Directed Ordered Acyclic Graphs (DOAGs)

• Directed Ordered Acyclic Graphs (DOAGs) are finite DAGs with an inherent partial order that guarantees all edge directions follow a prescribed, application-dependent sequence.
• They support efficient enumeration and random generation through recursive decomposition and anticipated rejection techniques, reducing complexity in structurally constrained graphs.
• Their use in causal inference, distributed algorithms, and neural architecture design demonstrates improved statistical power, computational efficiency, and interpretability.

A directed ordered acyclic graph (DOAG) is a finite directed acyclic graph (DAG) in which the vertices admit a partial order that is explicitly linked to the directions of edges: for every edge $(u \to v)$, one requires $u \prec v$, and the reachability relation coincides with the given order. Many research communities use the term “DOAG” to emphasize the crucial connection between graph topology and a partial (often application-dependent) ordering on nodes, a feature not present in arbitrary DAGs. This class of graphs arises in numerous domains, including causal modeling, scheduling, distributed algorithms, knowledge representation, and machine learning architectures, where order constraints encode irreversibility, causality, or resource dependencies (Rey et al., 13 Jun 2025, Pépin et al., 2023, Luo et al., 2022, Shojaie et al., 2024, Rahman et al., 2019, Gupta et al., 2023).

1. Formal Definitions and Structural Properties

A DOAG is formally defined as a triple $(V, E, \prec)$, where $V$ is a finite set of nodes, $E \subseteq V \times V$ is a set of directed edges, and $\prec$ is a partial order on $V$ such that $(u \to v) \in E$ implies $u \prec v$ (Gupta et al., 2023, Pépin et al., 2023, Shojaie et al., 2024). The reachability relation in the graph—that is, the existence of a directed path from $u$ to $u \prec v$0—necessarily extends the partial order, and in standard constructions, the two are taken to coincide.

• Partial order: A binary relation $u \prec v$1 that is transitive, antisymmetric, and reflexive (or irreflexive by convention). Not all node pairs need be comparable.
• Acyclicity: The absence of directed cycles implies that the partial order is well-defined, and a topological ordering exists for every DOAG.
• Order preservation: All edges go "upward" in the partial order, and every walk following directed edges must eventually reach a sink (a maximal element).

DOAGs generalize the combinatorial notion of DAGs by requiring an explicit (often external) ordering constraint, which may be provided as a total or partial causal, temporal, or functional ordering. Specialized variants also incorporate, at each node, a total order on the set of outgoing edges and a total order on the sources of the graph, resulting in the “ordered out-edge DOAG” model relevant for sampling and enumeration (Pépin et al., 2023).

2. Enumeration, Asymptotics, and Random Generation

The enumeration and random generation of DOAGs are significantly more complex than for classical DAGs due to the extra ordering structure (Pépin et al., 2023). Key combinatorial results are:

• The number of DOAGs with $u \prec v$2 vertices, $u \prec v$3 edges, and $u \prec v$4 sources, denoted $u \prec v$5, satisfies a recurrence based on the decomposition of the smallest source and the ordered reconnection of outgoing edges.
• The total number $u \prec v$6 of DOAGs on $u \prec v$7 vertices grows super-factorially:

$u \prec v$8

for an explicit constant $u \prec v$9, with further refinement for fixed $(V, E, \prec)$0.

• For large $(V, E, \prec)$1, random DOAGs are almost always singly-sourced and have nearly $(V, E, \prec)$2 edges.

Efficient uniform random generation is achieved through recursive decomposition methods and an “anticipated rejection” algorithm based on variation matrices, both with polynomial expected complexity in $(V, E, \prec)$3 and full control on the number of edges and sources (Pépin et al., 2023).

Parameter	Standard DAGs	DOAGs
Typical enumeration	$(V, E, \prec)$4	$(V, E, \prec)$5
Out-edge ordering	Absent	Partial/total at each node, sources
Uniform sampler complexity	$(V, E, \prec)$6	$(V, E, \prec)$7 random bits

3. Learning Algorithms and Estimation under Ordering Constraints

Learning the structure of a DOAG from data occurs in varied regimes, each exploiting a different form of ordering information. A key paradigm is the incorporation of known partial or partition-based ordering to dramatically constrain the feasible edge set and search space (Shojaie et al., 2024, Rahman et al., 2019, Zantedeschi et al., 2023).

• Partition-DAG and PODAG: When the node set is partitioned into ordered blocks, efficient penalized likelihood estimators impose block-wise lower triangularity of the Cholesky factor, with within-block structure arrived at by coordinate minimization, lasso screening, or PC-style search (Rahman et al., 2019, Shojaie et al., 2024).
• Permutahedron Optimization: Direct optimization over the permutahedron (the polytope of all node orderings) ensures that candidate graphs are always acyclic by construction, with edge-parameter learning modularized to accommodate arbitrary loss functions and regularizers (Zantedeschi et al., 2023).
• Algorithmic Search Space Reduction: Known partial ordering (layers, partitions, or externally supplied relations) restricts candidate parent sets and greatly reduces the conditioning sets required in conditional independence testing, conferring notable computational speedups and improved finite-sample accuracy (Shojaie et al., 2024).
• Identifiability and Consistency: Incorporation of ordering information narrows the Markov equivalence class and can guarantee (under additional faithfulness and eigenvalue conditions) consistent edge recovery even in high-dimensional settings (Shojaie et al., 2024, Rahman et al., 2019).

Empirical studies demonstrate substantial improvements in both statistical power and computational efficiency across biological and simulated network data, with performance monotonic in the informativeness of the partial order (Rahman et al., 2019).

4. Graph Signal Processing and Deep Neural Architectures

Recent advances extend graph convolutional and neural architectures to the DOAG setting, explicitly leveraging the partial order structure (Rey et al., 13 Jun 2025, Rey et al., 2024, Luo et al., 2022). Key innovations include:

• Causal Graph-Shift Operators: For a DOAG with adjacency $(V, E, \prec)$8, the weighted transitive closure $(V, E, \prec)$9 encodes that $V$0 iff $V$1. Node-indexed causal shift operators $V$2 (where $V$3 selects predecessors of $V$4) aggregate only from valid ancestors (Rey et al., 13 Jun 2025, Rey et al., 2024).
• Graph Convolutional Networks (DCN/PDCN): Layer constructions based on these shifts permit shift-invariant filtering, with theoretical guarantees of permutation equivariance and expressivity. PDCN decouples model size from graph size by sharing weights, allowing scaling to very large DOAGs (Rey et al., 13 Jun 2025).
• Transformers on DOAGs: Model adaptation of attention architectures employs k-hop reachability restriction on the self-attention mask, with positional encodings given by node depth in the partial order. This reduces per-layer complexity from quadratic to linear in the (averaged) receptive field size and yields systematic improvements over structure-agnostic baselines (Luo et al., 2022).
• Spectral Interpretation: The causal shift and transitive closure matrices act as "DOAG Fourier transforms," supporting spectral-domain reasoning and filter design (Rey et al., 13 Jun 2025, Rey et al., 2024).

In experiments on synthetic, biological, and engineering datasets, DOAG-aware architectures substantially outperform traditional GNNs, multilayer perceptrons, and even specialized DAG neural models, often with fewer parameters and improved robustness to noise (Rey et al., 13 Jun 2025, Rey et al., 2024, Luo et al., 2022).

Model Type	Key Mechanism	Complexity	Performance (relative)
DCN/PDCN	Causal graph-shift, filters	$V$5, can subset	Highest, parameter-efficient
DAG-Transformer	Masked attention, depth encoding	$V$6	Substantially higher than vanilla transformer/GNN
MLP/GNN	Structure-agnostic	$V$7	Lower

5. Representation, Embedding, and Theoretical Analysis

Embedding DOAGs in continuous spaces for downstream learning and reasoning has motivated new frameworks exploiting the underlying partial order (Suzuki et al., 2019). Leading approaches include:

• Disk Embeddings: Each node is mapped to a formal disk $V$8 in a quasi-metric (often hyperbolic) space, with $V$9 if and only if the disk of $E \subseteq V \times V$0 is included in that of $E \subseteq V \times V$1 ($E \subseteq V \times V$2). This formalism encompasses earlier order embeddings and hyperbolic entailment cones as special cases and captures "two-sided" exponential growth in both ancestor and descendant populations—phenomena observed in complex DAGs but not trees (Suzuki et al., 2019).
• Loss and Optimization: Disk inclusions are enforced by margin-based hinge losses on the disk protrusion, trained using Riemannian stochastic gradient descent in hyperbolic space (Suzuki et al., 2019).
• Representational Capacity: Hyperbolic disk embeddings achieve low-distortion representation of general DOAGs, including those with rapidly growing ancestor and descendant sets. Theoretical results clarify the isomorphism between disk inclusion in hyperbolic space, polyhedral quasi-metrics, and previously introduced DAG embedding frameworks (Suzuki et al., 2019).
• Empirical Evaluation: On general DAGs (e.g., reversed WordNet), disk embedding models strictly outperform order embeddings and hyperbolic cones, achieving F1 ≈ 88–90% where earlier models degrade sharply.

6. DOAGs in Distributed and Asynchronous Algorithms

A foundational application appears in the analysis and construction of asynchronous, self-stabilizing distributed algorithms (Gupta et al., 2023). Here, system states are themselves organized as vertices of a DOAG (the "$E \subseteq V \times V$3-DAG"), where the partial order is induced from local state updates:

• DOAG-Inducing Problems: Problems are classified as DAG-inducing if, for every suboptimal state, there exists an "impedensable" node whose progress is necessary to globally increase the system order. Self-stabilizing algorithms are then exactly those whose global state transition graph forms a DOAG whose sinks coincide with optimal states.
• Necessary and Sufficient Condition: An asynchronous algorithm converges correctly if and only if its global-state transition relation forms a DOAG; the existence of a partial order on local states ensuring all correction moves ascend the order is both necessary and sufficient for asynchrony-tolerance (Gupta et al., 2023).
• Reduction of Verification Complexity: Verification is reduced to reasoning about the local partial order, bypassing the combinatorial explosion of global state enumeration.

This abstraction provides a unifying language for the design and analysis of a broad class of asynchronous optimization, scheduling, and convergence algorithms.

7. Applications and Empirical Impact

DOAGs are central in modeling and solving problems where information or influence must propagate in a direction respecting some partial order:

• Causal Inference and Genomics: Bipartite and multi-layer DOAG structures reflect known biological hierarchies, reducing the search space for regulatory networks or eQTL mapping (Shojaie et al., 2024, Rahman et al., 2019).
• Neural Architecture Search and Scheduling: DOAGs model permissible architecture flows or job schedules subject to resource constraints (Rey et al., 13 Jun 2025).
• Hydrological Forecasting and System Identification: Time-respecting network diffusion processes naturally form DOAGs, where convolutional networks exploiting the order provide accurate, efficient estimators compared with classical or generic deep baselines (Rey et al., 13 Jun 2025).
• Knowledge Representation: Partial orderings and reachability in DOAGs are exploited for scalable embeddings and reasoning in large knowledge graphs (Suzuki et al., 2019).

Across these settings, DOAG-driven methods yield improvements in estimation accuracy, computational speed, and interpretability, especially where ordering information is strong or structured.

---

References

• (Rey et al., 13 Jun 2025) "Directed Acyclic Graph Convolutional Networks"
• (Pépin et al., 2023) "Asymptotic analysis and efficient random sampling of directed ordered acyclic graphs"
• (Luo et al., 2022) "Transformers over Directed Acyclic Graphs"
• (Shojaie et al., 2024) "Learning Directed Acyclic Graphs from Partial Orderings"
• (Rahman et al., 2019) "Estimation of Gaussian directed acyclic graphs using partial ordering information with an application to dairy cattle data"
• (Gupta et al., 2023) "DAG-Inducing Problems and Algorithms"
• (Zantedeschi et al., 2023) "DAG Learning on the Permutahedron"
• (Suzuki et al., 2019) "Hyperbolic Disk Embeddings for Directed Acyclic Graphs"
• (Rey et al., 2024) "Convolutional Learning on Directed Acyclic Graphs"