From paper summary to your first question

This paper investigates whether spontaneous, distributed brain activity contains low-dimensional rotational dynamics that predict responses to external cues and constrain learning. Conventional dimensionality-reduction methods emphasize large fluctuations and may overlook smaller, behaviorally relevant signals. The authors introduce rotational component analysis (RCA) and apply it to existing monkey and human motor-cortical recordings, mouse cortex-wide calcium imaging, and human fMRI and MEG data. RCA identifies a two-dimensional projection that maximizes the strength and consistency of directed angular current. These rotations represent circulation through a latent activity space, not necessarily waves physically rotating across the cortex. Its objective—an estimator of the asymmetry production rate (APR)—quantifies dynamical mirror-symmetry breaking using the squared mean accumulated current relative to its variance. Unlike principal component analysis (PCA), RCA does not maximize explained variance and requires neither behavioral labels nor trial averaging. In reaching benchmarks, RCA recovered more consistent single-trial circulation than jPCA, which fits rotational dynamics within a leading-PCA subspace. For representative conditions, current signal-to-noise ratios for RCA versus jPCA were 0.89 versus 0.05 in macaque recordings and 1.90 versus 0.98 in human recordings. The central behavioral analysis examined recordings from 44 cortical regions in 25 mice learning an auditory-cued lever-pull task over 15 days. A rotational plane learned exclusively from resting activity explained only 0.6% of total variance, compared with 87% for the leading PCA plane, yet clearly distinguished cue-onset states associated with subsequent responses. Its standardized response-gradient magnitude was 0.151, compared with 0.050 for PCA and 0.078 for jPCA, placing it near the upper extreme among 2,000 random planes. In held-out-session classification, the cue-onset RCA state achieved an area under the receiver operating characteristic curve (AUC) of 0.638. In a separate pupil-matched analysis, RCA outperformed pupil-indexed arousal, with AUCs of 0.611 and 0.535, respectively. A two-dimensional stochastic model explained this association through an interaction between ongoing rotational flow and cue-evoked drive: responses were more likely when the combined drift pointed strongly in the response direction. Empirical response probability increased from below 20% to above 70% across the observed drift-score range. A trained 64-unit recurrent neural network reproduced the main geometry without explicitly imposed rotation. Across mouse training, intrinsic drift remained comparatively stable, while cue-evoked drift strengthened. A phenomenological reinforcement model suggested that reward could preferentially strengthen responses compatible with existing dynamics. RCA also detected resting rotations in human fMRI and MEG data, explaining 0.2% and 2.6% of variance, respectively, with estimated periods of 51 and 4.4 seconds. In a comparison of 53 controls and 82 participants with early non-affective psychosis, APR was lower in the psychosis group (P = 5.22 × 10⁻⁷), whereas leading-PC variability did not differ significantly. Split-sample analyses supported the APR finding. An extension that detects two counter-rotating centers distinguished preparatory from post-cue activity in mice and fixation from social-task activity in human fMRI. Combining the mouse preparatory state measured 2.7 seconds before the cue with the cue-onset RCA state increased held-out AUC to 0.677. These findings support rotational flow as a compact coordinate linking ongoing activity, behavioral readiness, and learning. However, the observational reanalyses establish associations rather than causal constraints. The proposed learning mechanism remains hypothetical, predictive accuracy is modest, and the clinical differences do not establish diagnostic utility. APR should not automatically be interpreted as thermodynamic entropy production, and its moment-based estimator is subject to distributional qualifications. Finally, RCA’s superior rotation scores partly reflect the objective it directly optimizes, and code availability was deferred until publication.

1. Scientists use a tool called RCA to find tiny, circle-like patterns in brain signals, but these are not real circles moving through the brain.
2. RCA found patterns in monkeys, mice, and people that other tools sometimes missed.
3. In mice, these patterns gave small clues about whether a mouse would answer a sound, and learning made the brain’s push from the sound stronger.
4. A measure called APR was lower in people with a condition called psychosis, but this does not mean it can tell who has that condition.
5. These findings may help explain how brains get ready and learn, but they do not prove that the patterns cause actions or learning.

Definitions

• RCA: a tool that looks for circle-like changes in brain signals.
• Patterns: things that happen in a repeated or organized way.
• Brain signals: messages made by brain cells.
• Learning: getting better at something or finding out something new.
• APR: a number showing how much brain-signal patterns look different when time runs backward.
• Psychosis: a condition that can make it hard to tell what is real.
• Cause: to make something happen.

• Rotational component analysis (RCA) identifies low-dimensional rotational dynamics in ongoing brain activity without behavioral labels or trial averaging, revealing signals that variance-focused methods may miss.
• RCA maximizes consistent directed circulation in a two-dimensional latent space; these rotations do not necessarily correspond to physical waves across the cortex.
• Across monkey and human motor recordings, RCA recovered more consistent single-trial rotations than jPCA, although this advantage partly reflects RCA’s optimization objective.
• In 25 mice learning an auditory-cued task, a rotational plane derived solely from resting activity explained just 0.6% of total variance yet predicted subsequent responses better than PCA or jPCA states.
• Behavioral prediction was modest: cue-onset RCA states achieved held-out AUC of 0.638, increasing to 0.677 when combined with earlier preparatory states; RCA also outperformed pupil-indexed arousal in a separate analysis.
• Modeling linked response probability to the interaction between ongoing rotational flow and cue-evoked drive, with observed response probability rising from below 20% to above 70% across drift scores.
• During learning, intrinsic drift remained relatively stable while cue-evoked drift strengthened, suggesting—but not proving—that learning preferentially reinforces responses compatible with existing dynamics.
• RCA detected resting rotations in human fMRI and MEG, and an extended method distinguished preparatory from post-cue activity and fixation from social-task activity.
• Participants with early non-affective psychosis had lower asymmetry production rate (APR) than controls, despite no significant difference in leading-PC variability; this does not establish diagnostic utility.
• The findings support rotational flow as a compact link between ongoing activity, behavioral readiness, and learning, but remain observational; causal constraints and the proposed learning mechanism are unproven.
• APR measures dynamical mirror-symmetry breaking and should not automatically be interpreted as thermodynamic entropy production; its estimator has distributional qualifications, and code release was deferred until publication.

The Brain’s Quiet Rotations May Shape Its Next Move

Why does the same signal prompt action one moment and go unanswered the next? Part of the answer may lie in what the brain is doing before the signal arrives. A study suggests that spontaneous brain activity contains subtle rotational patterns that help predict whether an animal will respond—and may help explain how learning builds on existing neural dynamics.

In “Spontaneous brain-wide rotational dynamics constrain cue-evoked behavior,” Sihan Chen, Yunlong Xu, Doruk Efe Gökmen, Michel Fruchart, Brent Doiron, and Vincenzo Vitelli introduce a method for finding these patterns. Their results suggest that some behaviorally informative signals occupy surprisingly little of the brain’s overall activity.

Finding Signals That Bigger Fluctuations Can Hide

Brain recordings contain many signals unfolding together. To make sense of this complexity, researchers often use dimensionality reduction: mathematical tools that compress activity into a few coordinates. Principal component analysis, or PCA, prioritizes the directions with the largest fluctuations. But the biggest fluctuations are not necessarily the most relevant to behavior.

The authors developed rotational component analysis, or RCA, to look for something different: consistent circulation through a two-dimensional activity space. Imagine representing the brain’s changing state as a moving dot. RCA searches for a view in which that dot reliably travels around rather than simply wandering.

These rotations are mathematical patterns in a latent activity space—not necessarily waves physically spinning across the cortex. RCA requires neither behavioral labels nor trial averaging, allowing researchers to identify patterns without first telling the method which moments correspond to particular actions.

A Measure of Directed Circulation

RCA optimizes an estimator called the asymmetry production rate, or APR. It compares the squared mean accumulated current with its variance, rewarding circulation that has a strong, consistent direction. This captures dynamical mirror-symmetry breaking: activity favors one direction of circulation over its mirror opposite.

In reaching datasets, RCA recovered more consistent single-trial circulation than jPCA, a method that fits rotational dynamics within a leading-PCA subspace. For representative macaque conditions, current signal-to-noise ratios were 0.89 for RCA and 0.05 for jPCA; in human recordings, they were 1.90 and 0.98. These comparisons demonstrate RCA’s focus, although stronger rotation scores partly reflect the objective it explicitly optimizes.

A Tiny Signal Linked to Behavioral Readiness

The central behavioral analysis used recordings from 44 cortical regions in 25 mice learning an auditory-cued lever-pull task over 15 days. Crucially, the researchers learned the rotational plane exclusively from resting activity, then examined whether the brain’s position in that plane at cue onset predicted a subsequent response.

The RCA plane captured just 0.6% of total activity variance, compared with 87% for the leading PCA plane. Yet it separated cue-onset states associated with different response probabilities more clearly. Its standardized response-gradient magnitude was 0.151, versus 0.050 for PCA and 0.078 for jPCA, placing it near the upper extreme among 2,000 random planes.

Prediction on held-out sessions was meaningful but modest: the cue-onset RCA state achieved an area under the receiver operating characteristic curve, or AUC, of 0.638, where 0.5 indicates chance performance. In a separate pupil-matched analysis, RCA also outperformed pupil-indexed arousal, with AUCs of 0.611 and 0.535.

When a Cue Pushes With the Flow

A two-dimensional stochastic model offered an intuitive explanation. The cue adds a push to activity already moving through the rotational plane. Responses become more likely when ongoing motion and cue-driven motion combine to point strongly toward the response direction.

Consistent with this account, empirical response probability rose from below 20% to above 70% across the observed drift-score range. A trained 64-unit recurrent neural network reproduced the main geometry without having rotation explicitly imposed.

Across mouse training, intrinsic drift remained comparatively stable while cue-evoked drift strengthened. A phenomenological reinforcement model suggested that reward could preferentially strengthen responses compatible with existing dynamics. This is a proposed learning mechanism, however, rather than an experimentally established causal explanation.

Beyond the Mouse Task

RCA also detected resting rotations in human fMRI and MEG recordings. These accounted for 0.2% and 2.6% of variance, with estimated periods of 51 and 4.4 seconds, respectively.

In a comparison involving 53 controls and 82 participants with early non-affective psychosis, APR was lower in the psychosis group, with split-sample analyses supporting the result. Leading-PC variability did not differ significantly. This suggests a potentially informative dynamical difference, not an established diagnostic test.

An extension detecting two counter-rotating centers distinguished preparatory from post-cue activity in mice and fixation from social-task activity in human fMRI. Combining the mouse preparatory state measured 2.7 seconds before the cue with its cue-onset RCA state improved held-out AUC to 0.677.

A Promising Coordinate, Not a Final Explanation

The study makes a compelling case for looking beyond high-variance signals. Still, these observational reanalyses establish associations, not causal constraints. APR should not automatically be equated with thermodynamic entropy production, and its moment-based estimator has distributional qualifications. Code availability was also deferred until publication.

The central insight remains intriguing: the brain’s readiness to act may depend partly on where it already is—and which way it is moving—within a quiet, low-dimensional flow.

Uncovering low-dimensional representations of brain-wide neural dynamics and their relationship to behavior remains a central challenge. Here, we show that spontaneous brain-wide rotational dynamics constrain cue-evoked behavior and learning. To systematically establish these findings, we developed a dimensionality-reduction technique called Rotational Component Analysis (RCA) that optimizes the detection of low-variance rotational modes even within individual trials. We applied RCA to motor-cortical spike-train data from monkeys and humans, mouse cortex-wide calcium imaging, and human functional magnetic resonance imaging (fMRI) and magnetoencephalography (MEG). In mice, a rotational mode identified from activity recorded at rest predicts cue-evoked behavior: the phase of rotation when a randomly timed cue arrives predicts whether the animal subsequently pulls the lever. We explain this surprisingly strong correlation using both a recurrent neural network (RNN) and a minimal two-dimensional model of response learning, in which the intrinsic rotational mode serves as a dynamical constraint on response and learning. We further identify double-rotation modes containing two centers of rotation that separate preparatory from task dynamics in both mouse and human data. Remarkably, corresponding modes also organize neural dynamics during a non-motor human task, and rotational irreversibility captures early psychosis even when principal component analysis shows no group difference. Our work establishes rotational dynamics as low-dimensional coordinates that constrain how ongoing brain states are converted into behavior. More broadly, RCA provides a versatile tool for detecting hidden nonreciprocal dynamics in biological and physical datasets.