DAG Projections in Graphs & Optimization

• DAG projections are transformations that map cyclic graphs or weighted matrices to directed acyclic graphs under acyclicity constraints while preserving key structural properties.
• They employ explicit combinatorial methods, low-rank additive models, and convex projections to achieve scalable and efficient algorithmic implementations.
• These techniques underpin advancements in causal inference, optimization, and large-scale graph algorithms by ensuring accurate distance, flow approximations, and reduced computational complexity.

A DAG projection is a transformation or mapping that associates a possibly cyclic or general directed graph, or a weighted adjacency matrix, with a directed acyclic graph (DAG) or its corresponding adjacency structure under explicit or implicit constraints. DAG projections are foundational in large-scale structure learning, optimization, and combinatorial graph algorithms, as they enable computational methods, acyclicity enforcement, and approximation reductions across various disciplines, particularly in causal inference, optimization, and graph algorithms.

1. Formal Definitions and General Concepts

DAG projections arise in multiple domains:

• Pure Graph-theoretic Projections: Given a directed graph $G=(V,E)$, a DAG projection is a construction of a DAG $D=(V',E')$ with a surjective map $\pi:V'\to V$ such that every edge in $D$ corresponds to a valid edge in $G$. The central structural property is acyclicity of $D$, with the additional goal of preserving combinatorial or metric features (e.g., distances, flows) up to a controlled approximation (Haeupler et al., 6 Apr 2026).
• Matrix Projections in Optimization: Given a real (possibly cyclic) weighted adjacency matrix $W\in\mathbb{R}^{d\times d}$, the projection seeks a matrix $A$ representing a DAG that is optimal with respect to a metric (Frobenius norm), subject to explicit algebraic acyclicity constraints.

Two principal categories emerge:

• Combinatorial DAG projections: Transform graph objects for algorithmic or approximation purposes.
• Optimization-based DAG projections: Map general matrices to DAG-feasible regions for statistical modeling and causal structure learning.

2. Algebraic and Convex Characterizations of Acyclicity

Enforcing the DAG constraint in optimization typically leverages smooth or convex algebraic surrogates:

• Exponential Trace Constraint ("No-Tears"): $$h_\mathrm{exp}(W) = \mathrm{Tr}[\exp(W\circ W)] - d$$. Here, $W$ is a weighted adjacency matrix (elementwise squared), and $D=(V',E')$0 if and only if $D=(V',E')$1 encodes a DAG (Ziu et al., 2024, Dong et al., 2022).
• Log-Determinant Constraint ("DAGMA"/Convex Form): $D=(V',E')$2, $D=(V',E')$3, also exactly characterizes acyclicity. For non-negative $D=(V',E')$4, $D=(V',E')$5 with equality if and only if $D=(V',E')$6 is nilpotent (i.e., defines a DAG) (Rey et al., 2024).

These relaxations enable the formulation of continuous optimization problems and are central to recent algorithmic advancements.

3. Projection Operators and Algorithms in Optimization

3.1 Explicit Combinatorial Projection

The $D=(V',E')$7DAG framework (Ziu et al., 2024), instead of imposing soft penalties, introduces an explicit projection operator $D=(V',E')$8:

• Given $D=(V',E')$9, construct a permutation $\pi:V'\to V$0 (heuristic topological ordering) by recursively removing "least-connected" nodes (based on row/column norms), forming a mask $\pi:V'\to V$1 such that $\pi:V'\to V$2 only if $\pi:V'\to V$3. The projected matrix is $\pi:V'\to V$4.
• This operator runs in $\pi:V'\to V$5 time and guarantees exact feasibility with respect to DAG constraints.

3.2 Low-rank Additive Model (LoRAM)

LoRAM projects $\pi:V'\to V$6 to a "closest" DAG matrix $\pi:V'\to V$7 by minimizing $\pi:V'\to V$8 subject to $\pi:V'\to V$9 (Dong et al., 2022):

• $D$0, with $D$1 (low-rank) and $D$2 (sparse) restricted to the support of $D$3.
• Gradients involving the matrix exponential are efficiently approximated using fast sparse-dense products and truncated Taylor expansions; total per-iteration complexity is $D$4.

3.3 Convex Projections for Non-negative Weights

For $D$5, one uses the log-determinant constraint $D$6; the projection solves:

$D$7

This is solved efficiently via method-of-multipliers algorithms, with the acyclicity set being convex (Rey et al., 2024). When the input is non-negative, this leads to global convergence.

4. DAG Projections in Combinatorial Graph Algorithms

The notion of a DAG projection extends to graph algorithmics, where a possibly cyclic graph $D$8 is mapped to a DAG $D$9 (possibly with vertex replication via $G$0), such that structural properties (distances, flows) are approximately preserved (Haeupler et al., 6 Apr 2026). The main formalizations are:

Type	Guarantee	Projection Size
(1+$G$1)-distance preserving	$G$2	$G$3 edges
$G$4-congestion-preserving	$G$5	$G$6 edges

• Constructions are recursive and employ low-diameter decompositions and expander decompositions to preserve structure within acyclic overlays, at near-linear (in $G$7) runtime and size.

5. Algorithmic Complexity, Parallelization, and Scalability

• For optimization-based DAG projections, per-iteration complexity has evolved from $G$8 (matrix exponentials and penalties) to $G$9 (explicit projection) or $D$0 (low-rank), removing the cubic barrier and enabling scalability to $D$1--$D$2 nodes (Ziu et al., 2024, Dong et al., 2022).
• Parallel graph-theoretic DAG projections support $D$3 work and $D$4 depth, provided fast oracles exist for SSSP or MFMC on DAGs (Haeupler et al., 6 Apr 2026).

6. Guarantees, Applications, and Reductions

• Structure learning: Projected methods (e.g., $D$5DAG) provably converge to feasible local minima with optimality rates in expectation, and empirically outperform penalty-based methods when scaling to large $D$6 (Ziu et al., 2024).
• Distance and flow approximation: (1+$D$7)-distance DAG projections enable preservation of pairwise distances and efficient construction of preservers, hop-sets, and approximate min-cuts in general graphs at near-DAG bounds (Haeupler et al., 6 Apr 2026).
• Reductions to open problems: Major parallel graph problems, such as exact single-source shortest path or maximum flow in general graphs, are reduced to near-equivalent problems on DAGs, facilitating algorithmic transfer from the acyclic setting.

7. Comparative Perspective and Implications

• Penalty-based or augmented Lagrangian approaches are sensitive to ill-conditioning and hyperparameter choice, and scale poorly, as matrix-exponential or log-determinant computations are $D$8 (Ziu et al., 2024).
• Explicit and projection-based methods remove auxiliary parameters, avoid vanishing-gradient issues near cycles, and enforce acyclicity exactly and efficiently (Ziu et al., 2024, Dong et al., 2022).
• Combinatorial DAG projections generalize prior “DAG covers” and tree-sparsifier techniques to the directed setting, facilitating algorithmic simplifications and enabling approximation guarantees in otherwise difficult domains (Haeupler et al., 6 Apr 2026).
• The convexity of the acyclicity set under non-negativity (log-det formulation) provides a rare regime where global optimality is achievable in a problem that is normally non-convex and combinatorial (Rey et al., 2024).

In conclusion, DAG projections provide a unifying framework bridging combinatorial graph algorithms and non-convex continuous optimization. Recent advances have dramatically improved scalability, theoretical guarantees, and practical applicability across a range of problems in statistical modeling, causal inference, and large-scale algorithmics (Ziu et al., 2024, Dong et al., 2022, Rey et al., 2024, Haeupler et al., 6 Apr 2026).