Effective resistance treats a brain's wiring as an electrical circuit. White-matter connection strength plays the role of conductance, and the "distance" between two regions is the voltage drop when unit current flows between them. Unlike shortest-path length, it counts every parallel route. It has become a useful input to brain-network models: AFR-Net uses it inside a differentiable circuit solver for psychiatric and neurological diagnosis. These models always compute it on a parcellated atlas, meaning each subject's scan is warped into a standard template and chopped into 100–400 regions. The observation behind this note is simple. Building a region graph by adding up voxel-to-voxel connections between regions is electrically the same as short-circuiting every voxel inside a region into one node. By a classical monotonicity fact (Rayleigh's law: shorting wires can only lower resistance), the region-level resistance therefore systematically underestimates the resistance of the underlying tissue. The size of that underestimate depends on how big and how internally connected each parcel is. Parcel size is exactly what varies across atlas resolutions and, after spatial normalization, across subjects.
We propose to measure this "shorting bias," correct it, and test whether the correction makes resistance-based brain features stable across atlas resolution without losing diagnostic accuracy. Recursive spectral mapping plays two roles. First, it supplies graphs where the bias is known exactly: on self-similar graphs, resistance between old nodes rescales by closed-form factors under refinement. Second, it supplies the nested block-matrix recursion we use to compute the corrected quantity at every level of the atlas hierarchy in one pass.
Core Claim
Let $L$ be the Laplacian of a subject's fine-grained (voxel- or sub-parcel-level) structural graph and $L^\dagger$ its pseudoinverse. For two atlas regions $A$ and $B$, let $p_A$ and $p_B$ be uniform probability vectors over their voxels. We define the distributed resistance
$$\Omega(A,B) = (p_A - p_B)^\top L^\dagger (p_A - p_B), \qquad \Omega^{\mathrm{short}}(A,B) \le \Omega(A,B),$$
where $\Omega^{\mathrm{short}}$ is the ordinary resistance on the region graph whose edge weights are summed voxel connections. The inequality holds because shorted resistance is the minimum flow energy over all ways of spreading current across $A$ and $B$ (Thomson's principle), and uniform spreading is one such way.
The claim has two parts. First, the gap $\Omega - \Omega^{\mathrm{short}}$ is large and varies across regions, tracking parcel volume and internal connectivity rather than between-region wiring. Second, $\Omega$ computed at different atlas resolutions and mapped to common coarse regions agrees far better than $\Omega^{\mathrm{short}}$ does, and this consistency carries over to downstream models.
What is new is not faster resistance computation. It is the identification of a resolution artifact baked into how atlas-level resistance is currently defined, together with a principled replacement.
Research Program
The first component is theory on graphs where the answer is exact. The recursive-spectral literature gives closed-form resistance laws under uniform refinement:
- On $q$-subdivision graphs, where each edge becomes $q$ parallel length-2 paths, resistance between original nodes scales as $2/q$, as shown by Zeng and Zhang.
- On growing simplicial networks, precursor-node resistance scales by $2/(q+2)$ each generation, and average resistance converges to a closed-form constant, per Zhu et al..
- Under iterated triangulation, normalized-Laplacian eigenvalues halve, per Xie, Zhang and Comellas.
We treat each generation-$(t-1)$ node together with the nodes grown around it as a "parcel" of the generation-$t$ graph. We then derive $\Omega^{\mathrm{short}}$ and $\Omega$ in closed form, using the nested block structure of Qi et al. and Vieta-style coefficient sums as in Julaiti et al.. These papers supply the method. None of them considers set-to-set resistance or parcel-level aggregation. The result is an exact prediction of the shorting gap as a function of $q$ and generation. It serves as a unit test for any estimator we build.
We then drop uniformity, which is the real open problem these papers leave. In an atlas, different regions are refined by different amounts. We will derive how the gap decomposes into per-region terms, each close to an intra-region "spread resistance" $p_A^\top L^\dagger p_A$ minus its shorted counterpart. We will also bound the cross terms. If the cross terms prove non-negligible, we report the gap only as an empirical quantity rather than claiming a clean decomposition.
The second component is computation. The Allen CCF ontology and nested cortical parcellations form a tree of regions. We compute $\Omega$ for all levels by recursive Schur complements (eliminating a region's interior onto its boundary) up this tree. This is the empirical analogue of the block nesting used in the recursive-spectral papers, and it is related in spirit to the effective-resistance coarsening of HyperEF. HyperEF coarsens to preserve spectra of abstract netlists. We instead tie coarsening to a fixed anatomical hierarchy and track a set-level quantity. Speed is a secondary benefit, and we compare it honestly against sparse solvers.
The third component is empirical. We use PPMI and ABCD diffusion and functional MRI as used by AFR-Net, with streamline-based structural graphs built at a fine level and at Schaefer-style 100/200/400 resolutions.
Baselines
The naive reference point is AFR-Net's own pipeline: shorted resistance on the parcel graph, which is the quantity we claim is biased. Its reported results, from AFR-Net, are AUC 90.38% on PPMI (next best baseline 71.88%) and F1 70.62% on ABCD-OCD (prior baselines 64.13%).
We will rerun AFR-Net's released configuration on identical splits, varying only the resistance input. Each variant gets the same number of seeds, the same hyperparameter search budget, and the same parcellation. For the downstream comparison, "beating" AFR-Net means one of two things:
- matching its AUC and F1 within seed variance while substantially improving cross-resolution transfer, i.e., training at 200 parcels and testing at 400 parcels mapped back; or
- exceeding it outright.
A shortest-path-distance variant is included as a floor. The argument of Velingker et al. that resistance captures path redundancy that hop distance misses suggests it should be clearly worse.
The strongest competitor on the resolution-stability question is global normalization: dividing each subject's parcel resistances by their mean or by the Kirchhoff index. This removes any uniform rescaling for free. If refinement effects were uniform, as in the exact recursive laws, it would erase the artifact entirely.
The key ablation replaces uniform injection $p_A$ with injection concentrated at a single representative voxel per region. This separates the benefit of avoiding shorting from the choice of spreading current within a region.
On the mouse side, WiCAT's zero-shot decoding numbers from WiCAT (R² 0.327 on Musall, 0.184 on Kondo) serve only as a reference for a secondary exploratory test of resistance-defined patches. We do not treat that test as a primary claim.
Why not just…?
Why not just compute exact effective resistance on the voxel graph with a sparse solver? This is the right way to compute $\Omega$, and we use it as ground truth. It answers the compute question, not the definitional one. Solving on the fine graph still leaves the question of what "resistance between region A and B" means. The usual answer, the parcel graph, is the shorted quantity. AFR-Net's stated scaling limit (100–400 parcels) is therefore not the central problem. A 10⁴–10⁵-node sparse solve is routine, and Velingker et al. report resistances on a 169k-node graph in about an hour on a laptop.
Why not just globally rescale parcel resistances? This is the real threat. On the exactly solvable recursive graphs, uniform refinement produces a single global factor ($2/q$, $2/(q+2)$), so rescaling would suffice there. The proposal survives only if real atlases show region-dependent gaps. Experiment 2 below measures the residual non-uniformity after global rescaling. If that residual is small, we will report that global normalization suffices and the corrected metric is unnecessary.
Why not just use JL sketching of resistance embeddings? Sketching, as in Velingker et al., approximates node-to-node resistance within $(1\pm\epsilon)$. Applied to the parcel graph, it approximates the biased quantity. Applied to the fine graph, it could approximate $\Omega$ through the embedding of $p_A - p_B$, which is compatible with our proposal. We include it as an alternative way to compute $\Omega$, not as a competitor to the definition.
Proposed Experiments
Experiment 1 checks the theory. On $q$-subdivision, simplicial, and triangulated graphs up to generation 8, we verify that the Schur-recursion estimate of $\Omega$ and $\Omega^{\mathrm{short}}$ matches the closed forms to machine precision. We then randomize which edges are refined and measure how the gap departs from the uniform law.
Experiment 2 is the decisive empirical check. On PPMI and ABCD, we compute three quantities: the fraction of variance in $\Omega - \Omega^{\mathrm{short}}$ explained by parcel voxel count and intra-parcel density; the residual non-uniformity after global rescaling; and the rank correlation of region-pair resistances across 100/200/400-parcel resolutions mapped to common coarse regions. If the corrected metric does not raise cross-resolution agreement beyond what global rescaling achieves, the core claim fails.
Experiment 3 covers downstream models and circuit stability. We rerun AFR-Net with shorted, globally rescaled, single-representative, and distributed resistance under the fair-comparison protocol above. We report AUC, F1, cross-resolution transfer, and the overlap across resolutions of the top circuits the model identifies. The OCD "check-and-doubt" circuit is a named test case.
Experiment 4 is a runtime comparison: the hierarchical recursion versus a sparse Cholesky solve at each level, on the same hardware. It is reported but not used to justify the proposal.
References
- Uncovering Latent Communication Patterns in Brain Networks via Adaptive Flow Routing
- Spectra, Hitting Times and Resistance Distances of q- Subdivision Graphs
- Resistance Distances In Simplicial Networks
- On the spectrum of the normalized Laplacian of iterated triangulations of graphs
- Structural and spectral properties of a family of deterministic recursive trees: rigorous solutions
- Eigenvalues of normalized Laplacian matrices of fractal trees and dendrimers: Analytical results and applications
- HyperEF: Spectral Hypergraph Coarsening by Effective-Resistance Clustering
- Affinity-Aware Graph Networks
- Cross-Subject Modeling for Widefield Calcium Imaging via Atlas-Aligned Spatiotemporal Tokenization