deterministic) pushdown automata, stack‐based recurrent neural architectures, and the algebraic and combinatorial structures underlying stack‐sorting processes and their connections to probability theory. Stack theory thus bridges automata theory, enumerative combinatorics, and algebraic probability, providing tools for both the theoretical analysis and efficient simulation of highly structured processes such as language recognition, stack‐sorting maps, and related combinatorial phenomena.

1. Nondeterministic Stack Data Structures and Pushdown Automata

In formal language theory, stack theory is rigorously formalized through the study of (nondeterministic) pushdown automata (PDA), described as $M = (Q, \Sigma, \Gamma, \delta, q_0, \bot)$ where $Q$ is the set of states, $\Sigma$ is the input alphabet, $\Gamma$ is the stack alphabet (with initial symbol $\bot$), and $$\delta: Q\times\Gamma\times\Sigma\times Q\times\Gamma^*\to\mathbb{R}_+$$ is a probabilistic transition function. Stack manipulation admits three atomic operations: push, pop, and replace (no-op). Each succession of operations over an input string yields an exponential number of possible stack configurations and machine states as a function of input length. Classical simulation is thus intractable; however, Lang’s algorithm enables simultaneous encoding and manipulation of this exponential set of stacks at each time step using a weighted finite automaton (WFA) whose states are triples $(i, q, x)$, representing having read position $i$ in state $q$ with top-of-stack symbol $x$ (DuSell et al., 2020).

A fundamental insight of stack theory is that the combinatorial explosion of configurations can be compressed to polynomial resource requirements (specifically, $Q$0 transitions) using dynamic programming recurrences. This representation underpins powerful simulation and learning frameworks, including differentiable implementations that are amenable to gradient-based optimization.

2. Lang’s Algorithm and Differentiable Stack Simulation

Lang’s algorithm provides a grammar-like, cubic-time simulation for nondeterministic PDAs. The algorithm constructs, for each input position $Q$1, a WFA $Q$2 representing all reachable stack configurations, via three inference rules reflecting push, replace, and pop operations. The dynamic programming variables

• $Q$3: parameterizes probabilistic transitions,
• $Q$4: accumulates transition weights for one move,
• $Q$5: forward weights aggregating probabilities for being in state $Q$6 at position $Q$7,

are updated using recurrences derived from Lang’s rules, efficiently capturing exponentially many latent computations per time step.

All stack operations and recursion are differentiable, permitting integration with trainable controllers, such as RNNs. When the transition tensor $Q$8 is controlled by a neural network and optimized using back-propagation, the resulting process—Nondeterministic Stack RNN (NS-RNN)—models distributions over all potential stack traces, supporting learning of context-free languages and robust trainability even on nondeterministically structured tasks (DuSell et al., 2020).

3. Stack-Sorting Maps, Troupes, and Cumulants

Stack-sorting theory studies the combinatorics of the stack-sorting map $Q$9, which acts on permutations and is central in the theory of sorting with stacks. Defant introduced the concept of “troupes,” which are insertion- and decomposition-closed families of colored binary plane trees satisfying a black-peaked property (all degree-2 vertices are black) (Defant, 2020). These combinatorial models provide an abstract, algebraic structure for analyzing stack-sorting operations.

A key result of stack theory is the connection between troupes and cumulants—specifically, between sequences that count tree families (free cumulants) and their “decreasing” counterparts (classical cumulants). Two novel explicit conversion formulas are established:

1. As a sum over noncrossing partitions (Corollary 15).
2. As a sum over $\Sigma$0-avoiding valid hook configurations (Theorem 25).

These formulas facilitate combinatorial and probabilistic translation between stack-sorting enumeration and free probability theory.

4. Valid Hook Configurations and Fertility Formulas

Valid hook configurations (VHCs) arise as a universal structure in the enumeration of preimages of postorder permutations under stack-sorting. Each VHC corresponds to a canonical decomposition of decreasing trees (troupe elements) and can be related bijectively to acyclic orientations on associated graphs. The tree decomposition lemma, the tree fertility formula, and their generalizations for troupes enable explicit calculations for permutation fertility—the number of preimages under the stack-sorting map.

These results subsume and strengthen prior techniques, providing streamlined, conceptual proofs and generalized enumerative machinery for fertile, uniquely sortable, or pattern-constrained permutations (Defant, 2020).

5. Enumeration, Asymptotics, and Structural Properties

Stack theory provides precise enumerative and asymptotic results for families sorted or sortable by stacks:

• The generating function for 2-stack-sortable alternating permutations of odd length is algebraic, satisfying degree-5 equations, with exponential growth rate $\Sigma$1 and explicit subexponential factors.
• For permutations whose descents are all peaks, analogous formulas and rates ($\Sigma$2) are established.
• The expected number of descents after stack-sorting is $\Sigma$3, with asymptotic variance $\Sigma$4.
• The number of sorted permutations $\Sigma$5 grows at exponential rate, bounded below by $\Sigma$6 and above by $\Sigma$7 (as $\Sigma$8).
• The degree of noninvertibility of $\Sigma$9 satisfies $\Gamma$0 (Defant, 2020).

These results rely essentially on VHC-based methods, fertility formulas, and the translation of combinatorial structures into analytic enumerations.

6. Applications in Learning and Expressivity

Differentiable versions of stack-based models (specifically, NS-RNNs) provide practical algorithms for learning probabilistic context-free languages by embedding the exponential space of PDA configurations in compact, tractable automata. Empirically, the NS-RNN architecture achieves:

• Faster and more reliable convergence to optimal solutions on deterministic CFLs,
• Lower cross-entropy on inherently nondeterministic context-free language tasks,
• Enhanced trainability and greater expressivity due to global, non-greedy gradient paths (DuSell et al., 2020).

These advantages derive from representing all possible stack traces jointly, rather than selecting a single latent run, leading to improved optimization landscapes.

7. Open Problems and Future Directions

Several open directions remain in stack theory:

• Algebraicity criteria for generating series of troupes, and explicit dependence on branch-generator series,
• Real-rootedness and unimodality conjectures for descent polynomials associated with classical troupes,
• Distributional convergence of descent statistics under stack-sorting, including central limit behaviors,
• Closed-form enumeration of alternating permutations in certain preimage classes,
• Structural analysis and combinatorial interpretation of arc-poset linear extensions related to noncrossing partitions,
• Generalizations of the fertility formula and decomposition techniques to broader classes of pattern-avoiding and uniquely-sorted permutations.

Continued investigation into these problems will further unify stack theory, automata, combinatorics, and free probability, fostering new methodologies for analysis and simulation of stack-driven processes (Defant, 2020).