What Does the Local Field Potential Measure?
Companion post to:
Local field potentials primarily reflect inhibitory neuron activity in human and monkey cortex
Bartosz Telenczuk, Nima Dehghani, Michel Le Van Quyen, Sydney S. Cash, Eric Halgren, Nicholas G. Hatsopoulos & Alain Destexhe
Scientific Reports 7, 40211 (2017).
DOI: https://doi.org/10.1038/srep40211
What Does the Local Field Potential Measure?
Local field potentials are among the most widely used signals in systems neuroscience. They sit between scales. They are not single-neuron spikes, but they are also not as spatially diffuse as scalp EEG or MEG. They are recorded close to neural tissue, often with intracortical or depth electrodes, and they are usually interpreted as a measure of local population activity.
But this phrase, local population activity, hides a difficult biophysical question: which cells, which currents, and which spatial scales actually generate the LFP?
This paper was motivated by that question. More specifically, we asked whether one can identify the field contribution associated with the spike of a single neuron in vivo, and whether this contribution differs between putative excitatory and inhibitory neurons.
The answer was not simply that spikes generate nearby LFP deflections. The first answer was more cautionary: a standard spike-triggered LFP average is dominated by broad, population-level, and sometimes non-causal components. The second answer required a spatial unmixing step. After adapting a whitening filter to the covariance structure of the ongoing LFP, the spike-triggered field became much more focal. In that focal component, putative inhibitory neurons contributed earlier and more directly to the LFP than putative excitatory neurons.
The central conclusion is therefore not that the LFP is “just inhibition.” Rather, the conclusion is more precise:
The focal component of the LFP associated with single-neuron spiking in human and macaque cortex appears to be dominated by inhibitory interneuron activity, while the apparent spike-triggered LFP before spatial correction contains large diffuse components reflecting population correlations, recurrent activity, and extracellular mixing.
This matters because much of neurophysiology depends on how we interpret field signals. If the LFP is treated as a generic population readout, we risk missing the fact that the signal is shaped by cell type, synaptic geometry, extracellular filtering, volume conduction, and the recurrent structure of the local circuit.
The problem: the LFP is easy to record but difficult to interpret
The LFP is generated by transmembrane currents. Synaptic currents, return currents, active dendritic conductances, spike-related currents, and extracellular tissue properties can all shape the measured field. In practice, the LFP recorded at an electrode is not the activity of one neuron, one synapse, or one layer. It is a spatially filtered and temporally filtered mixture of many currents.
This is why the LFP occupies an ambiguous position in neural data analysis. It is more local than EEG or MEG, but it is not a simple microscopic signal. It is closer to a mesoscopic variable: a field signal created by many microscopic events and transformed by the tissue and recording geometry.
A useful way to think about the problem is:
\[\mathrm{LFP}_i(t) = \sum_{\alpha} G_{i\alpha} * I_{\alpha}(t),\]where (I_{\alpha}(t)) denotes a microscopic or mesoscopic current source, (G_{i\alpha}) describes how that source contributes to electrode (i), and (*) indicates temporal filtering. The measured field is therefore not the current itself. It is a mixture of currents through a spatial and biophysical observation operator.
The question is whether one can invert part of this mixture. Can we recover the field component specifically associated with the spike of one neuron?
Spike-triggered LFP: a tempting but dangerous estimate
The natural first tool is the spike-triggered LFP average. For a trigger neuron with spike times ({t_k}_{k=1}^{N_s}), the spike-triggered LFP at electrode (i) is
\[\mathrm{stLFP}_i(\tau) = \frac{1}{N_s} \sum_{k=1}^{N_s} \mathrm{LFP}_i(t_k + \tau).\]This gives the average field around the spike time of a neuron. If the spike reliably causes postsynaptic currents in nearby targets, one might expect the spike-triggered average to reveal the unitary LFP associated with that neuron.
But this is only true under strong assumptions. The spike-triggered average is a correlation measure. It does not automatically distinguish:
- currents caused by the trigger spike,
- currents that helped cause the trigger spike,
- activity of other neurons correlated with the trigger neuron,
- common-input effects,
- recurrent network activity,
- passive spread of extracellular fields.
In the paper, this became visible immediately. The raw spike-triggered LFP was often spatially broad. It could appear over much of the Utah array within only a few milliseconds. Some components even preceded the trigger spike. These components cannot be interpreted as postsynaptic consequences of that spike. A field deflection appearing before the spike, or appearing nearly simultaneously at distant electrodes, is not a simple causal footprint of that neuron.
This is an important point for field-potential analysis more generally. A spike-triggered field can reveal the embedding of a neuron in a larger dynamical population, but it does not by itself isolate the unitary contribution of that neuron.
The experimental setting
The data came from intracortical Utah-array recordings in human temporal cortex and macaque dorsal premotor cortex. These arrays provided simultaneous LFP and single-unit spiking activity across a dense grid of electrodes, with inter-electrode spacing of approximately (400\,\mu\mathrm{m}).
The human recordings came from patients implanted for clinical monitoring, while the macaque recordings provided a comparison in non-human primate cortex. Single units were classified into putative regular-spiking and fast-spiking groups based on extracellular spike waveform features. Regular-spiking units were interpreted as putative excitatory pyramidal neurons; fast-spiking units as putative inhibitory interneurons.
This classification is imperfect, especially in primate cortex, where some excitatory neurons can have relatively narrow spikes and some inhibitory neurons can have broader waveforms. But the classification is biologically motivated and supported by prior work using spike waveform and short-latency interactions.
The goal was not merely to compare spike-triggered averages between these two classes. The goal was to ask whether the local field contribution attributable to single-neuron spiking differs between putative excitatory and inhibitory neurons.
Spatial whitening: separating focal fields from diffuse population components
The key methodological move in the paper was to apply a spatial whitening filter to the spike-triggered LFP.
The logic is straightforward. If much of the spike-triggered LFP is dominated by covariance structure common to the ongoing LFP, then one can use the covariance of the ongoing LFP to suppress broad shared components. Let
\[C_{\mathrm{LFP}} = \left\langle x(t)x(t)^\top \right\rangle_t\]be the spatial covariance matrix of the ongoing LFP across electrodes. If
\[C_{\mathrm{LFP}} = E \Lambda E^\top,\]then the whitening operator is
\[W = C_{\mathrm{LFP}}^{-1/2} = E \Lambda^{-1/2} E^\top.\]The whitened spike-triggered LFP is then
\[\mathrm{wstLFP}(\tau) = W\,\mathrm{stLFP}(\tau).\]This operation is not just a generic spatial derivative. The resulting filters resemble Laplacian-like filters, but they are adapted to the empirical covariance structure of the ongoing field. In other words, they are data-driven spatial filters designed to suppress shared, broad components and preserve focal deviations.
Conceptually, this is a form of partial unmixing. The raw LFP is a spatially mixed signal. The spike-triggered average inherits that mixing. Whitening reduces the effect of the dominant covariance modes of the LFP and reveals components that are more spatially localized around the trigger neuron.
The paper also validated this approach with a simple linear model of LFP generation. In that model, the standard spike-triggered LFP was broader than the underlying unitary kernel, while whitening recovered a spatial profile closer to the true kernel.
What changed after whitening?
After spatial whitening, the spike-triggered field became much more focal. Instead of spreading over much of the array, the whitened spike-triggered LFP was largely restricted to the neighborhood of the trigger neuron, with spatial decay within approximately (1\,\mathrm{mm}).
This is the critical distinction:
\[\mathrm{stLFP} = \text{focal spike-related component} + \text{diffuse population/covariance component}.\]The raw (\mathrm{stLFP}) includes both terms. The whitened (\mathrm{wstLFP}) is an attempt to isolate the first.
This matters because a broad spike-triggered LFP could easily be misinterpreted as evidence that the spike of a single neuron has a spatially extensive direct field consequence. The whitening analysis shows that much of this apparent spatial extent is not local in that sense. It reflects the structure of the ongoing population signal and the physics of field propagation.
The spatial decay of the whitened signal was much sharper than that of the raw spike-triggered LFP. In the paper, the whitened LFP decay constants were approximately (0.15)–(0.20\,\mathrm{mm}) for fast-spiking neurons and (0.20)–(0.25\,\mathrm{mm}) for regular-spiking neurons. These values are consistent with a focal local-circuit contribution rather than a diffuse array-wide event.
Latency and propagation: why timing matters
Spatial localization alone is not enough. A focal component could still reflect some artifact or non-synaptic effect. The timing of the field is therefore essential.
The paper examined the latency of the trough of the spike-triggered field as a function of distance from the trigger neuron. If the field reflects a propagating synaptic process, one expects latency to increase with distance. A simple model is
\[\tau(d) = \tau_0 + \frac{d}{v},\]where (d) is distance from the trigger neuron and (v) is an effective propagation speed.
For the whitened spike-triggered LFP, the inferred propagation speeds were in the range
\[v \approx 0.08-0.29 \ \mathrm{m/s},\]which is consistent with propagation along unmyelinated axons and synaptic delays. This supports the interpretation that the whitened component is closer to a synaptically mediated local field contribution.
The raw spike-triggered LFP, by contrast, can include fast components produced by passive field spread and common population activity. This can make its apparent propagation difficult to interpret.
Thus, the whitening step changes not only the spatial structure of the signal, but also the interpretation of its timing.
Inhibitory neurons dominate the focal LFP contribution
The main biological result concerns the difference between putative inhibitory and excitatory neurons.
Fast-spiking neurons produced focal whitened spike-triggered LFPs with shorter latencies than regular-spiking neurons. In human cortex, this latency difference was especially clear near the trigger neuron. In macaque cortex, the pattern was also present, though with some differences across distances.
The polarity of the whitened LFPs was also important. If excitatory and inhibitory neurons directly generated fields through their own postsynaptic currents in the simplest possible way, one might expect opposite polarities. But the observed whitened fields for regular-spiking and fast-spiking neurons had the same polarity.
This suggests a circuit mechanism. The contribution associated with regular-spiking neurons may be mediated disynaptically through inhibitory interneurons:
\[\text{RS spike} \rightarrow \text{interneuron activation} \rightarrow \text{inhibitory postsynaptic currents} \rightarrow \text{LFP contribution}.\]In this interpretation, inhibitory neurons contribute more directly to the focal LFP, while excitatory neurons contribute partly through recruitment of inhibitory circuitry. The extra latency of the regular-spiking contribution is then naturally explained by synaptic delay through an intermediate inhibitory population.
This provides a mechanistic interpretation of the title: the LFP, at least its focal spike-related component in these data, primarily reflects inhibitory neuron activity.
Why this does not mean that excitation is absent
It is important not to overstate the result. The conclusion is not that excitatory neurons are irrelevant to LFP generation. Excitatory neurons shape cortical activity, drive recurrent interactions, recruit inhibition, and contribute to synaptic currents. The cortex is an interacting E/I system, not two independent populations.
The more subtle claim is that the focal unitary component recovered from spike-triggered LFPs appears to be dominated by inhibitory postsynaptic effects. Excitation may appear in the field through indirect pathways, through dendritic currents, through recurrent recruitment of inhibition, and through population-level correlations.
In other words, the LFP is not a transparent readout of either excitation or inhibition. But when the diffuse field correlations are reduced, the remaining local spike-associated component points strongly toward inhibitory circuitry.
This is especially relevant for interpreting rhythms. Oscillations in the LFP are often discussed in terms of excitation, inhibition, synchrony, and recurrent loops. If inhibitory currents have a privileged role in shaping the field, then LFP rhythms may often be closer to a readout of inhibitory timing and local circuit control than to a simple measure of population firing.
Relation to excitation-inhibition balance
This paper also connects naturally to a broader theme in cortical dynamics: the balance and interaction of excitation and inhibition.
In a balanced cortical network, excitation and inhibition are not merely opposing forces. They are dynamically coupled processes. Inhibition controls gain, timing, synchrony, and the spatial spread of activity. It can stabilize recurrent excitation, sculpt oscillatory events, and constrain the effective dimensionality of population dynamics.
From that perspective, the result that focal LFPs primarily reflect inhibitory activity is not surprising. Inhibitory neurons are positioned to coordinate local population timing. They receive strong excitation, respond rapidly, and impose structured currents onto local populations. Their synaptic effects are often temporally precise and spatially organized. These are exactly the kinds of currents that can dominate a local field measurement.
Thus the paper provides a bridge between microscopic cell-type physiology and mesoscopic field dynamics. It suggests that when we observe an LFP deflection, especially a focal spike-triggered component, we may often be observing the signature of inhibitory control over the local circuit.
A measurement lesson: fields are not populations without physics
A major lesson of the paper is methodological. Neural signals are not only generated by circuits; they are transformed by measurement physics.
The LFP is often treated as if it were a direct population variable. But the paper shows that the observation process matters. The raw spike-triggered LFP contains broad, non-local, and even non-causal components. Only after accounting for the spatial covariance of the ongoing field does a more biologically interpretable local component emerge.
This is a general lesson for electrophysiology:
\[\text{measured signal} \neq \text{neural variable of interest}.\]Rather,
\[\text{measured signal} = \text{neural dynamics} \circ \text{biophysical mixing} \circ \text{recording geometry} \circ \text{analysis method}.\]For LFP, ECoG, EEG, and MEG, this distinction is fundamental. These signals are not wrong or inferior because they are mixed. They are powerful precisely because they expose collective dynamics. But their interpretation requires explicit attention to the transformation between microscopic currents and macroscopic measurements.
The whitening approach used here is one way to make that transformation more explicit.
Implications for systems neuroscience
For systems neuroscience, the paper has several implications.
First, spike-field coupling should not be interpreted too quickly as a direct causal relation. A neuron can be correlated with a field because it is embedded in a larger population rhythm, because it receives common input, or because its spike recruits local synaptic currents. These are different mechanisms.
Second, spatial scale matters. A raw spike-triggered LFP extending across millimeters does not necessarily mean that one neuron directly drives a millimeter-scale field. Much of that extension can be due to shared population activity and extracellular mixing.
Third, cell type matters. The same field signal can have different interpretations depending on whether it is related to inhibitory or excitatory spiking. A field potential is not just “population activity”; it is a cell-type-weighted, geometry-dependent, biophysically filtered population signal.
Fourth, invasive human recordings can reveal principles of cortical physiology that are difficult to obtain otherwise. The combination of human Utah-array recordings and macaque data allowed the paper to test whether the observed spike-field relationship generalized across species and cortical areas.
Implications for computational neuroscience and modeling
For computational neuroscience, the result argues against overly abstract interpretations of LFP as a generic low-pass version of firing rate.
A useful model of the LFP should include at least four ingredients:
- cell-type-specific synaptic currents,
- spatial arrangement of neurons and synapses,
- extracellular and dendritic filtering,
- recurrent population correlations.
A simplified firing-rate model may reproduce some low-frequency features of a field signal, but it will not explain why inhibitory spikes have earlier focal field signatures, why raw spike-triggered LFPs are spatially broad, or why spatial whitening changes the apparent propagation structure.
This is especially important for models that try to connect spiking networks to LFP, ECoG, EEG, or MEG. The mapping from spikes to fields is not a simple summation. It is a structured observation model.
A minimal conceptual model might be written as
\[\mathbf{x}(t) = \mathbf{G} \left[ \mathbf{I}_E(t) + \mathbf{I}_I(t) \right] + \boldsymbol{\eta}(t),\]where (\mathbf{x}(t)) is the recorded field, (\mathbf{I}_E(t)) and (\mathbf{I}_I(t)) are excitatory and inhibitory current sources, (\mathbf{G}) is the spatial mixing operator, and (\boldsymbol{\eta}(t)) includes background and measurement noise. The result of the paper suggests that (\mathbf{I}_I(t)), especially when recruited locally and rapidly, can dominate the focal spike-associated component of (\mathbf{x}(t)).
Implications for NeuroAI
There is also a NeuroAI lesson here, though it should be stated carefully.
The paper is not about artificial neural networks. But it speaks to a broader issue: biological computation is not only about excitation, firing rates, or feedforward transformations. It is also about inhibitory control, recurrent timing, and local circuit motifs.
If field potentials partly expose inhibitory control signals, then LFP-like objectives or neural-data constraints used in NeuroAI models may preferentially constrain the inhibitory and recurrent structure of the model. This matters for biologically inspired RNNs, E/I-constrained networks, and models that attempt to learn from neural recordings.
In many artificial systems, inhibition is treated as a sign constraint or a stabilizing term. In cortex, inhibition is more than that. It shapes gain, synchrony, timing, competition, oscillations, and local routing. The present result reinforces that inhibitory circuitry is central to the observable dynamics of cortex.
For NeuroAI, one possible lesson is that biologically realistic models should not only match firing rates or representational geometry. They should also match the field-generating dynamics of local circuits. A model that reproduces spikes but fails to reproduce inhibitory field signatures may be missing an important part of cortical computation.
Limitations and caution
Several cautions are important.
First, the spike-triggered LFP and whitened spike-triggered LFP are still correlational. Whitening improves spatial specificity, but it does not create a direct causal perturbation. The ideal experiment would trigger spikes externally at random times and measure the resulting field, but that is technically difficult in intact human and macaque cortex.
Second, cell-type classification was based on extracellular spike waveform. This is a standard and useful approach, but not perfect. Some excitatory neurons can have narrow waveforms, and some inhibitory neurons can have broader waveforms. The main conclusions are robust to reasonable levels of misclassification, but the classification should still be understood as putative.
Third, the human recordings came from patients with epilepsy, even though the analyzed periods excluded epileptic activity and the arrays were not placed in the epileptic focus. The agreement with macaque recordings helps support generality, but this remains an important consideration.
Fourth, the cortical layer and spatial distribution of synapses matter. A full biophysical account would require more detailed knowledge of axonal arborization, synaptic placement, dendritic geometry, and layer-specific current flow.
These limitations do not weaken the central result. They clarify its interpretation: the paper provides strong evidence that the focal spike-associated LFP component in these recordings is dominated by inhibitory circuitry, but it does not claim to fully solve the inverse problem of LFP generation.
The broader picture
The LFP is often treated as a convenient intermediate signal: more accessible than spikes at population scale, more local than EEG, and easier to analyze than full biophysical current-source models. This convenience is real. But it can also be misleading.
The paper shows that to interpret the LFP, one must separate at least two things:
\[\text{local unitary field} \quad \text{versus} \quad \text{diffuse population field}.\]The raw spike-triggered LFP is dominated by the second. Spatial whitening helps reveal the first. Once the first is isolated, inhibitory neurons emerge as the primary contributors to the focal field.
This gives a more precise interpretation of what the LFP is measuring. It is not simply “nearby neural activity.” It is a field generated by structured currents in a recurrent circuit, transformed by tissue and measurement geometry. In the local spike-associated component, inhibitory interneurons appear to play a dominant role.
For anyone using LFP, ECoG, EEG, or MEG to reason about cortical computation, the message is clear: field potentials are not just signals to decode. They are biophysical observables. To understand them, we must model the circuit and the measurement together.
Short summary
Local field potentials are widely used as mesoscopic readouts of cortical activity, but their cellular origin is difficult to interpret. In this paper, we studied spike-triggered LFPs in human and macaque cortex using Utah-array recordings. Raw spike-triggered LFPs were broad, sometimes non-local, and partly non-causal, indicating strong contamination by ongoing population correlations and field mixing. By applying a spatial whitening filter adapted to the covariance of the ongoing LFP, we recovered a more focal spike-associated field. This whitened component was restricted to the local neighborhood of the trigger neuron, propagated at speeds consistent with unmyelinated axonal transmission, and showed earlier contributions from putative inhibitory neurons than from putative excitatory neurons. The results suggest that focal LFPs in human and monkey cortex primarily reflect inhibitory interneuron activity and that interpreting field potentials requires explicit attention to biophysical mixing, cell type, and local circuit dynamics.