Brain-imaging analyses almost always assume that every voxel turns neural activity into a blood-oxygen signal through the same fixed filter, the "canonical" hemodynamic response function (HRF). The HRF is the delayed, smeared blood-flow echo that functional MRI (fMRI) measures in place of neural firing. This assumption is the neuroimaging counterpart of the simplified channel models used in wireless engineering. A single smooth fading law is convenient to analyze, but systems tuned to it fail once the real medium scatters signals along several distinct paths, much as Beer's-law attenuation overestimates underwater quantum key rates (Monte Carlo-Based Performance Analysis for Underwater Continuous-Variable Quantum Key Distribution). The fMRI version of that failure is clearest in white matter. There the HRF has a pronounced early negative dip, a lower peak and a narrower width, and its shape changes along each fiber tract and covaries with diffusion-MRI measures of microstructure. In white matter the dip correlates with neurite density in the opposite direction from gray matter (Anomalous and heterogeneous characteristics of the BOLD hemodynamic response function in white matter).
We propose to borrow the wireless remedy that replaced dense channel grids: describe the channel as a handful of physical "paths," each with a delay, a spread and a gain, and track only those few parameters (AFDM for LEO Inter-Satellite Links: Path-Level CSI Prediction and CRLB-Guided Pre-Equalization). We then add the ingredient fMRI makes available and radio does not. Each voxel's path parameters are predicted in advance from its diffusion-MRI microstructure, using a mapping learned on resting-state data from other people. In short, a voxel's HRF becomes a small, physically structured kernel whose likely shape can be read off the tissue before any task data are fit. Task data then only need to correct it. The payoff is largest exactly where current methods break: low-signal voxels and white matter, where flexible per-voxel estimates are too noisy and template libraries contain no correct shape.
Core Claim
A three-component kernel whose parameters are predicted from tissue class, neurite density, fractional anisotropy, radial diffusivity and along-tract position should do three things in held-out subjects. It should predict held-out task fMRI better than a canonical HRF, a free-shape per-voxel HRF, or a library of template HRFs. It should give more reliable single-trial response estimates. Its advantage should concentrate in white matter and low-signal gray-matter voxels. The new element is not the multi-component shape itself. It is the cross-subject, microstructure-conditioned prior over that shape, which lets kernels transfer to voxels with too little signal to estimate them alone.
Research Program
For voxel $v$, the kernel is a sum of $P=3$ smooth components (an early dip, a main positive response and a late undershoot):
$$h_v(t) = \sum_{p=1}^{P} a_{v,p}\, g\!\left(t;\ \tau_{v,p}, \sigma_{v,p}\right),$$
where $g(t;\tau,\sigma)$ is a unit-area gamma-shaped bump with peak delay $\tau$ and width $\sigma$, and $a_{v,p}$ is the signed gain of component $p$. Collect the nine parameters into $\theta_v$, and write $m_v$ for the voxel's microstructure vector. We fit voxel parameters, a shared map $f$, and condition amplitudes $\beta_v$ by minimizing
$$\sum_v \Big[\ \| y_v - X(h_v)\,\beta_v \|^2 + \lambda\, s_v\, \| \theta_v - f(m_v) \|^2 \Big],$$
where $y_v$ is the voxel time series, $X(h_v)$ is the design matrix convolved with the kernel, and $s_v$ is a weight that grows as the voxel's estimated signal-to-noise falls. Noisy voxels are therefore pulled toward the microstructure prediction. Clean voxels can depart from it.
The map $f$ is a linear model with tissue-specific slopes, or at most a small two-layer network. It is fit first on resting-state HRFs, which are estimated by blind deconvolution in training subjects following Schilling et al.'s pipeline. It is then frozen or lightly refined on task data. This design tests something not yet tested in the file's literature: whether HRF shapes measured at rest, together with their microstructure dependence, carry over to task responses. As a sanity check, the learned slopes should reproduce Schilling et al.'s finding that the dip–microstructure relationships have opposite signs in white and gray matter. If they do not, the mapping is fitting noise.
The approach relates to existing estimators as follows. Pedregosa et al.'s R1-GLM shares one free-shape HRF across conditions within a voxel, and that rank-1 constraint is what keeps its estimates stable (Data-driven HRF estimation for encoding and decoding models). We keep that sharing, but replace the free shape with nine physical parameters and add a prior that pools across voxels and subjects; R1-GLM treats each voxel independently. GLMsingle chooses each voxel's HRF from 20 fixed curves and adds denoising and ridge regularization (Improving the accuracy of single-trial fMRI response estimates using GLMsingle). Our continuous kernel can produce dips that the library cannot, and GLMsingle's denoising stage can be used unchanged on top of it. Learned per-region HRF layers inside deep translation models such as SAMBA (Latent Representation Learning for Multimodal Brain Activity Translation) and the HRF generator in CausalMamba (CausalMamba: Scalable Conditional State Space Models for Neural Causal Inference) motivate the need for variable kernels. Neither ties the kernel to measured tissue, and both work at the level of brain regions rather than voxels.
Two related problems are left out of scope deliberately: non-linear responses to brief contrast decrements (Voxel-Wise Linearity Analysis of Increments and Decrements in BOLD Responses in Human Visual Cortex Using a Contrast Adaptation Paradigm) and drift across runs, for which online conformal methods from wireless link adaptation are a natural later step (Distribution-Free Adaptive Modulation and Coding via Online Conformal Inference under Non-Stationary Fading). The model stays linear and time-invariant within a run.
Baselines
The naive reference is a standard general linear model (GLM), the per-voxel regression of the time series on the experimental design, using the canonical double-gamma HRF. A second simple baseline adds temporal and dispersion derivative regressors to that canonical HRF. A third is a finite impulse response (FIR) model, which estimates one free coefficient per time point; it bounds how much flexibility alone can buy. The strongest flexible-shape competitor is R1-GLM. On the Kay et al. natural-images data it improved image identification by up to 8.2% over fixed-HRF models, and it found that only 50% of voxels peak at the canonical time (Pedregosa et al.). The strongest single-trial competitor is GLMsingle, which roughly tripled image-level decoding accuracy in the Natural Scenes Dataset (NSD) and doubled it in BOLD5000 relative to a baseline GLM (Prince et al.). These reports are relative, so we will rerun every baseline ourselves rather than compare against published figures.
The most dangerous baseline is our own construction: GLMsingle with its library extended by 20 additional curves, made by clustering the resting-state HRFs of training subjects, white matter included. It receives the same training data as our prior. Two ablations isolate the key ingredient. The first uses the same three-component kernel with a prior equal to the population mean for each tissue class, removing microstructure conditioning. The second uses the microstructure-predicted kernel with no per-voxel correction.
For fairness, all methods see identical preprocessed runs and identical nuisance regressors (GLMdenoise components are given to every method). Hyperparameters, meaning ridge strength, $\lambda$ and the library size, are tuned on the same validation runs with the same search budget. We count a win if our method beats the augmented library in held-out-run prediction and in split-half beta reliability in white matter, with paired tests across subjects, while not losing in gray matter.
Why not just…?
Why not just add white-matter-shaped curves to GLMsingle's library? This is the real threat, and it is why the augmented library is a baseline rather than a straw man. A library learns the typical shapes but not which voxel should get which shape. In low-signal voxels, choosing among 40 curves from task data is itself noisy. Our prior instead chooses using the tissue. If the augmented library matches us, the conclusion is that microstructure conditioning adds nothing, and we will report that.
Why not just use R1-GLM, which already allows any shape? R1-GLM estimates roughly 20 free shape values per voxel from that voxel alone. Its rank-1 constraint helps when a voxel has many conditions and strong signal, which describes the visual cortex data it was tested on, not white matter. The comparison stratified by signal-to-noise will show whether its advantage disappears in low-signal voxels, as we predict.
Why not just use the canonical HRF with derivatives? Derivatives describe small shifts of about a second around the canonical peak. They cannot produce a negative early dip of the size Schilling et al. report. Held-out prediction in white matter tests this directly, and the tissue-mean ablation shows how much of any gain comes merely from allowing a dip.
Proposed Experiments
The primary data are the Human Connectome Project releases, which pair resting-state fMRI, diffusion MRI with neurite-density measures, and task fMRI for the same subjects. Schilling et al. analyzed the resting-state and diffusion data from 199 of these subjects; CausalMamba used the task data from 1,081. We learn $f$ on resting-state data from a training split and evaluate on task runs (working memory and motor) from disjoint subjects. Metrics are held-out-run variance explained per voxel, split-half reliability of condition and trial amplitudes, and the number of white-matter voxels with replicable task responses. Results are stratified by tissue class, tract and voxel signal-to-noise.
For gray-matter single-trial estimation we use NSD and BOLD5000, the datasets on which GLMsingle reported its decoding gains. There we run the tissue-class version of our model, and the microstructure-conditioned version wherever diffusion data are available for the same subjects. Image-level decoding accuracy and test-retest reliability, scored exactly as in Prince et al., are the outcomes.
Three results would settle the claim. The first is whether the microstructure-predicted kernel alone, with no task fitting, beats the canonical HRF in held-out white-matter prediction. That would show transfer from rest to task. The second is whether the full model's margin over the augmented library and the tissue-mean ablation grows as voxel signal-to-noise falls, the signature of a useful prior rather than of extra flexibility. The third is whether the learned slopes reproduce the opposite-sign white/gray microstructure relationships. If the first fails, rest-derived kernels do not describe task responses, and the program reduces to a better per-voxel shape model. If the second fails, the wireless-style few-path parameterization is merely a tidier library, and we will say so.
References
- Monte Carlo-Based Performance Analysis for Underwater Continuous-Variable Quantum Key Distribution
- Anomalous and heterogeneous characteristics of the BOLD hemodynamic response function in white matter
- AFDM for LEO Inter-Satellite Links: Path-Level CSI Prediction and CRLB-Guided Pre-Equalization
- Data-driven HRF estimation for encoding and decoding models
- Improving the accuracy of single-trial fMRI response estimates using GLMsingle
- Latent Representation Learning for Multimodal Brain Activity Translation
- CAUSALMAMBA: SCALABLE CONDITIONAL STATE SPACE MODELS FOR NEURAL CAUSAL INFERENCE
- Voxel-Wise Linearity Analysis of Increments and Decrements in BOLD Responses in Human Visual Cortex Using a Contrast Adaptation Paradigm
- Distribution-Free Adaptive Modulation and Coding via Online Conformal Inference under Non-Stationary Fading