How do hippocampal theta states reconfigure population activity by context?

Two theta states split by power and frequency

The paper asks whether oscillations relate to neural firing in a way that goes beyond phase. Phase precession, the classic hippocampal example, links the timing of spikes within a theta cycle to a single neuron's place field. Here the authors instead ask whether the power and frequency of theta itself co-vary with how the whole population of neurons fires. They fit a hidden Markov model — a data-driven method that segments a continuous signal into recurring states without being told what those states should be — to the local field potential recorded from dorsal CA1 of five rats. The model was run independently for each animal.

Across subjects it consistently returned two states. One combined low power with a lower theta frequency (LPLT); the other combined high power with a higher theta frequency (HPHT). The states were not brief artefacts: LPLT was typically active around 60% of the time and HPHT around 40%, their lifetimes were similar, and most switches between them occurred below 2 Hz, meaning at most two switches per second. Because the model had no knowledge of the task, the recurrence of the same two spectral peaks across animals is the main reason to treat them as real network states rather than fitting noise.

The states are not separate neuron subpopulations

A natural first guess is that each state recruits its own set of cells. The data argue against that. Many neurons did fire at significantly different rates between the two states, and HPHT was generally associated with higher firing rates. But most neurons were active during both states, and an attempt to predict which state was active from population activity performed close to chance. So the states do not map onto distinct subpopulations.

The authors also standardised the spike density across neurons for each trial and time point before the main analyses. This removes the trivial link that would otherwise arise because HPHT simply has more overall firing. What remains is the pattern of relative activity across neurons, which is the quantity the later decoding tests actually compare.

Movement explains part of the picture, not all of it

The rats were freely moving, so movement is an obvious confound. The authors separated locomotion along the track from local head movement at the port. The two states did not differ significantly in mean head position, but in three of four subjects they differed in the spread of positions, with HPHT visits covering a wider range. That is consistent with a partial link to movement, not with movement being the primary driver of state assignment.

They then removed trials containing clear running bouts — 13, 2, 18 and 2 trials for subjects 1 to 4 — and regressed head position out of the spike density data. Head position alone predicted trial outcome above chance, particularly 1–2 s after response, which fits reward-related licking on successful trials. After deconfounding, decoding of trial outcome from neural activity was slightly reduced but stayed above chance, while shuffled neural activity sat at chance. Movement contributes; it does not account for the neural signal on its own.

Outcome decoding differs between the two states

The central test asked whether success and failure are represented differently depending on the active theta state. The authors used state-conditioned decoding: train a classifier to read trial outcome from population activity during one state, then test it during the same state (within-state) or the other state (cross-state). If the population code were fixed, a decoder trained in one state should transfer cleanly to the other.

Cross-state accuracy was relatively high but significantly lower than within-state accuracy, and the gap was significant at some time points after the response. Per-state accuracy also fluctuated across time rather than sitting at a constant offset, which argues against the states merely differing in signal-to-noise ratio. The authors describe the pattern as ensemble pattern reconfiguration: the representation of success versus failure by the whole population changes with the state. They are explicit that significance was not reached in every subject under their permutation testing, so the effect is a hint at differential representation rather than a uniform result.

What the two-state model cannot settle

The authors state plainly that they do not claim exactly two theta states exist in the hippocampus. Two is the granularity the current data support, and the same effect could arise from variation along a continuous distribution rather than from categorical states. They also note that frequency and power were not identified as independent features, so this analysis cannot say whether one or both drives the effect.

The causal direction is left open as well. The paper describes an oscillation-to-neuron relationship but does not claim a unidirectional flow from oscillation to firing; the two are treated as properties of the same system emerging together. Mechanistic inference would need interventions and controlled task manipulations, and the authors suggest head-fixed preparations to separate cognitive from motor contributions more cleanly. The small number of animals also limits generalisation to other animals, brain regions or tasks.

Why a state-dependent code matters

The result challenges static models of neural coding, in which a population's representation of a behavioural variable is assumed to be fixed. If the same neurons represent success and failure differently depending on theta power and frequency, then a decoder or an experimenter who pools trials across states is averaging over two different codes. That is the practical consequence the paper's own comparison supports: within-state decoding outperformed cross-state decoding, so the state label carries information about how outcome is encoded.

The authors connect this to frequency sliding, the observation that oscillation frequency is not constant but drifts within a small range, and speculate that slower and faster frequencies could trade firing variability against precision. They also note that the two spectral peaks resemble the classical Type I and Type II hippocampal theta modes, though the states here are not solely movement-driven, since both are involved during trials performed at a constant spatial location.

The priority is state-aware analysis

The most useful takeaway is a comparison, not a new catalogue of states: within-state decoding beat cross-state decoding at post-response time points, and that gap is the paper's evidence for reconfiguration. Anyone analysing hippocampal population data from a similar task should therefore check whether outcome-related decoding changes when trials are split by theta power and frequency, rather than assuming one pooled decoder describes the population. The paper's own caveats set the threshold for how far that goes — the effect was not significant in every subject, two states are a chosen granularity, and power and frequency remain entangled — so the appropriate next step is a state-aware reanalysis plus an intervention or head-fixed design that can separate cognitive from motor contributions.

Sources used in this answer