Vector search engines answer "find the items most similar to this one" by walking a proximity graph, a network in which each stored vector links to a few nearby vectors and a query hops greedily toward its target. These graphs work well when they are built once and left alone. They degrade when data is deleted: removing a node cuts the routes that passed through it, and queries begin to dead-end. Current systems sit at two poor extremes. They either repair every deletion eagerly with an expensive global search, or they keep deleted nodes as "ghost" waypoints and clean up in large batches, wasting memory and letting recall collapse under heavy deletion. We propose a third option. A cheap physical quantity, the electrical effective resistance of the graph around each deletion, measures how much routing damage that deletion caused. A leaky integrate-and-fire accumulator borrowed from spiking neurons then decides when a region has collected enough damage to be worth repairing.
The intuition comes from analog computing. If you treat graph edges as resistors, effective resistance between two nodes is small when they are joined by many short paths and large when they are joined by few long ones. It is what a resistive network computes for free when it settles to minimum power dissipation. Effective resistance therefore rises sharply when a deletion removes a bottleneck or a long-range shortcut, and barely changes when a deletion removes one of many redundant routes. A node's lost in-degree, by contrast, looks the same in both cases. If this signal tracks recall loss, repair work can go only where routing broke. This should give fresher indices at a fraction of the repair cost of eager methods, without the recall collapse of lazy ones.
Core Claim
The claim is that local increase in effective resistance, accumulated per graph region by a leaky threshold rule, is a better trigger for proximity-graph repair than either always repairing or repairing on structural counts such as degree loss. "Better" means higher recall at a matched repair budget, or the same recall at lower budget, under realistic churn. The work would be new in two ways. Physical energy minimization serves here as a sensor of topological damage, not as the model being computed. Spiking dynamics serve as a scheduler for index maintenance, not as a feature extractor.
Concretely, give each edge $(i,j)$ a conductance $g_{ij}$ that decreases with vector distance. The effective resistance between nodes $a$ and $b$ is
$$R_{ab} = (e_a - e_b)^\top L^{+} (e_a - e_b),$$
where $L$ is the weighted graph Laplacian built from the $g_{ij}$, $L^{+}$ its pseudoinverse, and $e_a$ the indicator vector of node $a$. Equivalently, $R_{ab}$ is the voltage that appears when one unit of current is pushed from $a$ to $b$ and the network relaxes to minimum dissipated power. When node $v$ is deleted, we compute the damage $D(v)$: the mean increase in $R_{ab}$ over pairs $a,b$ of $v$'s former neighbors. The computation is restricted to a $k$-hop ball whose boundary is held at a fixed voltage, so it costs a few local relaxation sweeps, in the style of the local push updates of DynamicPPE, and never a global solve. Each region $r$ of the index (a partition cell) keeps an accumulator
$$V_r \leftarrow \lambda V_r + \sum_{v \in r} D(v) - \sum_{u \in r} H(u),$$
where $\lambda<1$ is a leak, $D(v)$ sums over deletions in $r$ since the last step, and $H(u)$ is the resistance drop caused by insertions in $r$. Insertions near damage therefore partially heal it, and the accumulator gives them credit. When $V_r$ crosses a threshold $\theta$, the region fires. Its orphaned in-neighbors are re-linked from candidates in their two-hop neighborhood, and only nodes whose post-repair resistance remains high are escalated to a full global re-insertion. Then $V_r$ resets.
Relation to Prior Work
Three lines of work motivate the gap. On the maintenance side, IPGM proves that only the incoming neighbors of a deleted vertex need repair, and it re-inserts them through full global search. That is correct but eager: it pays the full price for every deletion, whether or not routing was harmed. DiskANN's FreshVamana does the opposite. It keeps tombstones as waypoints and consolidates in batches. LSM-VEC reports that under delete-heavy workloads this batch-and-tombstone approach lets DiskANN's recall fall to 61% from fragmentation, while LSM-VEC keeps 77.4%. Neither paper asks which deletions actually hurt. We keep IPGM's locality result as the repair primitive and add a decision about when and where to apply it.
The physical-computing papers supply the measurement, and they also mark a boundary. Resistive networks for vision such as Computing Motion Using Resistive Networks show that minimum co-content relaxation solves Laplacian-type problems directly. NP-GL, EADS and InstaTrain map graph learning onto conductances with large reported speed and energy gains, but all three assume a fixed topology and cannot add or remove nodes. We do not try to make the substrate change its wiring. We use the same relaxation physics to quantify the effect of a topology change that happens in ordinary memory.
The neuromorphic papers supply the scheduling rule and a hint of hardware fit. The spiking dynamic-graph models SpikeNet, Dy-SIGN and ChronoSpike use leaky integrate-and-fire neurons to learn embeddings over graph snapshots. None of them maintains an index. MorphSNN rewires a network's internal feature graph with Dirichlet-energy diffusion. Its graph is a latent representation, not an external data structure, and it offers no connectivity guarantees. Pohoiki Springs shows neuromorphic nearest-neighbor search with constant-time insertion. It uses brute-force dot products, though, with no graph and no deletion support. Our accumulator is the same integrate-to-threshold primitive, applied to maintenance events. Any hardware mapping is a secondary projection, not the central claim.
Research Program
The first stage is a diagnostic that can kill the idea cheaply. We build Vamana and HNSW indices on SIFT1M, GloVe, NYTimes and GIST, the datasets IPGM uses. We then apply random deletions and cluster deletions, the latter removing a contiguous region of vector space, as in IPGM's evaluation. After each deletion batch we measure, for every partition region, the recall loss of held-out queries whose ground-truth neighbors lie in that region. We also measure four candidate damage signals: lost in-degree, increase in shortest-hop distance among former neighbors, local effective resistance increase, and the number of orphaned in-neighbors. The question is which signal best ranks regions by recall loss, scored by Spearman correlation and by the precision of the top-$k$ ranked regions.
The second stage builds the trigger-and-repair loop described above on one shared in-memory codebase, so that every policy uses identical search, pruning (Vamana's $\alpha$ fixed at 1.2) and re-linking code. Only the decision of what to repair and when differs between policies. We tune $\lambda$, $\theta$ and the ball radius $k$ on SIFT1M alone and transfer them unchanged to the other datasets.
The third stage scales to a 100M-vector subset of SIFT1B under the balanced and delete-heavy workloads used by LSM-VEC. It also includes a burst test: deletions arrive in short spikes, and update tail latency matters as much as recall. A short final analysis estimates what the damage computation would cost on a resistive substrate of the NP-GL kind, where the local relaxation is a physical settling step. This estimate is clearly labeled as a projection.
Baselines
The naive reference point is doing nothing: deleted nodes are removed and their edges simply dropped, with no repair until a full periodic rebuild. Rebuilding from scratch after every batch serves as the opposite reference. It is the recall ceiling for a fresh graph and the cost ceiling for maintenance. Both run on the same codebase and hardware.
The strongest existing method on this question is IPGM's eager global reconnect, which repairs every orphaned in-neighbor of every deletion by full re-insertion. IPGM reports that it beats periodic rebuilding by up to 18.8% in query efficiency on SIFT and keeps recall high across 10 update batches. This is the right competitor because it already exploits the locality of deletion damage and is the most accurate published repair. We run it with the same search parameters and the same 10-batch and 50-cycle churn schedules. To beat it, our method must reach recall within 0.5 points of IPGM at matched query throughput while using at most half of IPGM's distance computations for repair. Alternatively, it must achieve higher recall when both are capped at the same repair budget.
The second strong baseline is FreshVamana's tombstone-and-consolidate policy from DiskANN, run with its consolidation interval tuned on SIFT1M under the same budget as our hyperparameters. DiskANN reports above 95% recall across 50 churn cycles on its own workloads. LSM-VEC reports 82.0% on balanced and 61% on delete-heavy 100M SIFT1B workloads for DiskANN, against 88.4% and 77.4% for LSM-VEC. On the 100M runs we treat these as published reference points. Our method's head-to-head comparison is with FreshVamana as reimplemented in our codebase, because LSM-VEC's gains come partly from its on-disk storage design, which is orthogonal to the repair trigger.
The key ablations isolate the two ingredients. The degree-triggered ablation keeps the accumulator but replaces resistance increase with lost in-degree. The hop-distance ablation uses increase in shortest-path length instead. The no-accumulator ablation repairs immediately whenever a single deletion's resistance increase passes the threshold, with no leak and no credit for insertions. If resistance does not beat degree and hop distance at matched budget, the physical measurement is not earning its place.
Why not just…?
Why not just trigger repair on in-degree loss? This is the most serious threat, because degree is free to compute and correlates with damage in uniform random deletion. Our argument is that degree cannot tell a redundant edge from a long-range shortcut. Fragmentation of the kind LSM-VEC reports for DiskANN under delete-heavy load is a connectivity failure that per-node counts miss. That argument is a hypothesis, not evidence. The stage-one correlation study and the degree-triggered ablation settle it directly. If degree matches resistance within noise on both random and cluster deletions, we will report that and recommend the degree trigger.
Why not just repair everything eagerly, as IPGM does? Eager repair is the accuracy gold standard, and when deletions are rare its cost is acceptable. The case against it is bursts. When thousands of deletions arrive at once, each one triggers a global search, and update tail latency explodes. Our burst experiment measures exactly this. If IPGM's cost stays tolerable at realistic burst rates, the advantage of selective triggering shrinks to a memory and throughput argument, and we will say so.
Why not just monitor recall with probe queries and repair where they fail? Probe queries measure the target quantity directly, which makes them attractive. However, a failing probe only signals that some region along its route is broken, so locating the damage requires many probes per region, and coverage scales with the number of regions. A resistance measurement is local to the deletion that caused it. We include a probe-triggered policy at matched compute (probe distance computations counted against the repair budget) to test whether this localization advantage is real.
Why not just rebuild periodically? Periodic rebuilding is simple and always yields a fresh graph. IPGM already shows incremental repair beating it by up to 18.8% in query efficiency on SIFT, and full reconstruction does not fit continuous churn at the 100M scale. We keep rebuilding only as the ceiling reference.
Proposed Experiments and Decisive Checks
Every experiment reports recall@10 against queries per second, distance computations spent on repair per deleted node, update throughput, and 99th-percentile update latency. The workloads are random deletion, cluster deletion and mixed insert/delete churn on SIFT1M, GloVe, NYTimes and GIST over 10 batches and 50 cycles, plus balanced, delete-heavy and burst workloads on 100M SIFT1B vectors.
Three results decide the project. First, in stage one, local resistance increase must rank damaged regions better than lost in-degree, with a clearly higher Spearman correlation against per-region recall loss on cluster deletions. If it does not, we stop before building the system. Second, on the repair-budget-versus-recall curve, the resistance-triggered policy must dominate the degree-triggered, hop-distance and probe-triggered policies. It must also reach IPGM-level recall with substantially fewer repair computations. Third, on the delete-heavy 100M workload, it must clearly avoid the fragmentation collapse that LSM-VEC reports for DiskANN, holding recall well above the 61% published figure while staying within FreshVamana's memory footprint. We expect gains to be largest under cluster deletion, where damage concentrates on bottlenecks, and smallest under uniform random deletion. Seeing that pattern would itself support the claim that what matters is where routing breaks, not how many edges disappear.
References
- Subset Node Representation Learning over Large Dynamic Graphs
- Proximity Graph Maintenance for Fast Online Nearest Neighbor Search
- The DiskANN library: Graph-Based Indices for Fast, Fresh and Filtered Vector Search
- LSM-VEC: A Large-Scale Disk-Based System for Dynamic Vector Search
- Computing Motion Using Resistive Networks
- Extending Power of Nature from Binary to Real-Valued Graph Learning in Real World
- An Expressive and Self-Adaptive Dynamical System for Efficient Function Learning
- InstaTrain: Adaptive Training via Ultra-Fast Natural Annealing within Dynamical Systems
- Scaling Up Dynamic Graph Representation Learning via Spiking Neural Networks
- Dynamic Spiking Graph Neural Networks
- ChronoSpike: An Adaptive Spiking Graph Neural Network for Dynamic Graphs
- MorphSNN: Adaptive Graph Diffusion and Structural Plasticity for Spiking Neural Networks
- Neuromorphic Nearest Neighbor Search Using Intel's Pohoiki Springs