Directed Delay: Graph & Categorical Causality

• Directed Delay (Causal Sieve) is a framework that formalizes causal computations and constant-delay enumeration in directed graphs via default edges and recursive branching.
• It employs a delayed trace operator in category theory to model feedback where each output depends solely on past inputs, ensuring causality and guarded computation.
• The approach bridges discrete algorithmic techniques for m-length walk enumeration with continuous differentiable programming, supporting efficient backpropagation through time.

Directed Delay, also known as the Causal Sieve or Delayed Trace, refers to a mathematical structure and algorithmic technique that formalizes causality, guarded feedback, and efficient enumeration within directed graph and categorical computational frameworks. This concept arises in two major contexts: (1) the design of constant-delay enumeration algorithms for walks in directed graphs, and (2) the categorical formalization of causal computations via delay-induced feedback, especially as it relates to differentiable programming and temporal sequence processing.

1. Enumeration with Directed Delay: Constant-Delay m-Length Walks in Directed Graphs

Given a directed graph $G=(V,E)$, a walk of length $k$ is a sequence of edges $((v_1,v_2), (v_2,v_3), ..., (v_k, v_{k+1}))$ such that each $(v_i,v_{i+1}) \in E$. The enumeration problem seeks to output, with worst-case $O(1)$ time ("constant delay") after preprocessing, all walks of fixed length $m$ in $G$, each represented succinctly (Adamson et al., 2024). Direct output of explicit walks incurs $\Theta(m)$ time per walk; thus, an implicit representation is adopted, encoding each walk as a triple $$(\ell_{shared}, \text{non-default-edge}, \ell_{tail})$$ that allows walk reconstruction in constant time from the previous output.

Algorithmic Preprocessing

1. Longest Walks and Edge Sorting: Compute $\pi(v)$, the length of the longest walk starting at each $v$ (or $\infty$ if $v$ reaches a cycle), and for each vertex, construct $L_v$, the sorted list of outgoing edges by "longest-walk-if-used" value.
2. Default Edge Pseudoforest: For each $v$, the first entry in $L_v$ is designated as the default edge $d(v)$, forming a pseudoforest $D$.
3. Component Decomposition: Each component of $D$ is either a directed cycle (possibly with attached trees) or an isolated tree. For trees, level-ancestor data structures are built to permit $O(1)$ jumps to ancestors at specified depths; for cycles, a doubly-laid-out array enables index arithmetic for advances.
4. Best Branch Extension Precomputation: For efficient branching, $w(v)$ records the best extension length via non-default edges; the routine $\mathrm{PMN}(s,\ell)$ computes, in $O(1)$, the optimal branch point along a default walk of length $\ell$ from $s$.
5. Data Synopsis: After $O(n + e)$ preprocessing, the algorithm maintains $\pi(v)$, $L_v$, $d(v)$, decomposition data, $w(v)$, and $\mathrm{PMN}$ support, ensuring constant-delay enumeration.

Enumeration Mechanism

The enumeration algorithm manages two stacks:

• $S$: Encodes the current walk as alternating default segments and non-default branches.
• $C$: Tracks recursion frames for restoring $S$ after branching.

Output is structured via triples indicating where to deviate from the prior walk (keeping the first $p$ edges), which non-default edge to take, and the length $\ell$ of the subsequent default segment. Internal iterative and recursive operations per output are strictly $O(1)$. The core theorems established are:

• Correctness: Each walk of length $m$ is enumerated exactly once in its succinct form.
• Constant Delay: Each output step occurs with fixed $O(1)$ delay following preprocessing (Adamson et al., 2024).

2. The Causal Sieve in Category Theory: Delayed Trace Construction

The delayed trace operator, or causal sieve, is a category-theoretic mechanism modeling causal computations wherein the $n$th output is a function only of inputs up to $n$ (Sprunger et al., 2019). It is formalized via the category $\mathbf{St}(C)$, constructed from any strict Cartesian category $C$.

Core Definitions and Structure

• Stateful Morphism Sequence: A morphism in $\mathbf{St}(C)$ from sequences $X = [X_0,X_1,...]$ to $Y = [Y_0,Y_1,...]$ consists of an initial state and a sequence of maps $$s_k : S_k \times X_k \rightarrow S_{k+1} \times Y_k$$, modeling state and output evolution.
• Truncation and Extensional Equality: Truncation to $n$ steps yields the finite composite of layers up to $n$, and two morphism sequences are extensionally equal if all truncations agree.
• Categorical Semantics: In $\mathbf{St}(\mathbf{Set})$, morphisms correspond identically to causal (history-dependent) functions $A^\mathbb{N} \rightarrow B^\mathbb{N}$ with the required dependency property.

The Delayed Trace Operator

The delayed trace, $\text{delayTr}_{T,p}(i,s)$, wraps feedback wires through a one-step register (initialized with $p$), enforcing causal feedback. This trace is not a standard monoidal trace:

• Axioms Satisfied: Target-naturality, source-naturality, superposing, vanishing (empty/nested traces).
• Axioms Failed: Yanking ($\text{delayTr}$ of a swap is not the identity but a delay gate), ordinary dinaturality; only a modified form holds.

This operator ensures:

• Causality: Each output at time $n$ is influenced solely by inputs through $n$ and the initial state.
• Guardedness: Feedback requires passage through at least one delay register, preventing instantaneity.

3. Differentiable Structure and Backpropagation Through Time

When $C$ is a Cartesian differential category, its structure lifts to $\mathbf{St}(C)$. Differential operators are defined so that both state and value derivatives are computed:

• The assignments satisfy all axioms of a Cartesian differential category.
• Chain rule and Schwarz theorem hold in $\mathbf{St}(C)$.
• Backpropagation Through Time (BPTT): Differentiation in $\mathbf{St}(C)$ corresponds to abstract BPTT, so unrolling followed by differentiation is rendered equivalent to differentiating the stateful morphism sequence directly—i.e., explicit unrolling is not required.

4. Practical Illustration: Running-Sum Transformer

As an exemplar, consider $C=\mathbf{Set}$ and the map $\phi: S \times X \rightarrow S \times Y$, with $\phi(s,x) = (s+x, s)$. With $p = 0$, the delayed trace yields a transformer computing the running sum $s_{k+1} = s_k + x_k$, outputting $s_k$ at each step. Its differentiated variant computes in parallel both the sum and its derivative with respect to perturbed inputs, with both tracked through corresponding delay-gates (Sprunger et al., 2019).

5. Relationship Between Algorithmic and Categorical Delay Frameworks

The constant-delay enumeration of walks in directed graphs and the categorical delayed trace share foundational concerns with causality, temporal dependency, and feedback guardedness. In enumeration, default-edge paths and branching echo the sequential layer application in stateful morphism sequences. Guarded feedback via delay-trace directly enforces no instantaneous cyclic dependencies, paralleling pseudoforest decomposition in the enumeration setting.

The delayed mechanisms in both contexts convert cycles (potentially permitting unbounded or circular dependencies) into well-structured, causally valid enumerations or computations, facilitating efficient traversal, representation, and differentiation.

6. Theoretical and Applied Significance

Directed delay frameworks have far-reaching implications:

• Algorithmic Enumeration: Enable practical enumeration of combinatorially rich path structures in linear or automaton-theoretic objects with optimal delay performance (Adamson et al., 2024).
• Causal Modeling in Computation: Provide category-level foundations for signal processing, temporal logic, and stateful system semantics, extending to differentiable systems for machine learning without explicit state unrolling (Sprunger et al., 2019).
• Feedback Guardedness: Establish intrinsic constraints for preventing unphysical or ill-defined instantaneous cycles, critical in both theoretical computer science and control theory.

A plausible implication is the transfer of categorical causal sieve concepts to efficient automata processing and sequence modeling, unifying discrete algorithmic and continuous computational perspectives.