Hankel Matrices for Weighted Visibly Pushdown Automata

Hankel matrices (aka connection matrices) of word functions and graph parameters have wide applications in automata theory, graph theory, and machine learning. We give a characterization of real-valued functions on nested words recognized by weighted visibly pushdown automata in terms of Hankel matrices on nested words. This complements C. Mathissen’s characterization in terms of weighted monadic second order logic.

💡 Research Summary

**
The paper establishes a precise algebraic characterization of the class of real‑ or complex‑valued functions on well‑nested words that are recognizable by Weighted Visibly Pushdown Automata (WVPAs). Classical results such as the Carlyle‑Paz theorem link finite‑rank Hankel matrices to weighted word automata, but these results do not directly apply to nested words because the usual Hankel matrix over the tagged alphabet would assign infinite rank to functions like the Dyck language characteristic, even though such functions are recognized by visibly pushdown automata.

To overcome this mismatch, the authors introduce the notion of a nested Hankel matrix (nH_f). Its rows and columns are indexed only by words over the tagged alphabet (\widehat{\Sigma}) that have well‑matched parentheses (i.e., encode well‑nested words). The entry at ((u,v)) is defined as (f(uv)), where (f) is the target function on well‑nested words (extended to the tagged representation). This restriction guarantees that functions recognizable by a WVPA can have a finite‑rank nested Hankel matrix.

The main contribution is Theorem 2 (the “Main Theorem”):
If a function (f) on well‑nested words is realized by a WVPA with (n) states, then the rank of its nested Hankel matrix satisfies (\operatorname{rank}(nH_f) \le n^2).
Conversely, if (\operatorname{rank}(nH_f) \le n^2) (over (\mathbb{R}) or (\mathbb{C})), then there exists a WVPA with at most (n) states that computes (f).

The forward direction is proved by constructing, for each ordered pair of states ((i,j)), an infinite row vector (v(i,j)) whose entry for a well‑nested word (w) is (\alpha^{\top} A(i,j) M_A(w) \eta). Here (A(i,j)) is the elementary matrix with a single 1 at position ((i,j)), (\alpha,\eta) are the initial and final weight vectors, and (M_A(w)) is the product of transition matrices defined by the WVPA. The set of (n^2) vectors ({v(i,j)}) spans all rows of (nH_f), establishing the rank bound.

The reverse direction relies on the Singular Value Decomposition (SVD), which is only guaranteed over fields such as (\mathbb{R}) or (\mathbb{C}). Assuming (\operatorname{rank}(nH_f)=r\le n^2), the authors apply SVD to obtain (nH_f = U\Sigma V^{\top}) with (r) non‑zero singular values. The left and right singular vectors are used to define a state space of dimension (r). Transition matrices for call, return, and internal symbols are then constructed from inner products of these vectors, ensuring that the resulting WVPA reproduces the original function exactly. The proof highlights that the SVD step is essential; without an analogue in general semirings, the theorem cannot be extended directly.

Beyond the core theoretical result, the paper discusses several applications and extensions:

1. Learning Theory – Existing spectral learning algorithms for weighted automata (e.g., Hsu et al., Balle & Mohri) exploit the connection between Hankel matrices and automata. The authors argue that the nested Hankel matrix provides a similar foundation for learning WVPAs. Lemma 11 sketches how to extract the matrices corresponding to call and return symbols from empirical data, suggesting a pathway to active or passive learning algorithms for visibly pushdown structures.

2. Semiring Generalization – Since SVD depends on additive inverses and inner products, the main theorem is limited to (\mathbb{R}) and (\mathbb{C}). The authors note that for the tropical semiring, an analogue of SVD exists via the “symmetrized max‑algebra” (De Schutter & De Moor). This hints that a finite‑rank characterization might be possible for WVPAs over tropical weights, which would be valuable for optimization‑oriented applications where max‑plus algebra is natural.

3. Relation to Logical Characterizations – Mathissen’s weighted monadic second‑order logic (WMSO) provides a logical description of WVPA‑recognizable functions. The Hankel‑based characterization complements this logical view, offering an algebraic criterion that can be checked via linear algebra rather than logical formulas.

4. Future Directions – The paper suggests extending the framework to more complex hierarchical structures (e.g., trees, graphs) by defining appropriate “nested” Hankel matrices, and investigating whether similar rank‑based characterizations hold.

In summary, the authors successfully bridge the gap between weighted visibly pushdown automata and linear‑algebraic tools by defining a tailored Hankel matrix for well‑nested words. The resulting rank condition provides both a necessary and sufficient criterion for WVPA recognizability (over (\mathbb{R}) or (\mathbb{C})), opens the door to spectral learning methods for pushdown‑type models, and points toward intriguing generalizations in semiring settings and richer hierarchical data structures.