Inherited Eigenmodes for Cross-Subject Cortical Surface Alignment

Treating a brain surface as a graph and using its "vibration modes" (the eigenvectors of its graph Laplacian, which act like a Fourier basis for that shape) is a fast way to put many brains into a shared coordinate system. The approach breaks down because each brain's modes are computed independently. Their signs flip, their order swaps, and nearly equal modes mix together, so mode 17 on one subject need not correspond to mode 17 on another. The current fix aligns only the first few modes with iterative point matching and discards the rest. We propose a different route. Each subject's high-frequency modes are built from coarse modes through the same algebra that mathematicians use to compute spectra of self-similar graphs exactly. Every fine mode then carries a lineage, a parent mode at a coarse level plus a branch label, and that lineage gives it an identity that is comparable across subjects.

In plain terms, the method shrinks each brain mesh in steps to a few hundred vertices. It matches the handful of very smooth modes at that small size, where matching is easy and stable. It then grows the modes back to full resolution using an exact formula that ties each fine mode to its coarse parent. Sign, order, and identity are fixed once, at the coarse level, and passed down to the fine level. No iterative matching happens at full resolution, so many more spectral coordinates become usable across subjects.

Core Claim

The tool that makes this possible is one identity from spectral decimation, the field that computes eigenvalues of recursively built graphs from those of the previous generation. Split a graph's vertices into a kept set (block 1) and an eliminated set (block 2). The Laplacian $L$ then has blocks $L_{11}, L_{12}, L_{21}, L_{22}$, and we define the frequency-dependent Schur complement

$$S(\lambda) = (L_{11} - \lambda I) - L_{12}\,(L_{22} - \lambda I)^{-1} L_{21},$$

where $\lambda$ is a candidate eigenvalue and $I$ the identity. Whenever $L_{22}-\lambda I$ is invertible, $\lambda$ is an eigenvalue of the fine graph exactly when $S(\lambda)$ has a null vector $x$. The full fine eigenvector is then $x$ on the kept vertices and $-(L_{22}-\lambda I)^{-1}L_{21}x$ on the eliminated ones. At $\lambda=0$, $S$ is exactly the Kron reduction used by Node Decimation Pooling, the operation that coarsens a graph while preserving effective resistances.

On exactly self-similar graphs, the spectral-decimation literature shows that $S(\lambda)$ factors as $\varphi(\lambda)\,(L_c - R(\lambda) I)$. Here $L_c$ is the coarse Laplacian, $R$ is a low-degree map such as $R(x)=2x^2-1$, and $\varphi$ is a scalar. Fine eigenvalues are therefore preimages of coarse ones under $R$ (The normalized Laplacian spectrum of subdivisions of a graph; Xie et al.). In this exact setting, each fine mode is literally a parent mode plus a choice of branch, such as the two square-root branches $f_1, f_2$.

Cortical meshes are not self-similar, so the factorization holds only approximately. Our claim is that at the low and middle frequencies used for alignment, the approximation is good enough for parent–child lineage to be well defined and stable across subjects. If that holds, lineage-labelled eigenmodes give a canonical spectral coordinate system that beats per-subject iterative point matching, both in how many modes can be aligned and in downstream parcellation accuracy. The claim fails cleanly if the "self-similarity defect" of real cortex is large. Measuring that defect is the first experiment.

Where This Sits

Graph Convolutions on Spectral Embeddings for Cortical Surface Parcellation motivates the proposal and is the method to beat. It computes normalized-Laplacian embeddings of each Mindboggle cortex (102k–185k vertices). It aligns them to a reference with Iterative Closest Point (ICP, repeated nearest-neighbour matching followed by a best-fit rotation) plus Procrustes analysis, then runs a graph CNN. The paper names eigenvector sign flips, reordering, and frequency mixing as the obstacle. It handles them empirically on a few modes and computes every subject's spectrum from scratch. We replace the fine-level ICP step with inherited labels and keep everything else fixed.

Spectral Demons also motivates the work. It shows that low Laplacian eigenmodes give global correspondences that survive large deformations, cutting transformation error by 73% relative to Log-Demons. Like the parcellation paper, it uses a generic eigensolver and treats each image independently.

Two lines of work supply the method. The decimation papers (Qi et al.; Zhou et al.; iterated triangulations) supply the parent–child structure. They also supply one fact we rely on: some fine eigenvalues have no parent. These are the "exceptional" eigenvalues, fixed at 1 and 3/2 under iterated triangulation, with multiplicities set by vertex and edge counts. Our lineage must therefore distinguish inherited modes from newborn, localized ones rather than pretend every mode has an ancestor. NDP supplies the coarsening recipe. It selects an eliminated set via a spectral approximation of MAXCUT, which is close to an independent set. That choice makes $L_{22}$ nearly diagonal, so the inverse in $S(\lambda)$ is cheap and local. NDP was only ever used for pooling in graph classification. It never tracked eigenvectors and was never applied to surfaces.

Two papers constrain the design. SUGAR and MS-SiT both use recursively subdivided icosahedral spheres. On a resampled icosphere, however, every subject has the identical graph. Decimating that graph's combinatorial spectrum carries no subject information, which rules out the "obvious" intersection idea of analytically decimating the icosphere. Subject information lives in geometry-dependent edge weights on the native mesh, and that is where our hierarchy is built.

Research Program

For each subject we build an $L$-level hierarchy on the native white or pial mesh. Edges are weighted by inverse distance, as in the parcellation paper. At each level we eliminate a MAXCUT-selected near-independent set, Kron-reduce, and sparsify following NDP, until roughly 300 vertices remain.

At the coarsest level we compute the lowest 6–10 eigenvectors exactly, which is trivial at this size. We align these to a template subject by Procrustes over sign and rotation, applied within near-degenerate clusters as subspace alignment. These smooth modes encode the large anterior–posterior and dorsal–ventral axes of the hemisphere and are well separated.

We then lift level by level. For each coarse eigenpair $(\mu, x)$, the candidate fine eigenvalues are the roots of a fitted local decimation map near $R^{-1}(\mu)$. In the exact case these come from $\det S(\lambda)=0$. Here we fit $R$ by least squares across the hierarchy and record the residual. The child vector is built from the formula above and refined with a few Rayleigh-quotient iterations on the fine Laplacian. Each refined child inherits its parent's sign and index and gets a branch label.

Fine modes that no lifted candidate converges to are flagged as newborn. They are labelled by where their energy concentrates at the coarse level, using the coarse coordinates already aligned. The output per subject is $K$ (say 50) spectral coordinate functions with identities comparable across subjects. These feed unchanged into the parcellation network of the baseline paper.

We also report the self-similarity defect for every level and frequency band. It is the relative residual $\|S(\lambda) x\| / \|x\|$ at the refined eigenpairs, together with the fraction of fine modes that are inherited rather than newborn. This is the scientifically new measurement. It tells us how far exact spectral-decimation theory transfers to real anatomy, which neither literature has tested.

Baselines

The naive reference point is the Euclidean graph CNN on raw coordinates. The parcellation paper reports 59.0% Dice on Mindboggle, with a 20.5% across-subject standard deviation. This floor shows what happens with no normalization at all.

The strongest existing method on this question is the numerical spectral embedding with ICP/Procrustes alignment from Graph Convolutions on Spectral Embeddings for Cortical Surface Parcellation. It reports 82.2% Dice without and 85.4% with MRF refinement, a 5.3% across-subject standard deviation, and about 18 seconds per subject. FreeSurfer reaches 83.2% Dice in about 3 hours. We will reimplement this pipeline on the standard 70/10/20 Mindboggle split. We give ICP and our method the same number of spectral coordinates, and we also run ICP with its originally reported small number of modes. Both pipelines use the same parcellation network and hyperparameter search budget, with and without the same MRF step. Beating it means higher Dice at matched mode count without MRF, together with lower across-subject variance and no greater runtime. A clear win would be Dice above 85.4% without MRF, or above it with MRF. Matching 82.2% while using only inherited labels would show feasibility but not superiority.

The ablation that isolates the key ingredient keeps the same hierarchy and the same coarse alignment but replaces Schur-complement lifting with plain interpolation of coarse eigenvectors followed by an independent fine eigensolve. Fine modes in this ablation are labelled by maximum correlation with the interpolated parent. A second ablation swaps MAXCUT elimination for random elimination of the same size. That tells us whether the near-diagonal $L_{22}$ matters for lineage stability or only for speed.

Why not just…?

Why not just register to a sphere with SUGAR or FreeSurfer and skip spectra entirely? Spherical registration is the field's default and the strongest practical competitor at scale. SUGAR processes 9,000 UK Biobank subjects in 32.57 minutes and outperforms FreeSurfer in 8 cortical areas. It does, however, require cortical inflation to a sphere first, uses Euler-angle deformation fields with known discontinuities, and enforces non-folding only by penalty. Our method works intrinsically on the native mesh with no inflation. We do not claim to beat SUGAR on raw throughput. We will compare parcellation Dice on Mindboggle using FreeSurfer-sphere-based labels as the alignment source. If SUGAR-style alignment plus the same network matches or exceeds lineage coordinates, the lineage method's value narrows to inflation-free settings and to its diagnostic role.

Why not just compute more eigenvectors with ICP and then match the higher ones by correlation? This is the real threat. Once ICP on the first few modes gives a rough vertex correspondence, one could compute 50 modes per subject and assign labels by Hungarian matching of correlations through that correspondence. We include exactly this as a baseline. Our prediction is that it degrades in near-degenerate bands, where correlation-based matching of individual vectors is ill-posed. There, lineage tracks a mode by its parent rather than by its shape. If the Hungarian variant equals ours on cross-subject mode consistency and Dice, the lineage contribution reduces to an elegant re-derivation, and we will say so.

Why not just use a multigrid eigensolver for speed? A standard coarse-to-fine eigensolver would also cut runtime, and the file gives no reason to think lineage is faster than one. Our claim is therefore about canonical labelling, not about speed. Runtime is reported only to show that lineage costs no more than the 15-second embedding step of the ICP pipeline.

Proposed Experiments

The first check decides whether to continue. On all 101 Mindboggle subjects we measure the self-similarity defect and the inherited fraction by frequency band. If fewer than about half of the lowest 50 modes are cleanly inherited, the core claim fails at this frequency range, and the project reports that negative result about transferring decimation theory to anatomy.

The second check measures mode identity directly. We use FreeSurfer's spherical registration purely as an external ground-truth correspondence. Through that correspondence we compute the across-subject correlation of each labelled mode $k$ for lineage, ICP, and Hungarian labelling. Success means lineage keeps high correlation further up the spectrum, especially inside near-degenerate clusters.

The third check is the main comparison: Mindboggle parcellation Dice, across-subject standard deviation, and per-subject runtime against all baselines above. Inputs, network, and tuning budget are held fixed, and only the alignment module is swapped.

The fourth check covers test–retest stability. On the CoRR-HNU subset of the CoRR/SALD cohort used by SUGAR, we measure whether a subject's labelled modes agree across sessions. SUGAR reports test–retest reliability only qualitatively, so there is no reported result on Lacuna for a numeric test–retest score to compare against. We will report our own agreement for both lineage and ICP.

The final check is a stress test on neonatal surfaces. On dHCP, where cortical folding changes rapidly with age, we feed lineage coordinates as positional encodings into a surface transformer for postmenstrual-age prediction. The reference is MS-SiT's 0.49 MAE on template data from The Multiscale Surface Vision Transformer. The goal is to see whether inherited modes stay consistent when anatomy is least alike.