HypPAu: Hyper Pattern Matching Tool

Updated 6 July 2026

HypPAu is a runtime verification tool for hyper pattern matching that computes concrete multi-trace witnesses for asynchronous hyperproperties.
It employs nondeterministic asynchronous finite automata to capture violations by extracting precise subword intervals from multiple traces.
Heuristics such as FJS-style skipping and projection-based pruning enhance its efficiency, despite the underlying NP-complete challenge.

Searching arXiv for "HypPAu" and closely related terms to ground the article in current papers. HypPAu most directly denotes a runtime verification tool for hyper pattern matching, introduced to compute concrete multi-trace witnesses of hyperproperties over sets of words or logs. In that usage, HypPAu operates with nondeterministic asynchronous finite automata (NAAs), returns tuples of subwords with start–end positions per trace, and targets properties such as noninterference, robustness under stuttering, and asynchronous request–response relations (Waga et al., 16 Jul 2025). The same label is also used in the supplied arXiv literature as a connection term in unrelated domains, including hyperparameter auto-tuning around HYPPO (Dumont et al., 2021) and Hybrid Adaptive Parallelism for Mixture-of-Experts inference in HAP (Lin et al., 26 Aug 2025). By contrast, the visually similar acronym HYPA names a different framework for path anomaly detection in temporal network data (LaRock et al., 2019).

1. Terminological scope and principal usages

Within the supplied literature, “HypPAu” is not a single unified formalism but a label attached to several distinct systems. The most formalized and implementation-specific usage is the runtime verification tool accompanying “Hyper pattern matching” (Waga et al., 16 Jul 2025).

Usage in the supplied literature	Meaning	Paper
HypPAu	Runtime verification tool for hyper pattern matching	(Waga et al., 16 Jul 2025)
HypPAu	Hyperparameter auto-tuning connection around HYPPO	(Dumont et al., 2021)
HypPAu	Hybrid Adaptive Parallelism for efficient MoE inference (HAP)	(Lin et al., 26 Aug 2025)
HypPAu	Informal connection to Hyper photometry in the supplied details	(Traficante et al., 2014)

A common misconception is to conflate HypPAu with HYPA because of orthographic similarity. HYPA, however, is a framework for unsupervised detection of anomalous temporal paths in graphs using higher-order De Bruijn graphs and a hypergeometric null model; it is not a hyper pattern matching system and does not compute multi-trace witness tuples in the sense used by HypPAu (LaRock et al., 2019).

Another potential source of confusion is HyPER. In the supplied literature, the HyPER paper explicitly states that the term “HypPAu” does not appear in the paper. The provided comparison only suggests that, if one informally uses “HypPAu” to mean “hypothesis path augmentation,” then HyPER’s expansion actions are conceptually related to that expand side, while HyPER itself is an expand-plus-reduce controller for LLM reasoning rather than a system officially named HypPAu (Qiu et al., 6 Feb 2026).

2. Formalization in hyper pattern matching

In its primary technical sense, HypPAu is built around hyper pattern matching over a set of words using nondeterministic asynchronous finite automata. The underlying motivation is that witnessing violations of hyperproperties requires evidence across multiple input words or across separated parts of the same word, which conventional single-trace pattern matching cannot express (Waga et al., 16 Jul 2025).

The formalism uses a finite alphabet $\Sigma$ , a finite set of directions or tapes $K = \{\chi_1,\chi_2,\ldots,\chi_k\}$ , and an NAA

$A = (\Sigma, K, S, I, F, \Delta),$

where $\Delta \subseteq S \times \Sigma \times K \times S$ is the asynchronous transition relation. Acceptance is defined through the underlying NFA over $\Sigma \times K$ : a tuple $(u_1,\ldots,u_k) \in (\Sigma^*)^k$ is accepted iff there exists an extended word in the underlying NFA whose projection onto each direction $\chi$ equals the corresponding component word (Waga et al., 16 Jul 2025).

This yields the central witness object:

$M(A, w) = \{ ((w_1, i^1, j^1), \ldots, (w_k, i^k, j^k)) \mid w_\chi \in w,\ 1 \le i^\chi \le j^\chi \le |w_\chi|,\ (w_1[i^1..j^1], \ldots, w_k[i^k..j^k]) \in L(A) \}.$

Each element of $M(A,w)$ is therefore a concrete multi-trace witness: a tuple of selected words together with intervals whose sliced subwords satisfy the asynchronous hyperpattern (Waga et al., 16 Jul 2025).

This formulation is significant because it localizes evidence. Rather than returning only a Boolean verdict or a large prefix set, HypPAu returns the exact subword intervals responsible for a violation or occurrence. In the supplied case studies, this supports direct witness extraction for noninterference, stuttering robustness, and request–response motifs in packet logs.

3. Algorithms, complexity, and search-space reduction

The baseline HypPAu procedure is a naive exhaustive algorithm that enumerates tuples of words and starting positions, maintains asynchronous configurations, extends reading positions direction by direction, and records a witness whenever an accepting state is reached. The initial combinatorial frontier is a priority queue over all choices of $k$ words with repetitions from the input set and all starting indices, and the algorithm simulates all asynchronous interleavings consistent with the NAA (Waga et al., 16 Jul 2025).

The resulting worst-case cost is

$K = \{\chi_1,\chi_2,\ldots,\chi_k\}$ 0

where $K = \{\chi_1,\chi_2,\ldots,\chi_k\}$ 1, $K = \{\chi_1,\chi_2,\ldots,\chi_k\}$ 2 is the number of words, $K = \{\chi_1,\chi_2,\ldots,\chi_k\}$ 3 is the number of automaton states, and $K = \{\chi_1,\chi_2,\ldots,\chi_k\}$ 4 (Waga et al., 16 Jul 2025). The associated nonemptiness problem—deciding whether $K = \{\chi_1,\chi_2,\ldots,\chi_k\}$ 5—is proved $K = \{\chi_1,\chi_2,\ldots,\chi_k\}$ 6-complete via a reduction from CNF SAT and a polynomial-time nondeterministic verification argument (Waga et al., 16 Jul 2025).

To make the search practical, HypPAu implements two correctness-preserving heuristic families. The first is FJS-style skipping, combining QS-style and KMP-style skip rules generalized to asynchronous multi-directional automata. The QS component uses shortest-match information and per-direction letter constraints to remove impossible start positions, while the KMP component uses partial-match states to rule out forbidden alignment windows after a partial run reaches a state $K = \{\chi_1,\chi_2,\ldots,\chi_k\}$ 7 (Waga et al., 16 Jul 2025).

The second family is projection-based pruning. For each direction $K = \{\chi_1,\chi_2,\ldots,\chi_k\}$ 8, the NAA is projected to a classical DFA $K = \{\chi_1,\chi_2,\ldots,\chi_k\}$ 9 by turning transitions on $A = (\Sigma, K, S, I, F, \Delta),$ 0 into symbol transitions and transitions on other directions into $A = (\Sigma, K, S, I, F, \Delta),$ 1-transitions before determinization. Classical per-direction matches are then computed, and the global candidate queue is restricted to starting positions that can participate in some projected match. HypPAu further includes a linear pass that marks relevant indices and masks irrelevant positions, reducing the Cartesian-product blowup before asynchronous matching begins (Waga et al., 16 Jul 2025).

These heuristics do not change the formal language accepted by the automaton. Their role is purely algorithmic: they prune candidate starts that cannot be extended to any true witness.

4. Implementation and empirical behavior

HypPAu is implemented in Rust, released as open-source under GPLv3, and exposes four algorithmic variants: naive, FJS, proj, and FJS+proj (Waga et al., 16 Jul 2025). The implementation represents NAAs directly, constructs projected DFAs for pruning, and manages the search frontier lazily so that invalidations from skipping heuristics do not require full materialization of the Cartesian product of candidate tuples.

The reported experiments cover three application-oriented benchmarks—Interference, Robustness, and PacketPairs—as well as the synthetic scaling benchmark ManyDirs. The evaluation uses word lengths of 500–5000 for Interference, 200–2000 for Robustness, and 1000–10000 for PacketPairs; numbers of words from 2–10; and ManyDirs with $A = (\Sigma, K, S, I, F, \Delta),$ 2 and word length 200. The reported hardware is an Intel Xeon w5-3435X 4.5 GHz, 63 GiB RAM, Ubuntu 24.04.2, with 10 repetitions per configuration and timeout 1800 s (Waga et al., 16 Jul 2025).

The results show that heuristic utility is workload-dependent. For Interference and Robustness, the supplied description states that QS/KMP skipping had little effect because skip values were often 1, making FJS slightly slower than naive due to overhead. For PacketPairs, by contrast, skip values were frequently 2, so FJS produced pronounced speedups, and projection pruning also helped because of clear per-direction anchors such as start/end markers (Waga et al., 16 Jul 2025).

The strongest quantitative illustration is the ManyDirs benchmark. At word length 200, the elapsed times are reported as follows: for $A = (\Sigma, K, S, I, F, \Delta),$ 3, naive 0.03, FJS 0.01, proj 0.02, FJS+proj 0.02; for $A = (\Sigma, K, S, I, F, \Delta),$ 4, naive 4.65, FJS 0.95, proj 1.34, FJS+proj 1.12; for $A = (\Sigma, K, S, I, F, \Delta),$ 5, naive timeout (>1800), FJS 95.48, proj 140.42, FJS+proj 130.05 (Waga et al., 16 Jul 2025). The log–log plots are described as approximately linear, with slopes between 1.6 and 2.3, suggesting roughly quadratic scaling in practice for $A = (\Sigma, K, S, I, F, \Delta),$ 6.

The practical interpretation is narrow but important: although the underlying decision problem is $A = (\Sigma, K, S, I, F, \Delta),$ 7-complete, HypPAu can still compute precise multi-trace witnesses over thousands of letters within about a minute for two directions, and for four directions at length around 200 within a few minutes, provided the automaton structure admits effective pruning (Waga et al., 16 Jul 2025).

A secondary usage in the supplied literature connects HypPAu to hyperparameter auto-tuning through HYPPO. In that setting, HypPAu denotes automatic discovery of good hyperparameter configurations for machine learning models with minimal human effort, and HYPPO instantiates it with three pillars: adaptive surrogate modeling, explicit uncertainty quantification, and asynchronous nested parallelism (Dumont et al., 2021). HYPPO formulates deep-learning HPO as a bilevel, bi-objective black-box optimization problem with stochastic evaluations, uses RBF and Gaussian Process surrogates, incorporates MC dropout plus repeated training for predictive uncertainty, and reports both an order-of-magnitude reduction in evaluations needed to reach the optimal region and a two-orders-of-magnitude throughput improvement in a CT case using up to 96 GPUs on NERSC/Cori (Dumont et al., 2021). This is conceptually unrelated to hyper pattern matching; the commonality is only the supplied HypPAu naming connection.

A further use appears in HAP, where the supplied literature names HypPAu as Hybrid Adaptive Parallelism for efficient MoE inference (Lin et al., 26 Aug 2025). HAP decomposes transformer layers into Attention and Expert modules, models compute latency as

$A = (\Sigma, K, S, I, F, \Delta),$ 8

models communication as

$A = (\Sigma, K, S, I, F, \Delta),$ 9

and solves an ILP to choose hybrid DP/TP/EP configurations under memory and divisibility constraints (Lin et al., 26 Aug 2025). Reported speedups over TP-based inference are 1.68x on A100, 1.77x on A6000, and 1.57x on V100, with <1 s solve time on common single-node 8-GPU settings (Lin et al., 26 Aug 2025). Again, this is a distinct technical program from runtime verification.

The supplied literature also connects the label to Hyper, the Hybrid Photometry and Extraction Routine for compact-source photometry in crowded fields and variable backgrounds (Traficante et al., 2014). Hyper combines multi-Gaussian fitting with aperture photometry, performs local background fitting with 2D polynomials of order 0–4, supports simultaneous multi-wavelength photometry with a fixed aperture across bands, and reports, for example, that in a 7-source blended simulation at 160 μm, deblending improved mean flux bias from about 43% to about 7.8% (Traficante et al., 2014). This use is astrophysical and photometric rather than automata-theoretic.

Finally, HyPER provides a useful negative comparison. The supplied details state that the term HypPAu does not appear in the HyPER paper. HyPER instead stands for Hypothesis Path Expansion and Reduction, a training-free controller for multi-path decoding in mixture-of-experts LLMs that dynamically chooses between None, Branch, MultiToken, and SingleToken actions using pool statistics such as confidence, entropy, diversity, and consensus (Qiu et al., 6 Feb 2026). The provided comparison suggests only a conceptual overlap with a hypothetical “hypothesis path augmentation” interpretation of HypPAu.

6. Distinctions, limitations, and outlook

The central distinction is therefore between HypPAu as a concrete runtime verification implementation and HypPAu as a broader naming label attached, in the supplied literature, to unrelated problems. The runtime verification system is formal, automata-based, and witness-oriented; the optimization and inference systems are resource-allocation frameworks for deep learning; the photometry system is an image-analysis pipeline; and HYPA is a statistically grounded anomaly detector for temporal paths in networks rather than a variant spelling of HypPAu (Waga et al., 16 Jul 2025).

For the runtime verification usage, the main limitation is computational hardness. The search problem inherits exponential sensitivity to the number of directions and candidate start positions, and the paper explicitly notes worst-case blowups from the naive queue size $\Delta \subseteq S \times \Sigma \times K \times S$ 0 (Waga et al., 16 Jul 2025). The heuristics mitigate this when automaton structure yields useful skip values or strong per-direction pruning, but they do not change the $\Delta \subseteq S \times \Sigma \times K \times S$ 1-completeness result.

The stated future directions for the runtime verification line are timed hyper pattern matching, approximate matching under error metrics such as edit or Hamming distances, and extension to more expressive hyperproperty classes (Waga et al., 16 Jul 2025). A plausible implication is that HypPAu’s longer-term significance lies not merely in deciding whether a hyperproperty is violated, but in making such violations inspectable through localized, asynchronous, multi-trace witnesses.

In that sense, the most stable encyclopedic characterization of HypPAu is narrow rather than umbrella-like: it is chiefly the tool associated with hyper pattern matching and the computation of concrete witnesses for asynchronous hyperproperties (Waga et al., 16 Jul 2025). The other supplied usages are best read as homonymous or connected labels in adjacent research contexts rather than as parts of a single unified framework.