Cortical null-space decomposition and the array layout

What the work claims

CANDLE (Cortical-geometry Aware Null-space Decomposition for Localized ESI) is a learning-based electrophysiological source imaging (ESI) model with a precise division of labor. Scalp EEG is generated as X = L Y plus noise, where L is the lead-field matrix mapping cortical sources to electrodes, derived from each subject's T1-weighted MRI using a three-layer boundary element head model. The source space has far more dimensions than the sensor space, so the inverse problem has no unique solution. CANDLE's move is to decompose any source pattern into two orthogonal components: the range-space part, which the sensors can in principle observe, and the null-space part, which produces exactly zero at every electrode and is therefore invisible to any measurement.¹

The observable part is recovered analytically, by pseudoinverse of the denoised EEG. The network learns only the unobservable part, constrained so its estimate stays inside the true null space of L. The supporting infrastructure is a whole-brain simulator: Jansen-Rit neural mass models coupled by structural connectivity, instantiated on 1,113 individual cortical geometries from the Human Connectome Project, with source configurations derived from clustering 26,273 statistical brain maps against 1,315 cognitive terms. The empirical claims: on 12,762 simulated triplets, CANDLE achieves an HD95 of 37.3 mm and a Dice score of 0.48, improving on the second-best baseline by 2.5 mm and 0.15 points at p below 0.001; applied without any tuning to two real-data tasks, it localizes intracranial electrical stimulation sites from simultaneous scalp EEG with the lowest spatial dispersion (54.1 mm) and peak distance (32.9 mm) among eight methods, and estimates epileptogenic zones from presurgical EEG with the lowest HD95 (61.7 mm) and centroid distance (35.4 mm).

How it works

The mathematical skeleton is a range-null space decomposition of the lead-field operator. Writing L-dagger for the pseudoinverse, any source Y splits uniquely into L-dagger L Y, the component observable through L, and (I minus L-dagger L) Y, the component in the null space of L, which by construction satisfies L times null component equals zero. In the noiseless case the first component is exactly recoverable as L-dagger X, so the only genuinely unknown degrees of freedom are those in the null space. CANDLE therefore trains a denoiser on the EEG, recovers the observable component analytically from the denoised signal, and uses a second network to estimate the null-space component, projected back through the null-space operator so the physics can never be violated. A three-term loss supervises the final estimate, enforces forward consistency between denoised EEG and simulated EEG, and supervises the null-space component directly.

The training data problem is the clever half. Ground-truth cortical source activity is unobservable in real recordings, so the model is trained entirely on simulation: one neural mass model per cortical region, coupled by the Hagmann structural connectome, run on each of 1,113 subject-specific geometries with subject-specific lead fields. Source configurations come from NeuroVault: 26,273 statistical maps are characterized by their correlation profiles with 1,315 Neurosynth cognitive terms, clustered into 200 representative patterns, then randomly pruned and perturbed into 800 distinct source regions, with regional activity modulated through the excitatory gain of the neural mass model. Each simulation yields a triplet of geometry, simulated EEG and ground-truth sources, 12,762 in total, split subject-wise 70/15/15.

Three results anchor the claims. On simulation, CANDLE beats LCMV, sLORETA, wMNE, eLORETA and the two learning-based baselines ConvDip and DeepSIF plus GBF, with statistical significance against all of them. On intracranial stimulation localization from the Parmigiani dataset (32 epilepsy patients, 291 stimulation runs, roughly 31.5 trials per run, evaluated in a 10 ms window around pulse onset), CANDLE's improvements are significant over five of seven baseline families but not over sLORETA or GBF, and it is second to GBF at shallow sites yet best at every depth range beyond 10 mm. On epileptogenic zone estimation from the Vorderwulbecke dataset (22 patients with favorable 12-month surgical outcomes, resection cavity segmented by the MELD-PostOp model as reference), it improves the second-best baseline by 0.9 mm HD95 and 3.4 mm centroid distance, significant against LCMV on both metrics and against GBF on HD95. The ablations attribute the gains correctly: removing the null-space estimator degrades peak distance by 4.8 mm, and removing the denoising loss degrades it by 13.1 mm.

Where a skeptic should push

The load-bearing assumption is that the T1-derived boundary element lead field is accurate enough that the null space is located correctly. The entire method rests on the claim that a specific, high-dimensional subspace of source patterns is exactly unobservable. That claim is only as true as the head model: errors in tissue conductivity, segmentation or electrode co-registration shrink or rotate the true null space, and sources the model declares invisible may in reality be weakly visible, while sources it treats as visible may be mapped wrong. The decomposition is exact for L as computed, not for the head as built.

The second push is sim-to-real. Everything the network knows about plausible source patterns comes from one canonical neural mass model, one connectome and additive Gaussian noise. Real EEG carries line noise, muscle artifact, reference montage effects and nonstationarities the simulator does not model; the authors themselves flag the single Jansen-Rit model as a limitation. The empirical validations are therefore localization tasks against surrogate references (a known stimulation site, a resection cavity that only approximates the true epileptogenic zone), not reconstructions of ground-truth source activity, which nobody possesses. Note also the margins: 0.9 to 3.6 mm on the empirical tasks, with several pairwise comparisons not significant, and an EZ-estimation HD95 of 61.7 mm, which is centimeters of error in absolute terms. The depth stratification is the most honest signal in the paper: where sensors genuinely carry the information (shallow sites), the learned prior adds nothing and GBF wins; where sensors are blind (beyond 10 mm), the null-space prior earns its keep. That is exactly the signature a mechanism working as advertised should produce, and it quietly concedes that the prior buys what physics cannot give and nothing more.

The null space of your electrode array

Transpose the problem and it describes every high-density microelectrode array. An MEA also has a lead field: the array's electrode positions and pitches, embedded in the tissue's volume conductor, define a linear mapping from current-source distributions in three-dimensional tissue to the potentials sampled at the electrodes. That mapping has its own null space: source configurations in the tissue that produce exactly nothing at any electrode, patterns of synchronous inputs that cancel at the sampling grid, structures smaller than half the electrode pitch, depths the planar geometry cannot reach. Array pipelines increasingly run learned decoding and source estimation on top of the raw traces, and nearly all of them learn the whole mapping end to end. CANDLE's discipline says that is wasted capacity and an open door to unphysical answers: the forward model already determines the observable component, so the network should learn only the unobservable one, with its output projected back through the array's own operator so the physics can never be violated. For source-density estimation on 3D organoids recorded from planar or perforated high-density arrays, that is a concrete blueprint: derive the tissue-specific forward model, split analytically, learn the null space, train on a biophysical simulator with organoid-specific geometries, and validate zero-shot against interventions with known ground truth, exactly the stimulation-localization protocol this paper uses.

The opportunity is a principled answer to the channel-count question. Today, how many electrodes an array needs is set by vendor roadmaps and marketing. The null-space decomposition turns it into a measurable quantity: the value of the next electrode, or the next thousand, is the dimensionality it removes from the null space, and that is computable from the geometry and the volume-conductor model before a single cell is plated. A vendor or lab could compare layouts by their observable-subspace coverage for a target tissue model, and stop buying channels when the marginal null-space reduction flattens. Simulation-first training also fits the organoid world unusually well, because the same ground-truth problem that forced CANDLE into simulation is even more acute in tissue cultures: nobody can record the true transmembrane current distribution inside a living organoid either.

The threat has two parts, and both are non-obvious. First, hallucination with a physics license: a learned null-space prior invents source configurations the hardware provably cannot see, and the decomposition guarantees only that the estimate is consistent with what was measured, never that the invisible part is true. In a closed loop where such an estimator drives stimulation, decisions can be made on structure no electrode could ever audit, and the confidence is borrowed from the prior, not earned from the data. Second, geometry lock-in: the prior is conditioned on a specific sensor montage, and the paper's own fixed-montage limitation is the mild form of a harder rule for arrays, whose layouts are fixed by fabrication. An inversion model trained against one array geometry does not transfer to the next generation, so the AI readout becomes a layout-specific asset that must be re-derived whenever the hardware changes. That quietly raises the switching cost between array platforms and hands incumbents a moat that has nothing to do with electrode quality. Both threats argue for the same discipline the paper models: keep the learned component scoped to what physics leaves open, and expose the null-space estimate as a separately flagged quantity rather than folding it silently into the source map.

The bottom line

Established: on simulated data with true ground truth, the range-null space split beats seven baseline methods on localization by margins that are small but statistically clean, and the ablations tie those gains to the null-space estimator and the denoising loss specifically. Established within limits: zero-shot transfer to real EEG for two localization tasks, with the caveat that both use surrogate references and several baseline comparisons are not significant. Asserted, not demonstrated: that the learned null-space prior is biologically true rather than merely plausible, which no available measurement can settle. For the array world, the transferable content is architectural: forward-model-exact observable component, learned-and-projected unobservable component, simulator-trained, intervention-validated. What would confirm the approach for arrays is a demonstration on a high-density MEA with a calibrated tissue volume-conductor model and optogenetic or electrical stimulation ground truth; what would break it is evidence that learned null-space priors diverge across tissue preparations, which would make the invisible component preparation-specific and far less trustworthy than the paper's clean cross-subject EEG results suggest.