Expander Pruning in Dynamic Graphs

• Expander pruning is a dynamic graph technique that preserves robust connectivity by selectively disabling vertices after adversarial deletions.
• The methodology employs local flow infeasibility, link-cut trees, and controlled recourse to maintain conductance and efficient routing in evolving graphs.
• Recent advances extend expander pruning to neural network sparsification, ensuring spectral gap preservation and improved post-training efficiency.

Expander pruning is a dynamic graph primitive for maintaining a robustly connected remainder of a graph after adversarial deletions by growing a monotone set of disabled or pruned vertices. In the standard formulation, one starts from a $\phi$-expander and seeks an online algorithm that preserves an induced subgraph of comparable expansion while keeping both the cumulative pruned volume and the per-update recourse small. Recent work has extended the objective beyond conductance preservation to the maintenance of routing structure, short paths, and explicit path access under deletions, while a separate line of neural-network research uses “expander pruning” to denote sparsification rules that preserve expander-like connectivity or spectral gap in layer graphs (Saranurak et al., 2018, Meierhans et al., 1 Apr 2025, Haeupler et al., 17 Jul 2025, Kalra et al., 2024, Bazarbachi et al., 13 Aug 2025).

1. Formal problem and terminology

For an undirected graph $G=(V,E)$, the volume of a set $S \subseteq V$ is

$\mathrm{vol}(S)=\sum_{v \in S}\deg(v),$

and the conductance of a cut is

$$\phi(S)=\frac{|E(S,\bar S)|}{\min\{\mathrm{vol}(S),\mathrm{vol}(\bar S)\}}, \qquad \bar S = V \setminus S.$$

The conductance of the graph is

$\phi(G)=\min_{\emptyset \neq S \subset V}\phi(S).$

A $\phi$-expander is a graph with $\phi(G)\ge \phi$. In dynamic settings, the expander pruning problem asks for a monotone pruning set $P \subseteq V$ such that, after each deletion, the induced graph on the remaining vertices continues to satisfy an expansion guarantee of roughly the original quality. The central performance measures are the growth of $P$, the recourse per update, and the update time (Meierhans et al., 1 Apr 2025).

This notion is closely related to, but distinct from, expander decomposition. Expander decomposition partitions a graph into clusters that are internally well connected and sparsely connected to one another; expander pruning is the repair mechanism that removes a small set of vertices so that a damaged cluster, or the whole graph, remains an expander after deletions (Saranurak et al., 2018).

A second formulation, developed for routing-sensitive settings, measures not only conductance but routability. In that setting, a demand $G=(V,E)$0 is unit if

$G=(V,E)$1

and a graph is an $G=(V,E)$2-length $G=(V,E)$3-congestion router for a node weighting $G=(V,E)$4 if every $G=(V,E)$5-respecting demand admits a length-$G=(V,E)$6 multicommodity flow of congestion at most $G=(V,E)$7. The same line distinguishes worst-case pruning ratio, meaning the maximum number of additional pruned vertices caused by a single deletion, from amortized pruning ratio, meaning total pruned vertices divided by the number of deletions (Haeupler et al., 17 Jul 2025).

2. Classical decremental expander pruning

A foundational formulation appears in “Expander Decomposition and Pruning: Faster, Stronger, and Simpler” (Saranurak et al., 2018). The paper works in an online deletion model on undirected, unweighted, simple graphs and uses the induced subgraph with self-loops $G=(V,E)$8, namely $G=(V,E)$9 augmented so that every $S \subseteq V$0 retains its degree in $S \subseteq V$1. This convention preserves volume while trimming vertices and is central to the dynamic guarantees.

Its main expander pruning theorem considers a $S \subseteq V$2-expander $S \subseteq V$3 with $S \subseteq V$4 edges and an edge deletion sequence of length $S \subseteq V$5. The algorithm deterministically maintains sets $S \subseteq V$6 such that $S \subseteq V$7 and $S \subseteq V$8, while for every prefix of deletions

$S \subseteq V$9

and the remainder $\mathrm{vol}(S)=\sum_{v \in S}\deg(v),$0 is a $\mathrm{vol}(S)=\sum_{v \in S}\deg(v),$1 expander. The total update time over $\mathrm{vol}(S)=\sum_{v \in S}\deg(v),$2 deletions is

$\mathrm{vol}(S)=\sum_{v \in S}\deg(v),$3

equivalently amortized $\mathrm{vol}(S)=\sum_{v \in S}\deg(v),$4 per deletion (Saranurak et al., 2018).

The same work provides a weighted extension. If $\mathrm{vol}(S)=\sum_{v \in S}\deg(v),$5 is a weighted $\mathrm{vol}(S)=\sum_{v \in S}\deg(v),$6-expander with total weight $\mathrm{vol}(S)=\sum_{v \in S}\deg(v),$7, and the total deleted weight is at most $\mathrm{vol}(S)=\sum_{v \in S}\deg(v),$8, then for cumulative deleted weight $\mathrm{vol}(S)=\sum_{v \in S}\deg(v),$9 the maintained pruning set satisfies

$$\phi(S)=\frac{|E(S,\bar S)|}{\min\{\mathrm{vol}(S),\mathrm{vol}(\bar S)\}}, \qquad \bar S = V \setminus S.$$0

and $$\phi(S)=\frac{|E(S,\bar S)|}{\min\{\mathrm{vol}(S),\mathrm{vol}(\bar S)\}}, \qquad \bar S = V \setminus S.$$1 remains a $$\phi(S)=\frac{|E(S,\bar S)|}{\min\{\mathrm{vol}(S),\mathrm{vol}(\bar S)\}}, \qquad \bar S = V \setminus S.$$2 expander, with total time $$\phi(S)=\frac{|E(S,\bar S)|}{\min\{\mathrm{vol}(S),\mathrm{vol}(\bar S)\}}, \qquad \bar S = V \setminus S.$$3 (Saranurak et al., 2018).

This line established the strong induced-expander guarantee as a practical dynamic objective. The remainder is not merely contained in some unknown expander; the induced graph itself is certified to be an expander. That distinction is important in downstream dynamic algorithms, because it permits local repair and recursion without an additional recovery step.

3. Worst-case recourse and update time

The next major step is “Expander Pruning with Polylogarithmic Worst-Case Recourse and Update Time” (Meierhans et al., 1 Apr 2025), which moves from amortized guarantees to worst-case guarantees per deletion. The paper considers an initially $$\phi(S)=\frac{|E(S,\bar S)|}{\min\{\mathrm{vol}(S),\mathrm{vol}(\bar S)\}}, \qquad \bar S = V \setminus S.$$4-expander $$\phi(S)=\frac{|E(S,\bar S)|}{\min\{\mathrm{vol}(S),\mathrm{vol}(\bar S)\}}, \qquad \bar S = V \setminus S.$$5 with $$\phi(S)=\frac{|E(S,\bar S)|}{\min\{\mathrm{vol}(S),\mathrm{vol}(\bar S)\}}, \qquad \bar S = V \setminus S.$$6 edges under an online sequence of up to $$\phi(S)=\frac{|E(S,\bar S)|}{\min\{\mathrm{vol}(S),\mathrm{vol}(\bar S)\}}, \qquad \bar S = V \setminus S.$$7 edge deletions and gives a deterministic algorithm that, for each deletion, runs in worst-case time $$\phi(S)=\frac{|E(S,\bar S)|}{\min\{\mathrm{vol}(S),\mathrm{vol}(\bar S)\}}, \qquad \bar S = V \setminus S.$$8, adds at most $$\phi(S)=\frac{|E(S,\bar S)|}{\min\{\mathrm{vol}(S),\mathrm{vol}(\bar S)\}}, \qquad \bar S = V \setminus S.$$9 vertices to the pruning set, and maintains that

$\phi(G)=\min_{\emptyset \neq S \subset V}\phi(S).$0

is an

$\phi(G)=\min_{\emptyset \neq S \subset V}\phi(S).$1

-expander at all times (Meierhans et al., 1 Apr 2025).

The paper also gives an amortized stepping-stone result with the same worst-case recourse bound but total processing time $\phi(G)=\min_{\emptyset \neq S \subset V}\phi(S).$2 over $\phi(G)=\min_{\emptyset \neq S \subset V}\phi(S).$3 deletions. The significance is not merely asymptotic. Earlier dynamic expander-pruning routines had good total guarantees but no non-trivial worst-case recourse bound; the 2025 result explicitly addresses that gap by growing the pruning set “slowly and explicitly” instead of rebuilding in the background and switching wholesale to a new solution (Meierhans et al., 1 Apr 2025).

A key technical convenience in the paper is expansion measured with respect to starting volume. If $\phi(G)=\min_{\emptyset \neq S \subset V}\phi(S).$4 is the initial graph, then one says that $\phi(G)=\min_{\emptyset \neq S \subset V}\phi(S).$5 is a $\phi(G)=\min_{\emptyset \neq S \subset V}\phi(S).$6-expander with respect to starting volume if every $\phi(G)=\min_{\emptyset \neq S \subset V}\phi(S).$7 with $\phi(G)=\min_{\emptyset \neq S \subset V}\phi(S).$8 satisfies

$\phi(G)=\min_{\emptyset \neq S \subset V}\phi(S).$9

This formulation is used to stabilize accounting as deletions reduce current degrees (Meierhans et al., 1 Apr 2025).

The paper also uses worst-case pruning as a building block for dynamic sparse graph primitives. Combined with recent techniques, it yields randomized algorithms that explicitly maintain an $\phi$0-edge graph $\phi$1 under fully dynamic updates, with at most $\phi$2 edge updates per adversarial update, supporting distance spanners, cut sparsifiers, and spectral sparsifiers with worst-case randomized update time $\phi$3 and high-probability guarantees against adaptive adversaries (Meierhans et al., 1 Apr 2025).

4. Self-pruning expanders and routing preservation

“Maintaining Routing Structures under Deletions via Self-Pruning” (Haeupler et al., 17 Jul 2025) reframes expander pruning around routing rather than pure conductance. The paper starts from the standard expander paradigm that a $\phi$4-expander routes any unit demand with congestion at most $\phi$5 over paths of length at most $\phi$6, then asks for a dynamic object that preserves these routing properties under deletions rather than merely preserving an expansion parameter.

Its core graph family is the $\phi$7-semi-hypercube, whose vertex set is $\phi$8. Clusters are prefixes $\phi$9, and between each pair of distinct child clusters $\phi(G)\ge \phi$0 and $\phi(G)\ge \phi$1 there is a perfect matching. The degree is $\phi(G)\ge \phi$2. On top of this base family, the paper defines noncritical and general $\phi(G)\ge \phi$3-semi-hypercubes by constraints on critical clusters, degenerate clusters, and marked vertices, with the invariant that each cluster has at most one $\phi(G)\ge \phi$4-critical child and that a critical child remains well connected to its siblings in aggregate (Haeupler et al., 17 Jul 2025).

For routing, the paper proves that if $\phi(G)\ge \phi$5 is a general $\phi(G)\ge \phi$6-semi-hypercube with $\phi(G)\ge \phi$7 and $\phi(G)\ge \phi$8, then there exists an oblivious routing $\phi(G)\ge \phi$9 with

$P \subseteq V$0

for unit demands $P \subseteq V$1, and path sampling takes expected time $P \subseteq V$2. A matching lower bound shows that for $P \subseteq V$3 and $P \subseteq V$4, some $P \subseteq V$5-semi-hypercubes require congestion at least $P \subseteq V$6 for some unit demand, so the exponential-in-$P \subseteq V$7 congestion factor is necessary up to constants (Haeupler et al., 17 Jul 2025).

For self-pruning, the paper gives a deterministic routine, EfficientSelfPruningSHC, initialized on a $P \subseteq V$8-semi-hypercube $P \subseteq V$9 with $P$0 and $P$1. Under vertex deletions Delete$P$2, it maintains an induced $P$3-semi-hypercube $P$4 while pruning additional vertices $P$5 with worst-case ratio

$P$6

per deletion, worst-case per-update work

$P$7

and depth $P$8 (Haeupler et al., 17 Jul 2025).

The same paper also maintains explicit deterministic routings. For a $P$9-semi-hypercube with $G=(V,E)$00 and an integral load-$G=(V,E)$01 demand $G=(V,E)$02, it maintains a routing of congestion at most $G=(V,E)$03 and length at most $G=(V,E)$04, with per-update recourse

$G=(V,E)$05

and work

$G=(V,E)$06

This explicit path maintenance distinguishes the routing-sensitive setting from earlier expander-pruning results, which preserve expansion but not direct path access (Haeupler et al., 17 Jul 2025).

Through embeddings, the semi-hypercube machinery transfers to arbitrary graphs. If a $G=(V,E)$07-semi-hypercube $G=(V,E)$08 is embedded into a graph $G=(V,E)$09 with path-length $G=(V,E)$10 and congestion $G=(V,E)$11, then under adversarial edge deletions one can deterministically maintain a vertex set $G=(V,E)$12 and an oblivious routing on $G=(V,E)$13 with congestion at most $G=(V,E)$14, dilation $G=(V,E)$15, path-sampling time $G=(V,E)$16, and worst-case pruning ratio at most

$G=(V,E)$17

vertices per deletion. For constant-degree $G=(V,E)$18-expanders, the resulting corollary gives worst-case pruning ratio and per-update work roughly $G=(V,E)$19 while maintaining routing length $G=(V,E)$20 (Haeupler et al., 17 Jul 2025).

5. Algorithmic techniques and proof architecture

The 2018 algorithm is based on local flow infeasibility and trimming. Given a candidate set $G=(V,E)$21, it sets source mass on boundary edges, sink capacity $G=(V,E)$22 at vertices, and edge capacity $G=(V,E)$23 on edges of $G=(V,E)$24. A push-relabel-style primitive, Unit-Flow, runs with height

$G=(V,E)$25

If the flow is feasible, $G=(V,E)$26 is certified as a $G=(V,E)$27 expander; if not, Unit-Flow returns a level cut $G=(V,E)$28 that is trimmed from $G=(V,E)$29. The amortized analysis shows that total work can be charged to total source mass, yielding the $G=(V,E)$30 bound over $G=(V,E)$31 deletions (Saranurak et al., 2018).

The 2025 worst-case algorithm retains the flow-certificate viewpoint but de-amortizes it by layering certificates across batches and repairing them incrementally. One central lemma states that if a flow in

$G=(V,E)$32

routes source

$G=(V,E)$33

to sinks $G=(V,E)$34 with edge capacities $G=(V,E)$35, then $G=(V,E)$36 is a $G=(V,E)$37-expander. To compose certificates over levels, the paper forms

$G=(V,E)$38

where $G=(V,E)$39. Link-cut trees support worst-case backtracking of certificate flow when deleted edges invalidate old routes, and vertices are added to the global pruning set only after a threshold counter exceeds $G=(V,E)$40 at a level-$G=(V,E)$41 certificate (Meierhans et al., 1 Apr 2025).

The self-pruning framework of 2025 uses a different architecture. Each cluster $G=(V,E)$42 stores a target child $G=(V,E)$43, a non-target set $G=(V,E)$44, shadow mark sets $G=(V,E)$45, mark-source information $G=(V,E)$46, size$G=(V,E)$47, a cycle length $G=(V,E)$48, and a multiplier

$G=(V,E)$49

The algorithm alternates between Retarget, Trim, and Delete. Retarget selects a new target child either by smallest nonempty size or by largest number of inherited marks, depending on the alternation schedule. The key invariant is “$G=(V,E)$50 critical child per cluster” together with a bounded fraction of marked vertices in the target child. Harmonic bounds show that with the stated $G=(V,E)$51, no non-target child becomes $G=(V,E)$52-critical and marks remain within a $G=(V,E)$53 fraction even through degenerate cascades. For routing, greedy paths are supplemented by escape moves that reduce isolation until the path reaches the root’s home cluster; for explicit routing, bucketed load balancers maintain per-matching balance within $G=(V,E)$54 of average (Haeupler et al., 17 Jul 2025).

Taken together, these techniques show three distinct proof paradigms. Classical expander pruning uses local flow infeasibility to identify cuts to trim. Worst-case pruning uses precomputed and repairable certificates to control recourse at every deletion. Self-pruning replaces generic repair by a graph family whose combinatorial structure is itself dynamically maintainable.

6. Neural-network uses, misconceptions, and open directions

In neural-network sparsification, “expander pruning” refers to a different object. The goal is not to maintain a pruning set under adversarial graph deletions, but to prune weights while preserving expander-like connectivity or spectral expansion of the layer graph. In “Graph Expansion in Pruned Recurrent Neural Network Layers Preserve Performance” (Kalra et al., 2024), recurrent layers are modeled as graphs with adjacency built either from the binary mask of nonzeros or from $G=(V,E)$55. The paper monitors two surrogates,

$G=(V,E)$56

and prunes $G=(V,E)$57 and $G=(V,E)$58 by iterative magnitude pruning while keeping $G=(V,E)$59 dense. Its empirical claim is that test accuracy is largely preserved while $G=(V,E)$60 and $G=(V,E)$61 remain positive, and that accuracy degradation begins at the first zero crossing, especially on noisy inputs and long sequences (Kalra et al., 2024).

A structured post-training variant appears in “EGGS-PTP: An Expander-Graph Guided Structured Post-training Pruning Method for LLMs” (Bazarbachi et al., 13 Aug 2025). There, a linear layer is modeled as a bipartite graph between input and output channels, and $G=(V,E)$62 sparsity is imposed row-wise over contiguous groups of $G=(V,E)$63 columns:

$G=(V,E)$64

The method combines channel permutation, Row Relative Importance, Relative Importance and Activation, and a connectivity-aware diagonal selection inside $G=(V,E)$65 blocks. The paper proves that the pruned bipartite graph is a two-sided $G=(V,E)$66-expander for some $G=(V,E)$67 with $G=(V,E)$68 and $G=(V,E)$69, and reports, for example, a runtime reduction on NVIDIA A40 GPUs for LLaMA2-13B from $G=(V,E)$70 dense to $G=(V,E)$71 under EGGS-PTP $G=(V,E)$72, a $G=(V,E)$73 speedup (Bazarbachi et al., 13 Aug 2025).

A common misconception is therefore that expander pruning always means preserving a conductance certificate under deletions. In graph algorithms, it is a decremental maintenance problem on an initially expanding graph. In neural-network pruning, it is a sparsification heuristic or structural rule guided by expansion or spectral-gap surrogates. The shared intuition is robustness of sparse connectivity, but the formal models, guarantees, and adversarial assumptions are different.

Several open directions remain explicit in the recent literature. In routing-sensitive pruning, achieving both worst-case pruning and near-original $G=(V,E)$74 while preserving short path access remains open, as do distributed implementations and extensions beyond undirected, uncapacitated graphs (Haeupler et al., 17 Jul 2025). In worst-case expander pruning for conductance, improving the retained expansion from $G=(V,E)$75 toward $G=(V,E)$76, and sharpening the $G=(V,E)$77 dependence or proving matching lower bounds, remain open problems (Meierhans et al., 1 Apr 2025). These questions indicate that expander pruning has shifted from a repair subroutine inside dynamic decompositions to a broader framework for maintaining sparse, high-connectivity structure under stringent update and path-access constraints.