NeuroTwin NFC Mathematical Constitution#
This is the canonical markdown entry point for the NeuroTwin NFC math. The longer LaTeX/PDF dossier remains in docs/research/neurotwin_nfc_research_dossier.tex and docs/research/neurotwin_nfc_research_dossier.pdf.
Primitive#
NFC treats each recording as a partial, noisy observation of a subject-specific latent neural field:
Here s indexes subject, x indexes neural location or parcel, t indexes time, and omega captures stochastic state.
Observation Operator#
Each modality is an observation operator over the same field:
A_s is anatomy or subject geometry, U is stimulus/task/context, and epsilon_m is modality-specific noise.
Generative Model#
The benchmark question is whether a field-mediated model explains held-out observations better than direct translation baselines.
Controlled Dynamics#
This says future field state depends on current field state, recent stimulus/context history, and subject anatomy.
Neural Field Dynamics#
Discretized Dynamics#
Low-Rank Pair Kernel#
This is where the old Pair-Operator idea survives: not as the main architecture, but as a low-rank relational field-update ablation.
fMRI Observation#
EEG and MEG Observation#
Spike, Calcium, and Behavior Observations#
These are theory entries unless corresponding adapters and tests are explicitly implemented.
Operator Learning Interpretation#
NFC is not direct modality fusion. It learns an inverse path from observations to latent field state and a forward path from field state to modality-specific readouts.
State-Space Interpretation#
NFC can be read as a controlled state-space model where F_t is the latent state and each modality supplies a partial observation channel.
Graph Calculus Interpretation#
Identifiability and Gauge Ambiguity#
Latent fields are identifiable only up to transformations unless constrained by architecture, observations, and regularizers.
Uncertainty and Calibration#
Uncertainty artifacts are not claim evidence unless they use actual uncertainty outputs and a documented calibration target.
Synthetic Proving Ground#
The NFC synthetic suite is a gate, not a result. It must test true field-grounded tasks, no-observation and no-pair ablations, no NaNs, strict shape contracts, and no target leakage.
Why Direct Fusion Is Not Enough#
Direct fusion can predict one modality from another, but it does not force a shared field explanation. NFC’s scientific bet is that field-mediated translation is more robust under held-out subject/site/dataset splits.
Why Pair-Operator Is a Submodule#
Pair-Operator captures low-rank relational updates. NFC needs that idea only as one possible kernel inside a broader latent-field and observation-operator model.