Graph Reconstruction: Models & Techniques

• Graph Reconstruction is the process of inferring unknown graph edges or structures from diverse partial observations such as oracle queries, noisy traces, and density functions.
• It integrates methods from combinatorics, topology, and geometry to recover properties like connectivity, loop structure, and metric approximations.
• Recent advances focus on efficient algorithms that balance query or sample complexity while ensuring accurate recovery of both combinatorial and geometric properties.

Graph reconstruction (GR) denotes a family of inverse problems in which one seeks to recover a graph, or recover structure encoded by a graph, from partial observations. Across current research, the observation model varies widely: the input may be oracle answers on vertex subsets, anonymized noisy traces, a density function concentrated around an embedded graph, path-correlation measurements, connected triples, vertex-deleted subgraphs, observed diffused signals, graph embeddings, or even model gradients (Konrad et al., 2024). The unifying questions are identifiability, algorithm design, query or sample complexity, and the extent to which reconstruction preserves topology, geometry, or combinatorial structure (Fasy et al., 2019).

1. Problem families and formal viewpoints

A central formulation treats GR as hidden-edge recovery. The vertex set $V$ is known, the edge set $E$ is unknown, and an algorithm interacts with an oracle. In the MIS-query model, a query $U \subseteq V$ returns a maximal independent set in $G[U]$ (Konrad et al., 2024). In the connected-components model, a query $S \subseteq V$ returns the partition of $S$ into connected components of $G[S]$ (Harviainen et al., 5 Sep 2025). In the distance-query model, the unknown graph is connected, unweighted, and bounded-degree, and the goal is to find every edge using distance queries between vertices (Mathieu et al., 2021).

A second major formulation is reconstruction from random partial subgraphs. In the noisy-trace model, each trace is produced by sampling each vertex independently with probability $p_v$, taking the induced subgraph, and then applying either edge deletions or edge flips; traces are anonymized, so matching vertices across traces is itself part of the problem (McGregor et al., 2024).

A third formulation is geometric or metric graph reconstruction. Here the hidden object is an embedded graph $G \subset \mathbb{R}^2$ or, more generally, a latent manifold whose geometry is indirectly encoded by the observed graph. Discrete Morse-theoretic methods reconstruct graph-like structures from a density function concentrated near the embedded graph (Dey et al., 2018), while random geometric graph methods reconstruct intrinsic and extrinsic geometry of the underlying manifold from the sampled graph alone (Huang et al., 2024).

A fourth formulation is reconstruction from combinatorial summaries. Examples include reconstruction from the set $T_3(G)$ of connected triples (Qi, 2023), from path correlation data measuring overlaps of boundary-to-boundary paths in directed weighted networks (Berkolaiko et al., 2018), and from the deck or $E$0-deck of vertex-deleted or induced subgraphs (Cotta et al., 2021). In the classical asymmetry-based program, unique induced subgraphs called anchors reduce reconstruction to smaller shadow graphs (Farhadian, 2016).

The term also appears in representation-learning and inverse-privacy settings. RESTORE uses graph reconstruction from embeddings as an intrinsic evaluation of graph embeddings (Yip et al., 2023), and GRAIN reconstructs graph structure and node features from GNN gradients in federated learning (Drencheva et al., 3 Mar 2025). By contrast, some adjacent literature uses “reconstruction” for signals or data defined on a known graph rather than reconstruction of the graph itself; this distinction is important in graph signal processing (Ferrer-Cid et al., 2022).

2. Geometric and topological reconstruction

In metric graph reconstruction from density data, discrete Morse theory has become a principal tool. The persistence-guided framework studied in “Graph Reconstruction by Discrete Morse Theory” simplifies the original algorithm by using persistence pairs from the lower-star filtration of the negated density $E$1, canceling low-persistence features, and outputting the union of 1-unstable manifolds of remaining critical edges (Dey et al., 2018). The paper proves that it suffices to cancel only vertex-edge pairs and that the simplified procedure outputs the same graph as the standard persistence-guided Morse cancellation algorithm, with time complexity $E$2 (Dey et al., 2018).

The theoretical guarantees in that framework are topological and geometric. Under the $E$3-approximation model, with $E$4, the reconstructed graph $E$5 satisfies

$E$6

and preserves loop structure in the sense that

$E$7

In the planar case, $E$8 deformation retracts onto both $E$9 and $U \subseteq V$0 (Dey et al., 2018). These statements formalize “geometric closeness” and “same loop structure” rather than exact embedding recovery.

“Threshold-Based Graph Reconstruction Using Discrete Morse Theory” extends this line by introducing a double-threshold noise model that separates vertex regions from edge regions (Fasy et al., 2019). The density function $U \subseteq V$1 is an $U \subseteq V$2-approximation if

$U \subseteq V$3

with $U \subseteq V$4 and $U \subseteq V$5 (Fasy et al., 2019). This model is designed to recover not only cycles but also degree-1 vertices or “leaves,” a limitation explicitly attributed to earlier single-threshold formulations (Fasy et al., 2019). The reconstruction algorithm computes persistence pairs on the super-level filtration, performs Morse cancellation for all pairs of persistence $U \subseteq V$6, and adds the stable manifolds of edges in persistence pairs with persistence $U \subseteq V$7; the output $U \subseteq V$8 has the same homotopy type as $U \subseteq V$9 and satisfies

$G[U]$0

The paper emphasizes recovery of the full structure of $G[U]$1, including vertices of all degrees (Fasy et al., 2019).

A different geometric regime appears in random geometric graphs. “Reconstructing the Geometry of Random Geometric Graphs” assumes a compact, connected, smooth, embedded $G[U]$2-dimensional manifold $G[U]$3, a measure $G[U]$4, and a strictly decreasing, smooth, Lipschitz link function $G[U]$5 (Huang et al., 2024). From the observed graph alone, the algorithm constructs clusters, estimates distances using common-neighbor statistics, builds a net, and outputs a weighted graph $G[U]$6 whose metric approximates both geodesic and Euclidean distances. With high probability, for every pair $G[U]$7,

$G[U]$8

and similarly for Euclidean distances (Huang et al., 2024). The paper positions this as a complement to manifold learning, where approximate distances are usually given rather than inferred from combinatorial structure.

3. Oracle and trace-based hidden-graph recovery

In oracle-based GR, performance is measured by query complexity. “A Simple Algorithm for Graph Reconstruction” analyzes a two-phase distance-query algorithm called Simple (Mathieu et al., 2021). It samples a random landmark set $G[U]$9, queries $S \subseteq V$0 for every $S \subseteq V$1 and $S \subseteq V$2, constructs a candidate set $S \subseteq V$3 by the rule

$S \subseteq V$4

and then verifies each candidate by querying $S \subseteq V$5 (Mathieu et al., 2021). On random $S \subseteq V$6-regular graphs with $S \subseteq V$7, the expected total query complexity is $S \subseteq V$8, i.e.

$S \subseteq V$9

while on arbitrary bounded-degree graphs, setting $S$0 yields $S$1 (Mathieu et al., 2021). The same framework gives $S$2 queries to an all-distances oracle and $S$3 queries to a betweenness oracle on random regular graphs, and it implies metric dimension at most $S$4 with high probability (Mathieu et al., 2021).

MIS queries provide a stronger oracle than ordinary IS queries. “Graph Reconstruction via MIS Queries” proves that a randomized non-adaptive algorithm using uniform random samples of $S$5 succeeds with high probability in $S$6 queries, while a deterministic non-adaptive algorithm uses $S$7 queries (Konrad et al., 2024). The lower bounds are

$S$8

for arbitrary randomized adaptive algorithms,

$S$9

for deterministic non-adaptive algorithms, and

$G[S]$0

even when the input graph is an $G[S]$1-vertex cycle (Konrad et al., 2024). The same paper states that MIS queries are strictly more powerful than IS queries when parameterized by maximum degree $G[S]$2, and cites the IS-query lower bound

$G[S]$3

for graph reconstruction (Konrad et al., 2024).

The connected-components oracle is stronger still in several parameter regimes. “Graph Reconstruction with a Connected Components Oracle” shows that an adaptive randomized algorithm reconstructs any graph using

$G[S]$4

CC queries, where $G[S]$5 is the number of edges, $G[S]$6 the maximum degree, and $G[S]$7 the treewidth (Harviainen et al., 5 Sep 2025). Matching lower bounds up to the $G[S]$8 factor are proved: no algorithm can solve the problem in $G[S]$9 CC queries, and any non-adaptive algorithm requires $p_v$0 CC queries even with $p_v$1 and $p_v$2 (Harviainen et al., 5 Sep 2025). The paper also states that CC is stronger than separation oracles, and for some instances stronger than MIS queries (Harviainen et al., 5 Sep 2025).

In the noisy-trace setting, “Graph Reconstruction from Noisy Random Subgraphs” distinguishes sharply between random and arbitrary graphs (McGregor et al., 2024). For $p_v$3, the edge-deletion model admits reconstruction from

$p_v$4

traces, and the edge-flip model from

$p_v$5

traces, under the parameter ranges stated in Theorem 1 (McGregor et al., 2024). The algorithmic core is consistent labeling across anonymized traces using pairwise matching, subgraph isomorphism, and vertex signatures (McGregor et al., 2024). By contrast, for arbitrary graphs, $p_v$6 traces are necessary even when $p_v$7 and $p_v$8 (McGregor et al., 2024).

A distributed analogue appears in the congested clique. “Graph Reconstruction in the Congested Clique” proves the information-theoretic lower bound

$p_v$9

where $G \subset \mathbb{R}^2$0 is the number of rounds and $G \subset \mathbb{R}^2$1 the per-link bandwidth (Montealegre et al., 2017). That bound is tight: hereditary graph classes admit a one-round private-coin randomized strong-reconstruction algorithm with cost $G \subset \mathbb{R}^2$2, and arbitrary graph classes admit a two-round deterministic weak-reconstruction algorithm and a two-round private-coin randomized strong-reconstruction algorithm with the same asymptotic cost (Montealegre et al., 2017). The construction combines randomized fingerprinting with error-correcting codes to separate graph inputs (Montealegre et al., 2017).

4. Reconstruction from partial combinatorial and network data

Path-based observations can determine a directed weighted network exactly under explicit structural conditions. “Graph Reconstruction from Path Correlation Data” defines PCD through lengths of paths $G \subset \mathbb{R}^2$3 between boundary vertices and lengths of path intersections such as $G \subset \mathbb{R}^2$4 and $G \subset \mathbb{R}^2$5 (Berkolaiko et al., 2018). The main theorem states that a weighted directed graph is uniquely reconstructible from PCD if and only if every edge is used by at least one observed path, every internal vertex is non-trivial, and every internal vertex is non-separable (Berkolaiko et al., 2018). The algorithm initializes each reconstructed boundary-to-boundary path, inserts labels for junctions at positions dictated by PCD, and recursively propagates those labels across all paths consistent with the measured overlaps (Berkolaiko et al., 2018). When the conditions fail, the algorithm returns the unique minimal cleaned network consistent with the same PCD (Berkolaiko et al., 2018).

Connected triples provide a much coarser summary. For a graph $G \subset \mathbb{R}^2$6,

$G \subset \mathbb{R}^2$7

(Qi, 2023). “Graph Reconstruction from Connected Triples” gives a complete characterization of strongly $G \subset \mathbb{R}^2$8-reconstructible graphs on $G \subset \mathbb{R}^2$9 vertices: $T_3(G)$0 is uniquely determined by $T_3(G)$1 if and only if no two vertices satisfy

$T_3(G)$2

and every edge inside a triangle is necessary in the sense of the paper’s induced-subgraph families (Qi, 2023). The paper also states uniqueness within the class of regular planar graphs, 5-connected planar graphs, certain strongly regular graphs, and complete multipartite graphs, and non-uniqueness for $T_3(G)$3-connected planar graphs with $T_3(G)$4, Eulerian graphs, and Hamiltonian graphs (Qi, 2023).

Classical reconstruction theory uses the deck of vertex-deleted subgraphs. “Reconstruction for Powerful Graph Representations” recalls the standard deck

$T_3(G)$5

and the $T_3(G)$6-deck

$T_3(G)$7

(Cotta et al., 2021). The same paper turns these objects into learnable graph representations, showing that $T_3(G)$8-reconstruction can approximate any $T_3(G)$9-reconstructible function and can boost GNN expressive power enough to solve seven graph property tasks not solvable by the original GNN while improving performance across nine real-world benchmark datasets (Cotta et al., 2021). This is not GR in the hidden-graph-recovery sense, but it directly imports reconstruction theory into graph representation learning.

An asymmetry-based approach appears in “Reconstruction of graphs via asymmetry” (Farhadian, 2016). A proper induced subgraph that occurs exactly once is called an anchor, and a subgraph that is distinguished by the way it attaches to the rest of the graph is a connective anchor (Farhadian, 2016). The paper shows that if $E$00 is an anchor of order at most $E$01 and the corresponding shadow graph is reconstructible, then $E$02 is reconstructible; it further reduces the general problem to balanced or quasi-balanced graphs and graphs with an $E$03-vertex anchor, and uses this framework to show that trees are reconstructible (Farhadian, 2016).

A more abstract variant replaces graphs by operator-algebraic invariants. “Graph Reconstruction and Quantum Statistical Mechanics” studies finite multigraphs with minimal degree three (Cornelissen et al., 2012). The boundary operator algebra

$E$04

reconstructs only the first Betti number $E$05, in the precise sense that

$E$06

under the conditions of Theorem 1.5 (Cornelissen et al., 2012). By contrast, the quantum statistical mechanical system $E$07 with time evolution

$E$08

reconstructs the full graph: $E$09 (Cornelissen et al., 2012). This establishes a sharp separation between purely topological information and dynamics-enhanced rigidity.

5. Signal, embedding, and gradient-based reconstruction

In some settings the graph itself is reconstructed from signals observed on its vertices. “Graph reconstruction from the observation of diffused signals” assumes

$E$10

where $E$11 is a matrix of i.i.d. initial signals, $E$12 is the diffusion matrix, and $E$13 is known (Pasdeloup et al., 2016). Since

$E$14

the eigenvectors of the empirical covariance reveal those of $E$15, while the eigenvalue signs are recovered by solving linear constraints encoding zero diagonal, nonnegative off-diagonal entries, and normalization (Pasdeloup et al., 2016). The recovered matrix $E$16 is then thresholded to obtain $E$17 (Pasdeloup et al., 2016). The method is exact under the stated assumptions, including distinct absolute eigenvalues and sufficient sample size, but it is restricted to non-bipartite graphs and a fixed known $E$18 (Pasdeloup et al., 2016).

RESTORE uses reconstruction as an intrinsic diagnostic for graph embeddings rather than as a hidden-graph solver. Given embeddings $E$19, it reconstructs adjacency scores by

$E$20

and evaluates the result with $E$21, mean average precision,

$E$22

and semantic Euclidean distance (Yip et al., 2023). On the CommonSense Knowledge Graph, the deep-learning method SDNE is reported as best for 2-hop and 3-hop topological reconstruction, with mAP $E$23 and $E$24, while the factorization method HOPE is best for semantic preservation, with average Euclidean distance $E$25, $E$26, and $E$27 for 1-hop, 2-hop, and 3-hop reconstruction, respectively (Yip et al., 2023). The results are explicitly described as modest, leaving room for further research (Yip et al., 2023).

A privacy-oriented inversion problem is developed in “GRAIN: Exact Graph Reconstruction from Gradients” (Drencheva et al., 3 Mar 2025). In the honest-but-curious federated setting, the attacker observes client gradients for GCN or GAT models and reconstructs both graph structure and node features (Drencheva et al., 3 Mar 2025). The central primitive is the span check

$E$28

which filters candidate node features and subgraphs by compatibility with the low-rank structure of GNN gradients (Drencheva et al., 3 Mar 2025). The pipeline generates building blocks, glues overlapping subgraphs, and verifies candidates by gradient matching; the paper reports exact reconstruction for up to $E$29 of all graphs and states that this significantly outperforms the baseline, which achieves up to $E$30 correctly positioned nodes (Drencheva et al., 3 Mar 2025).

A related but distinct literature concerns reconstruction on a fixed graph rather than reconstruction of the graph. “Graph filtering for data reduction and reconstruction” formulates dimensionality reduction and reconstruction as graph filtering operations based on the adjacency matrix $E$31, optimizing mean-square error and using gradient descent in the graph spectral domain (Schizas, 2018). “Graph Signal Reconstruction Techniques for IoT Air Pollution Monitoring Platforms” compares Laplacian interpolation, low-pass GSP reconstruction, and kernel-based graph signal reconstruction on O$E$32, NO$E$33, and PM$E$34, reporting the superiority of kernel-based methods and showing that clustering can improve scalability (Ferrer-Cid et al., 2022). These works are part of the broader reconstruction vocabulary around graphs, but their target is signal recovery rather than GR in the sense of hidden-edge inference.

6. Separations, boundaries, and adjacent reconstruction programs

Several recurring separations structure the field. The random-versus-worst-case divide is pronounced: noisy random subgraphs suffice for reconstruction of $E$35 with $E$36-scale dependence modulated by $E$37, $E$38, or $E$39, yet arbitrary graphs can require $E$40 traces (McGregor et al., 2024). Oracle strength also matters: MIS queries are strictly more powerful than IS queries for bounded-degree GR (Konrad et al., 2024), while CC queries are stronger than separation oracles and, for some instances, stronger than MIS queries (Harviainen et al., 5 Sep 2025). In geometric settings, topological recovery and geometric recovery are distinct notions: discrete Morse methods may guarantee matching homotopy type and Hausdorff closeness without claiming stronger metrics such as Fréchet distance, which is explicitly noted as future work (Fasy et al., 2019).

The literature also shows that “reconstruction” is not limited to recovering graphs from partial graph data. In polytope theory, the direction is reversed: the graph can reconstruct a richer object. “Graphs, Skeleta and Reconstruction of Polytopes” surveys the Blind–Mani–Kalai theorem, according to which the graph of a simple polytope determines its face lattice, while broader classes may require more skeletal information or admit counterexamples (Bayer, 2017). The survey also records that zonotopes are determined by their graphs, but reconstruction from skeleta fails in general for Eulerian posets (Bayer, 2017). This suggests that GR methods and intuitions extend naturally to incidence structures beyond ordinary graphs, although the target object changes.

Open problems remain explicit in many subareas. The MIS-query work asks whether a randomized algorithm with $E$41 queries exists and whether adaptivity can help deterministic algorithms (Konrad et al., 2024). The CC-oracle work leaves the $E$42 gap between upper and lower bounds open and highlights deterministic constructions and query-size restrictions as further directions (Harviainen et al., 5 Sep 2025). The diffused-signal model does not address unknown $E$43, broader graph classes such as bipartite graphs, or robustness to model mismatch (Pasdeloup et al., 2016). The discrete Morse literature notes future work toward stronger geometric guarantees (Fasy et al., 2019). In polytope reconstruction, there is no known characterization of which graphs are graphs of $E$44-polytopes for $E$45 (Bayer, 2017).

Taken together, these results show that GR is not a single theorem or algorithmic template. It is a research area organized around how much structural information remains in a chosen observation model, which invariants are recoverable, and what computational resources are needed to extract them. Depending on the model, the answer may be exact edge-set recovery, recovery up to homotopy type, recovery of geometry, recovery of only the first Betti number, or intrinsic evaluation of how much information a learned representation preserves (Cornelissen et al., 2012).