betaFeedback

WorksAgricultural and Biological Sciences

Spontaneous brain-wide rotational dynamics constrain cue-evoked behavior

Abstract

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.

Although neural activity is high-dimensional, the behaviorally relevant part can often be captured by low-dimensional latent dynamics [1, 2, 3, 4, 5, 6]. Rotation is a recurrent motif in such low-dimensional neural dynamics. In this context, rotation arises in cyclic activities produced by central pattern generators and rhythmic motor control [7], but also in non-cyclic actions such as individual reaches [8, 9, 10, 11, 12, 13]. At the brain-wide scale, oscillations are believed to play a central role in cognitive processes [14, 15, 16]. Similarly, rotating and spiral waves propagating through physical brain space have been suggested to coordinate long-range information flows [17, 18]. These elements suggest rotation as a potential organizing principle for brain-wide neural dynamics.

Rotations in ongoing spontaneous neural activity have received far less attention. Spontaneous activity refers to intrinsic neural activity that is not a direct response to external cues [19, 20, 21, 22, 23, 24]. By definition, it cannot be temporally aligned with known causes, making it more challenging to study. Whether rotational dynamics can be robustly identified in spontaneous activity, and if so, whether they can be linked to future behavior, remain open questions.

To address these questions, we introduce rotational component analysis (RCA), an unsupervised dimensionality-reduction method that explicitly prioritizes rotational dynamics. RCA quantifies rotation through the breaking of mirror symmetry in the distribution of neural trajectories and searches for a low-dimensional projection that maximizes a measure of this asymmetry, rather than maximizing explained variance or reconstruction of the full neural activity [25, 26, 27, 28, 29]. Because of its prioritization, RCA can reveal low-variance rotational modes that may be underrepresented by other methods. This feature is particularly important for spontaneous activity, in which uninstructed behavior and other ongoing processes generate large rotation-irrelevant fluctuations that cannot be suppressed by trial averaging. RCA can therefore be applied directly to continuous recordings or pooled individual trials, without event alignment, trial averaging or behavioral labels.

Here, we first benchmark RCA using human and macaque motor-cortical spiking activity during reaching, a setting in which rotational population dynamics are well established. We then apply RCA to mouse cortex-wide calcium imaging, human functional magnetic resonance imaging (fMRI) and human magnetoencephalography (MEG), and succeed in detecting spontaneous rotational dynamics in all three datasets. In mice, the rotational mode identified during rest exhibits an unusually strong association with cue-evoked behavior during task sessions. We propose that the intrinsic rotation serves as a dynamical constraint on both response and learning: cue-evoked responses are more likely, and more readily learned, when they follow the pre-existing rotational flow. We reproduce this relationship in a recurrent neural network (RNN) and a minimal two-dimensional model. In human fMRI, we find that rotational dynamics are weaker in participants with early psychosis than in controls. We further find that double-rotation modes, defined by circulation around two distinct centers, separate preparatory and task dynamics in both mouse and human data. Altogether, our results suggest that rotation provides a brain-wide dynamical coordinate associated with both behavior and learning.

Results

APR and RCA identify rotational irreversibility in reduced phase space

Our goal is to develop a dimensionality-reduction method that projects high-dimensional neural activity onto the latent plane exhibiting the most pronounced directed rotation. For a high-dimensional time series y(t)y(t), consider a candidate plane defined by two orthonormal projection vectors w1\mathbf{w}_1 and w2\mathbf{w}_2, with coordinates x1=w1⋅yx_1 = \mathbf{w}_1 \cdot\mathbf{y} and x2=w2⋅yx_2 = \mathbf{w}_2 \cdot\mathbf{y}. We quantify circulation in this plane using the centered angular current

j=(x1−c1)x˙2−(x2−c2)x˙1,(1)j = (x_1 - c_1)\dot{x}_2 - (x_2 - c_2)\dot{x}_1, \tag*{(1)}

where c=[c1,c2]\mathbf{c} = [c_1,c_2] is the empirical center. This quantity is the planar kinematic analogue of angular momentum, familiar from the orbital motion of a planet around the Sun: its sign distinguishes clockwise from counterclockwise circulation.

Angular current connects rotation to broken mirror symmetry. Dynamics are mirror-symmetric when reflecting their trajectories about an appropriate axis leaves their statistical behavior unchanged. Because reflection about any axis through c\mathbf{c} reverses the sign of jj and the direction of rotation, rotation can be quantified by a measure of dynamical mirror-symmetry breaking (Figure 1A and Extended Data Figure 1A). We refer to this quantity as the asymmetry production rate (APR); the formal definition extends to other spatial transformations (see Methods). In practice, we estimate APR from an ensemble of trajectory segments Γ\Gamma, each spanning two adjacent samples separated by a time interval τ\tau. The accumulated angular current along each segment is Jτ[Γ]=∫0τj(t) dtJ_\tau[\Gamma] = \int_0^\tau j(t)\,\mathrm{d}t, and the estimator is

Figure 1. RCA reveals rotational dynamics in reaching-related spiking activity. A, Schematic of rotational component analysis (RCA). High-dimensional data (e.g., spike-train activity) are projected onto a two-dimensional plane optimized for mirror-symmetry breaking and directed circulation. B, Schematic comparison with principal component analysis (PCA). A low-variance rotational plane (orange) can be missed when the direction of greatest variance is orthogonal to that plane (PC1; blue). The corresponding PCA and RCA projections are shown at right. C–H, Macaque motor-cortical activity during delayed reaching. Condition-averaged trajectories are shown in RCA (orange) and jPCA (blue) planes fitted either to condition-averaged data (C,D) or to individual trials pooled across reach conditions (E,F). Pale curves show all reach conditions, and the bold curve highlights representative maze condition 1. G,H, Normalized cumulative angular current J(t)J(t) across individual trials from maze condition 1 in the single-trial-fit RCA (G) and jPCA (H) planes. Lines show the across-trial mean, and shading denotes ±1\pm1 standard error of the mean (s.e.m.). I–N, Corresponding analyses of human motor-cortical activity during reaching. Condition-averaged trajectories are shown in planes fitted to condition-averaged data (I,J) or pooled individual trials (K,L); pale curves show all eight reach conditions, and the bold curve highlights the representative bottom-target condition. M,N, Trial-level cumulative angular current for the bottom-target condition in the individual-trial-fit RCA (M) and jPCA (N) planes. In all trajectory panels, circles mark trajectory onset and arrowheads indicate temporal direction. The trial-to-trial signal-to-noise ratio (SNR) in G,H,M,N is the absolute mean final current divided by its across-trial standard deviation.

Extended Data Figure 1. Symmetry properties of single- and double-rotational currents. Red and blue indicate opposite directions of circulation. A, A single rotational current (top) is converted to the opposite direction by mirror reflection (middle), thereby breaking mirror symmetry. A half-turn about the center preserves its circulation pattern (bottom). B, An ideal pair of neighboring counter-rotating currents (top). Reflection across the midline exchanges the two currents and reverses their circulation directions, reproducing the original pattern (middle). A half-turn exchanges their positions without reversing their circulation directions and therefore produces a different pattern (bottom).

σ^(τ)=1τ2Jτ‾ 2ΔJτ,(2)\hat{\sigma}(\tau) = \frac{1}{\tau}\frac{2\overline{J_\tau}^{\,2}}{\Delta J_\tau}, \tag*{(2)}

where Jτ‾\overline{J_\tau} and ΔJτ\Delta J_\tau are the mean and variance, respectively, across the ensemble of segments. Persistent directed rotation produces a nonzero mean current, whereas rotation-irrelevant fluctuations increase its variance; σ^\hat{\sigma} therefore balances persistent directional current against rotation-irrelevant fluctuations.

We constructed rotational component analysis (RCA) by searching over pairs of orthonormal projection vectors w1\mathbf{w}_1 and w2\mathbf{w}_2 to maximize σ^\hat{\sigma}. The optimization was performed using gradient descent and automatic differentiation. RCA therefore identifies the plane that maximizes this current-to-fluctuation measure (Figure 1A). This objective differs from that of principal component analysis (PCA), which selects directions of maximal variance. A rotational mode can be missed by PCA when it occupies a low-variance plane approximately orthogonal to the leading principal component (Figure 1B). RCA is unsupervised by construction because no task variables or behavioral labels enter the optimization.

We benchmarked RCA using motor-cortical spiking activity during reaching, a setting in which rotational population dynamics are well established [8]. Previous studies typically extracted these rotations from activity averaged across repeated trials within each reach condition, partly due to the reduced rotation-irrelevant fluctuations in the averaged trajectories [8]. We first reproduced this conventional analysis by fitting RCA and jPCA to condition-averaged activity. jPCA is a PCA-based method that first projects population activity onto a leading principal-component subspace and then extracts rotational dynamics by fitting a skew-symmetric linear dynamical system [8]. Both methods recovered rotational trajectories in the macaque and human datasets (Figure 1C,D,I,J).

To determine whether these rotational modes remained detectable in the presence of trial-to-trial variability, we next fitted RCA and jPCA directly to individual-trial trajectories pooled across reach conditions, without averaging trials before dimensionality reduction (see Methods). For visualization, we projected each trial into the fitted plane and subsequently averaged the projected trajectories within each condition. In the macaque data, the RCA plane fitted to individual-trial activity retained coherent loops across reach conditions, whereas the corresponding jPCA plane did not show coherent rotational organization (Figure 1E,F). In the human data, both planes fitted to individual-trial activity contained visible rotational trajectories (Figure 1K,L), motivating a direct comparison of their consistency across individual trials.

We quantified rotation on each trial using the normalized cumulative angular current

Ji(t)=1R2∫t0tji(s;c) ds,(3)\mathcal{J}_i(t) = \frac{1}{R^2}\int_{t_0}^{t} j_i(s;\mathbf{c})\,\mathrm{d}s, \tag*{(3)}

where ii indexes trials, ji(s;c)j_i(s;\mathbf{c}) is the angular current at time ss, c\mathbf{c} is the empirical center, and R2R^2 is the mean-squared radius across trials and time. We defined the trial-to-trial signal-to-noise ratio (SNR) at the final time tft_f as

SNR=∣⟨Ji(tf)⟩i∣SDi[Ji(tf)].(4)\mathrm{SNR} = \frac{\left|\langle\mathcal{J}_i(t_f)\rangle_i\right|}{\mathrm{SD}_i[\mathcal{J}_i(t_f)]}. \tag*{(4)}

where SDi\mathrm{SD}_i denotes the sample standard deviation across trials. The SNR measures the consistency of rotation across individual trials. For macaque reaching behavior, RCA produced a coherent cumulative current (SNR, 0.89), whereas the jPCA current remained near zero (SNR, 0.05; Figure 1G,H). For human reaching behavior, the current increased in both projections, but was more consistent across trials in the RCA plane (SNR, 1.90) than in the jPCA plane (SNR, 0.98; Figure 1M,N). Thus, RCA better recovered rotational activity directly from non-averaged population activity and identified a shared rotational coordinate that was consistently expressed across individual trials. We show in the Methods that RCA can also identify rotational modes in non-neural datasets (Extended Data Figure 5).

Extended Data Figure 5. RCA beyond neuroscience: recovering rotational modes from noisy traveling patterns and ciliary beating. A–D, A simulated two-mode traveling pattern with weak Ornstein–Uhlenbeck noise (D=10−7D = 10^{-7}). A, Spatial profile u(x,t0)u(x,t_0) at one time point. B, Space–time evolution of u(x,t)u(x,t); diagonal bands indicate traveling waves. C, Spatial weights w1w_1 and w2w_2 of the two RCA coordinates. D, Probability density (color) and probability flux (white arrows) in the standardized RCA plane. E–H, Corresponding analysis after increasing the noise strength to D=0.1D = 0.1. Although the traveling pattern is obscured in the instantaneous profile (E) and space–time map (F), RCA recovers structured spatial weights (G) and a coherent circulating probability flux (H). I–L, RCA analysis of experimentally measured ciliary beating from Geyer et al. [45]. I, Example cilium and centerline representation used to measure tangent angle ψ(s)\psi(s) and curvature along arclength ss. J, Curvature κ(s,t)\kappa(s,t) along normalized arclength s/Ls/L over 100 consecutive frames from an example recording. K, Mean-subtracted RCA weights w1w_1 and w2w_2 along the cilium. L, Probability density (color) and probability flux (white arrows) in the RCA plane, estimated from concatenated 500-frame curvature segments.

A brain-wide rotational mode predicts mouse responses to a random cue

We next applied APR to wide-field calcium imaging of mouse brain activity. We analyzed the BraiDyn-BC dataset, which contains longitudinal recordings from 25 head-fixed mice expressing GCaMP6s in cortical excitatory neurons during rest and during acquisition of a tone-triggered lever-pull task over 15 training days [30]. We used hemodynamics-corrected fluorescence sampled at 30 Hz from 44 regions (22 per hemisphere) registered to the Allen Common Coordinate Framework (Figure 2A, left). The regions span the mouse dorsal cortex, including motor, somatosensory, olfactory, visual, auditory, and retrosplenial areas (Extended Data Figure 2).

Figure 2. Brain-wide rotational mode predicts cue-evoked behavior in mice. A, Analysis overview. Cortex-wide calcium activity was projected by rotational component analysis (RCA), which optimizes broken mirror symmetry, onto a plane with a circulating probability current. The state at cue onset distinguished response from nonresponse trials. This behavioral association was reproduced by a two-dimensional minimal model and a recurrent neural network (RNN). B,C, Resting-state probability density (color) and probability flux (white arrows) in the PCA (B) and RCA (C) planes. Annotations report the fraction of total variance explained by each projection and the estimated rotation period. Translucent white ellipses indicate the standard errors of the mean for the flux vectors; most are smaller than the arrowheads at this plotting scale. D, Lever-pull task. Mice were trained to respond to an auditory cue; a pull during the response window was followed by a water reward. The cue appeared after a random delay of 3–4 s following the preceding lever pull. The dataset comprised 25 mice recorded in 30-min task sessions over 15 training days. E,F, Empirical response probability conditioned on cue-onset state in the PCA (E) and RCA (F) planes. The arrow in F shows the fitted response-gradient direction. G, Response-probability surface fitted in the RCA plane; the arrow again indicates the directional gradient. H, Distribution of response-gradient magnitudes across 2,000 random planes. Vertical lines indicate the PCA (g=0.050g = 0.050), jPCA (g=0.078g = 0.078) and RCA (g=0.151g = 0.151) planes. All projection planes were orthonormalized and their projected coordinates were scaled to unit variance before fitting (Methods).

Extended Data Figure 2. Anatomical organization of the mouse resting-state rotational mode. A, RCA loading vectors w1w_1 and w2w_2 across the 44 recorded Allen Common Coordinate Framework regions (22 per hemisphere). Blue and orange bars indicate positive loadings for w1w_1 and w2w_2, respectively, whereas gray bars indicate negative loadings; suffixes l and r denote the left and right hemispheres. B, Directed interregional structure implied by the RCA loadings. We constructed the minimum-Frobenius-norm antisymmetric matrix A=2sgn⁡(ω)(w1w2T−w2w1T)/(∥w1∥2+∥w2∥2)A = 2\operatorname{sgn}(\omega)(w_1w_2^{\mathsf{T}} - w_2w_1^{\mathsf{T}})/(\lVert w_1\rVert^2 + \lVert w_2\rVert^2), which represents a unit rotation in the observed temporal direction in the original region space; ω\omega is the signed rotation coefficient in the RCA plane. Region color denotes node strength ∑j∣Aij∣\sum_j \lvert A_{ij}\rvert, and arrows show the 20 strongest pairwise couplings, oriented such that Aij>0A_{ij} > 0. Numbers identify the regions listed at right; anterior is at the top. This construction visualizes the interregional coupling pattern implied by the projection and is not an estimate of anatomical connectivity.

We first compared resting-state probability flux in the PCA and RCA planes (Figure 2B,C). In the PCA plane, where x1x_1 and x2x_2 are the first two principal components, the probability flux showed only weak circulation (Figure 2B). By contrast, the RCA plane revealed a strong clockwise probability current (Figure 2C). This circulation was identified from spontaneous activity and therefore did not depend on sensory stimulation or task events. Notably, the RCA plane accounted for only 0.6% of the total variance, compared with 87% for the PCA plane, highlighting how a rotational mode can be overlooked by conventional dimensionality-reduction methods. The regional composition of this mode and the minimal antisymmetric coupling pattern show loadings across the cortex, with large values in motor and somatosensory areas (Extended Data Figure 2). This latent rotation coordinates distributed brain regions that need not be spatially adjacent, distinguishing it from topographic rotating waves [18].

We next asked whether the rotational mode identified at rest was related to behavior during task sessions. An auditory cue initiated a response window in which a successful lever pull was followed by a water reward (Figure 2D). Each cue occurred after a random delay of 3–4 s following the preceding lever pull. This randomized timing sampled different states of the ongoing cortical dynamics across trials, allowing us to test whether the cue-onset state predicted task outcome. We defined Prsp(xcue)P^{\mathrm{rsp}}(x^{\mathrm{cue}}) as the probability that the mouse responds to a cue, conditioned on its state xcuex^{\mathrm{cue}} in the reduced plane at cue onset. Response probability varied only weakly and irregularly across the PCA plane (Figure 2E), but showed a strong directional dependence in the RCA plane: cues that occurred when cortical activity occupied the upper-right portion of the plane were followed by responses more often than cues that occurred when activity occupied the lower-left portion (Figure 2F). We described this pattern as a radially symmetric baseline plus a directional linear component,

Prsp(xcue)=P0(1−c0∥xcue∥)+g⋅xcue,(5)P^{\mathrm{rsp}}(x^{\mathrm{cue}}) = P_0(1-c_0\lVert x^{\mathrm{cue}}\rVert) + \boldsymbol{g}\cdot x^{\mathrm{cue}}, \tag*{(5)}

where xcuex^{\mathrm{cue}} is the cue-onset state and P0P_0, c0c_0 and g\boldsymbol{g} were fitted to the empirical data. The fit captured both the directional increase and the weaker radial dependence of response probability (Figure 2G). Here, g\boldsymbol{g} is the response gradient and points in the direction of increasing response probability.

To determine whether this predictive relationship was unusual among two-dimensional projections, we compared the magnitude of the response gradient with those obtained from 2,000 random planes (Figure 2H). The RCA gradient lay near the extreme upper tail of the random-plane distribution and was substantially larger than the PCA and jPCA gradients. Notably, the RCA plane was selected using only resting-state activity, without access to task or behavioral information. Thus, a spontaneous rotational mode identified without behavioral supervision nevertheless provided a strong neural correlate of subsequent cue-evoked behavior. We further asked whether this predictive relationship could be explained by arousal, for which pupil diameter provides a widely used proxy. On held-out sessions, the cue-onset RCA state predicted trial outcome better than pupil-derived arousal (Extended Data Figure 3).

Response probability conditioned on pupil-derived arousal and held-out receiver operating characteristic curves

Extended Data Figure 3. Response prediction from the RCA state is not explained by pupil-indexed arousal. A, Observed response probability PrspP^{\mathrm{rsp}} conditioned on the pupil-derived arousal zz score at cue onset. Points show binned observations from all matched trials, error bars indicate binomial standard errors and the black curve shows a cubic polynomial fit. This descriptive all-data fit was not used for held-out evaluation. B, Receiver operating characteristic curves evaluated on trials from held-out sessions for models using the cue-onset RCA response score (blue; AUC = 0.611), a cubic pupil-arousal model (orange; AUC = 0.535), or both predictors (green; AUC = 0.611). Models were trained on 35,547 trials from 267 sessions and evaluated on 12,272 trials from 89 held-out sessions. The diagonal indicates chance performance.

Minimal and recurrent-network models explain state-dependent responses

To identify a heuristic dynamical mechanism linking cue-onset state to behavior, we first constructed a two-dimensional minimal model of brain activity (Figure 3A):

Figure 3. Minimal dynamical model and an RNN reproduce the state-dependent response rule. A, Mouse minimal model. Pre-cue activity follows the intrinsic rotational drift fpref^{\mathrm{pre}}; adding the cue-evoked response drift frspf^{\mathrm{rsp}} gives the post-cue drift fpostf^{\mathrm{post}}. B, Empirical pre-cue drift field fpref^{\mathrm{pre}} in the rotational plane. C, Empirical cue-evoked response-drift field frspf^{\mathrm{rsp}}. Basis arrows define the response direction n^∥\hat{n}_{\parallel} and the orthogonal direction n^⊥\hat{n}_{\perp}. D, Response probability PrspP^{\mathrm{rsp}} (color) conditioned on the cue-onset state, overlaid with the post-cue drift component, f∥post=f∥postn^∥f^{\mathrm{post}}_{\parallel}=f^{\mathrm{post}}_{\parallel}\hat{n}_{\parallel} (black arrows). E, Trial-averaged trajectories (dark curves) projected along n^∥\hat{n}_{\parallel} and aligned to cue onset, separated by response outcome; translucent curves denote trajectories of each individual mouse. F, Response probability as a function of the post-cue drift component f∥postf^{\mathrm{post}}_{\parallel}. Points show binned data and the black curve shows the fitted response rule. G, Recurrent neural network (RNN) trained to produce a graded response following a randomly timed cue. The shaded region indicates the response window. H, Pre-cue rotational drift field generated by the RNN. I, Cue-evoked response-drift field generated by the RNN. J, RNN response strength vrspv^{\mathrm{rsp}} (color) conditioned on the cue-onset state, overlaid with f∥postn^∥f^{\mathrm{post}}_{\parallel}\hat{n}_{\parallel} (black arrows). K, Trial-averaged RNN trajectories (dark curves) projected along n^∥\hat{n}_{\parallel} and aligned to cue onset, separated into high- and low-response trials; translucent curves denote 10 example trials. L, Response strength as a function of f∥postf^{\mathrm{post}}_{\parallel}. Points show binned simulations and the black curve shows the fit. In B,C,H,I, arrows indicate the direction and magnitude of the corresponding drift fields.

dxt=f(xt,t) dt+dWt,(6)\mathrm{d}x_t = f(x_t,t)\,\mathrm{d}t + \mathrm{d}W_t, \tag*{(6)}

where f(xt,t)f(x_t,t) is the state- and time-dependent drift and WtW_t is a two-dimensional Wiener process representing stochastic fluctuations. Cue onset changes the drift field: before the cue, activity follows the pre-cue (i.e., intrinsic) drift fpre\boldsymbol{f}^{\mathrm{pre}}, whereas after the cue it follows the post-cue drift fpost\boldsymbol{f}^{\mathrm{post}}. We decomposed the latter into the intrinsic drift and a cue-evoked response component,

fpost(x)=fpre(x)+frsp(x).(7)\boldsymbol{f}^{\mathrm{post}}(x) = \boldsymbol{f}^{\mathrm{pre}}(x) + \boldsymbol{f}^{\mathrm{rsp}}(x). \tag*{(7)}

We estimated the empirical pre-cue and post-cue drift fields over windows of ten and five frames, respectively. The pre-cue field preserved the clockwise rotation identified at rest (Figure 3B). By contrast, the cue-evoked response drift pointed approximately along a common direction n^∥\hat{\mathbf{n}}_{\parallel}, whereas its magnitude varied primarily along the orthogonal direction n^⊥\hat{\mathbf{n}}_{\perp} (Figure 3C). We captured this organization with the two-dimensional vector field:

frsp(x)=f0rsp(1+c x⋅n^⊥)n^∥.(8)\mathbf{f}^{\mathrm{rsp}}(\mathbf{x}) = f_{0}^{\mathrm{rsp}}\left(1+c\,\mathbf{x}\cdot\hat{\mathbf{n}}_{\perp}\right)\hat{\mathbf{n}}_{\parallel}. \tag*{(8)}

To relate this drift geometry to behavior, we evaluated the post-cue drift at each cue-onset state and projected it onto the response direction, f∥post=fpost(xcue)⋅n^∥f_{\parallel}^{\mathrm{post}}=\mathbf{f}^{\mathrm{post}}(\mathbf{x}^{\mathrm{cue}})\cdot\hat{\mathbf{n}}_{\parallel}.

The corresponding vector component is f∥post=f∥postn^∥f^{\mathrm{post}}_{\parallel}=f^{\mathrm{post}}_{\parallel}\hat{n}_{\parallel}. The scalar f∥postf^{\mathrm{post}}_{\parallel} increased predominantly along the transverse direction n^⊥\hat{n}_{\perp} (Figure 3D). This transverse gradient is expected from the intrinsic rotational flow: its response-direction component, f∥pre=fpre⋅n^∥f^{\mathrm{pre}}_{\parallel}=f^{\mathrm{pre}}\cdot\hat{n}_{\parallel}, is positive in one half of the plane and negative in the other. Notably, the behavioral gradient g\mathbf{g} was also approximately aligned with n^⊥\hat{n}_{\perp}, and response probability was higher at states with larger f∥postf^{\mathrm{post}}_{\parallel} (Figure 3D). Thus, the response can be predicted from f∥postf^{\mathrm{post}}_{\parallel}, and cues arriving at different phases of the rotation encounter different response probabilities.

We tested this prediction using cue-aligned trajectories from response and nonresponse trials (Figure 3E). Both groups showed an abrupt increase in displacement along n^∥\hat{n}_{\parallel} during the post-cue period, but the increase was substantially larger on response trials. Consistent with the model, empirical response probability increased monotonically from below 20% to above 70% across the observed range of f∥postf^{\mathrm{post}}_{\parallel} (Figure 3F). This dependence was well described by a sigmoid:

prsp(f∥post)=Plow+(Phigh−Plow)×[1+e−λ(f∥post−s0)]−1.(9)p^{\mathrm{rsp}}\left(f_{\parallel}^{\mathrm{post}}\right)=P_{\mathrm{low}}+\left(P_{\mathrm{high}}-P_{\mathrm{low}}\right)\times\left[1+e^{-\lambda\left(f_{\parallel}^{\mathrm{post}}-s_{0}\right)}\right]^{-1}. \tag*{(9)}

Thus, the interaction between the rotational drift and the cue-evoked drift provides a compact account of the state-dependent response probability.

We next asked whether a similar relation between response and rotation could also emerge in a simpler artificial system: a recurrent dynamical system trained to generate a randomly cued response. We used a recurrent neural network (RNN) with dynamics

ht+1=(1−α)ht+αtanh⁡(Wrecht+Winut+b)+2α σ ξt,(10)h_{t+1}=(1-\alpha)h_t+\alpha\tanh\left(W_{\mathrm{rec}}h_t+W_{\mathrm{in}}u_t+b\right) +\sqrt{2\alpha}\,\sigma\,\xi_t, \tag*{(10)}

where hth_t is the network state, WrecW_{\mathrm{rec}} is the recurrent weight matrix, WinW_{\mathrm{in}} maps the scalar cue input utu_t into the network, bb is the bias vector, α\alpha is the leak parameter, and σ\sigma is the noise amplitude with ξt∼N(0,I)\xi_t\sim\mathcal{N}(0,I). The model reproduced the essential temporal structure of the mouse task [30] (Figure 3G; see Methods): a pulse in utu_t at a random time represented the randomly timed auditory cue, and the fixed input weights WinW_{\mathrm{in}} specified how that cue entered the network. The hidden state hth_t represented population activity, whose internal dynamics were governed by WrecW_{\mathrm{rec}} and bb. Activity was read out along a fixed direction WoutW_{\mathrm{out}}, and response strength was defined by the normalized net output displacement from the beginning to the end of the fixed post-cue response window. Thus, WinW_{\mathrm{in}} and WoutW_{\mathrm{out}} provided fixed input and output mappings, whereas training by backpropagation modified only WrecW_{\mathrm{rec}} and bb to increase response strength.

After training, we applied RCA to the spontaneous network dynamics in the absence of a cue and estimated the intrinsic and cue-evoked drift fields as in the mouse data. The RCA plane showed a pre-cue rotation and a cue-evoked drift oriented approximately along a common direction n^∥\hat{n}_{\parallel} (Figure 3H,I). Response strength varied directionally across cue-onset states. As in the mouse data, overlaying f∥postf_{\parallel}^{\mathrm{post}} on the response-strength map showed that larger responses occurred in states with larger f∥postf_{\parallel}^{\mathrm{post}}; the response-strength gradient was approximately perpendicular to n^∥\hat{n}_{\parallel} (Figure 3J). Moreover, high- and low-response trials followed distinct trajectories along n^∥\hat{n}_{\parallel} (Figure 3K), and response strength increased with f∥postf_{\parallel}^{\mathrm{post}} (Figure 3L), qualitatively reproducing the relationships observed in mice (Figure 3E,F). Thus, a recurrent system with no explicitly imposed rotational structure can reproduce the key geometry linking spontaneous flow and cue-onset state to response strength, as observed in mice.

Learning strengthens the cue-evoked response drift

The lever-pull response is a learned rather than an innate mouse behavior. We therefore asked how the response drift changed as the behavior was acquired. We inferred fpref^{\mathrm{pre}} and frspf^{\mathrm{rsp}} in five nonoverlapping three-day windows spanning the 15 training days. The intrinsic drift fpref^{\mathrm{pre}} was largely unchanged, whereas the cue-evoked field frspf^{\mathrm{rsp}} showed a general increase in magnitude (Extended Data Figure 4). This is evident in the amplitude parameter f0rspf_{0}^{\mathrm{rsp}}, which shows an overall increasing trend across training (Figure 4A). In parallel, the distribution of f∥postf_{\parallel}^{\mathrm{post}} shifted toward the response-generating regime during learning (Figure 4B), whereas the conditional drift-to-response mapping prsp(f∥post)p^{\mathrm{rsp}}(f_{\parallel}^{\mathrm{post}}) remained similar across training windows (Figure 4C). A cross-validated model that held this response rule fixed while allowing the inferred response drift to vary captured most of the improvement in observed performance (Figure 4D). These results suggest that learning primarily alters the cue-evoked dynamics, and hence the distribution of f∥postf_{\parallel}^{\mathrm{post}}, while preserving the mapping from this dynamical variable to response probability.

Figure 4. Learning strengthens the cue-evoked response drift within an existing rotational coordinate. A–D, Learning across 15 task days. A, Inferred response-drift amplitude f0rspf_{0}^{\mathrm{rsp}} in successive three-day windows. The dashed line indicates the estimate from all task days. B, Distribution of f∥cuef_{\parallel}^{\mathrm{cue}} in each window; horizontal bars indicate medians. C, Empirical response probability as a function of f∥postf_{\parallel}^{\mathrm{post}} for each training window. Symbols and error bars show binned estimates and their uncertainty. D, Observed response probability and the cross-validated probability predicted from window-specific drift estimates and a fixed response rule. Symbols and error bars show estimates and their uncertainty. E–G, Proposed learning mechanism. E, Intrinsic drift before learning (frsp=0f^{\mathrm{rsp}}=0). The diagonal boundary separates states with relatively high (“easy”) and low (“hard”) response probability; the arrow indicates the rewarded direction. F, Cue-onset states and drifts from randomly sampled trials, colored by reward outcome. The dashed line is the optimal linear boundary separating rewarded and unrewarded trials. G, Simulated response drift generated by reinforcing rewarded states and drifts; arrows indicate direction and color denotes ∣frsp∣\lvert f^{\mathrm{rsp}}\rvert.

Extended Data Figure 4. Intrinsic and cue-evoked drift fields across learning. A, Pre-cue intrinsic drift field fpref^{\mathrm{pre}} estimated independently in five nonoverlapping three-day windows spanning the 15 task days. Arrows indicate drift direction and magnitude, and color denotes ∣fpre∣\lvert f^{\mathrm{pre}}\rvert. B, Cue-evoked response drift field frspf^{\mathrm{rsp}} estimated in the same training windows. Arrows indicate drift direction and magnitude, and color denotes ∣frsp∣\lvert f^{\mathrm{rsp}}\rvert.

During learning, the direction of the transverse gradient of frspf^{\mathrm{rsp}} remained fixed, which is consistent with the intrinsic rotational dynamics. This gradient was not captured by the RNN (Figure 3I). This difference may reflect the distinct learning signals available to the two systems. The RNN was trained by backpropagation using an explicit, differentiable objective, whereas the mice presumably received sparse trial-level reward feedback. To explain how sparse reward feedback could produce this, we constructed a minimal reinforcement model.

Before a learned response drift develops, the intrinsic rotation alone occasionally carries activity through states whose velocity points along the rewarded response direction (Figure 4E). Because cues occur at random times, they sample both favorable and unfavorable phases of the rotation, producing rewarded and unrewarded trials. The states and drifts associated with reward therefore form a biased subset of all samples in the rotational plane (Figure 4F). Reinforcing this subset could progressively shape a state-dependent cue-evoked response field. We represented that field in the minimal form

frsp(x)=f0rsp(1+k⋅x),(11)f^{\mathrm{rsp}}(x)=f_{0}^{\mathrm{rsp}}(1+k\cdot x), \tag*{(11)}

where f0rspf_{0}^{\mathrm{rsp}} sets the mean response drift, f0rsp=∥f0rsp∥f_{0}^{\mathrm{rsp}}=\lVert f_{0}^{\mathrm{rsp}}\rVert denotes its amplitude and kk determines its state dependence. For each rewarded sample with state xsx^{s} and drift fsf^{s}, we updated these parameters according to

f0rsp←(1−a1)f0rsp+b1fs,k←(1−a2)k+b2xs.(12)f_{0}^{\mathrm{rsp}}\leftarrow(1-a_{1})f_{0}^{\mathrm{rsp}}+b_{1}f^{s},\qquad k\leftarrow(1-a_{2})k+b_{2}x^{s}. \tag*{(12)}

The decay terms stabilized the parameters, whereas the reward-gated terms reinforced the drifts and locations associated with successful trials. Simulating this rule with the inferred pre-cue dynamics produced an anisotropic response field resembling that observed in mice (compare Figure 4G to Figure 3C; see Methods). In this model, learning does not create a new dynamical coordinate. Instead, it strengthens a cue-evoked drive within the coordinate supplied by spontaneous rotation, while the intrinsic flow constrains which states and drifts are reinforced. The model thus provides a possible account of how animals can acquire the response without access to an explicit cost function.

Rotational currents generalize to humans and detect early psychosis

We next tested whether rotational modes can be generalized beyond mouse cortex and calcium imaging. We analyzed resting-state recordings from a large cohort of healthy young adults in the Human Connectome Project Young Adult (HCP-YA) dataset [31], which includes whole-brain fMRI and MEG data. Applying RCA to the HCP-YA fMRI recordings revealed a rotational current in a projection of cortical activity (Figure 5A,B). HCP-YA MEG recordings likewise contained a coherent rotational current (Figure 5C,D), despite MEG’s distinct spatial resolution and much faster sampling rate. The recurrence of organized currents in mouse calcium imaging, human fMRI and human MEG indicates that symmetry-guided projection is not restricted to one species, recording modality or timescale. The inferred rotation periods differed markedly, from 51 s in fMRI to 4.4 s in MEG, consistent with the distinct temporal resolutions and characteristic signal dynamics of the two modalities. Despite this difference, the rotational planes accounted for only 0.2% of the variance in fMRI and 2.6% in MEG, in agreement with the low-variance rotational mode identified in mouse cortex.

Figure 5. Brain-wide rotational dynamics differ in early psychosis. A, Schematic of human fMRI acquisition and illustrative orthogonal activation maps. B, Resting-state probability density (color) and probability flux (white arrows) in the RCA plane for the HCP Young Adult (HCP-YA) fMRI dataset. The annotation reports the fraction of total variance explained by the RCA plane and the estimated rotation period. C, Schematic of human MEG acquisition with an illustrative sensor topography and multichannel evoked responses. D, Resting-state probability density and probability flux in the RCA plane for the HCP-YA MEG dataset. The annotation reports the fraction of total variance explained and the estimated rotation period. E,F, Participant-level PCA variability, defined as the standard deviation along the leading principal axis (E), and RCA asymmetry production rate (APR; F) in control participants and participants with early psychosis from the HCP Early Psychosis (HCP-EP) dataset. Points indicate participants, violins show the group distributions, and black symbols and bars show group summaries. Brackets indicate group-comparison PP values. Scale arrows in B and D give the vector scale.

We then asked whether rotational irreversibility could act as a probe for psychiatric disorders. For this comparison, we used resting-state fMRI from the HCP Early Psychosis (HCP-EP) dataset, which includes control participants and participants with early psychosis [32]. Variability along the leading principal component did not differ significantly between groups (P=0.124P = 0.124; Figure 5E). By contrast, the APR of the RCA projection was markedly lower in the early-psychosis group (P=5.22×10−7P = 5.22 \times10^{-7}; Figure 5F). Participant-level split-sample analyses confirmed that reduced APR was robust to the group used for projection learning, whereas PCA variability showed at most a modest group difference (Extended Data Figure 7; see Methods). The rotational measure therefore revealed a group difference that was not apparent from PCA variability alone: reduced brain-wide rotational dynamics were associated with early psychosis.

Extended Data Figure 7. Split-sample analyses of reduced rotational irreversibility in early psychosis. A,B, Control-trained analysis. RCA and PCA projections were learned from 26 randomly selected control participants and evaluated in the remaining 27 controls and all 82 participants with non-affective psychosis. Participant-level standard deviation (STD) along the learned leading principal component did not differ between groups (A; P=0.592P = 0.592), whereas APR in the learned RCA plane was lower in participants with psychosis (B; P=0.000252P = 0.000252). C,D, Disease-trained analysis. Projections were learned from 41 randomly selected participants with psychosis and evaluated in the remaining 41 participants and all 53 controls. PCA standard deviation showed only a modest, single-asterisk difference (C; P=0.0275P = 0.0275), whereas APR showed a more pronounced group difference (D; P=8.29×10−5P = 8.29 \times10^{-5}). Points indicate participants, violins show the group distributions, and black symbols and error bars indicate means ±\pm s.e.m. Group comparisons used two-sided Welch’s tt-tests; ns, not significant.

Double-rotation modes separate preparation and predict future responses

We further generalized RCA to detect more complex dynamical modes. Here, we targeted double-rotational dynamics, characterized by oppositely directed circulation around neighboring centers, using double-rotational component analysis (DRCA). An ideal pair of counter-rotating currents can be mirror-symmetric. By contrast, the pair breaks half-turn symmetry: a 180° rotation exchanges the two cycles but preserves chirality and therefore does not reproduce the original current pattern (Extended Data Figure 1B). Consistent with this distinction, under a half-turn about the empirical center, the ordinary angular current jj is unchanged, whereas the spatially weighted current j∗=(x1−c1)jj^{*}=(x_{1}-c_{1})j changes sign. We therefore detect double-rotation modes by optimizing

σ^double=1τ2Jτ∗ΔJτ∗(13)\hat{\sigma}_{\mathrm{double}}=\frac{1}{\tau}\frac{2J_{\tau}^{*}}{\Delta J_{\tau}^{*}} \tag*{(13)}

where c=[c1,c2]c=[c_{1},c_{2}] is the empirical center. The factor x1−c1x_{1}-c_{1} changes sign across the midline, causing counter-rotating currents on the two sides to contribute with the same sign rather than cancel in the ordinary angular current.

Applied to mouse task activity, DRCA revealed a probability flux with two distinct centers of circulation (Figure 6A). Cue-aligned trajectories showed that the two centers organized different periods of each trial: during the pre-cue period, activity circulated around the lower-right center, whereas during the post-cue period it followed a broader loop around the upper-left center (Figure 6B). The double-rotation mode therefore separated ongoing preparatory activity from post-cue and response-related dynamics.

Figure 6. Double-rotation modes separate preparatory and task-related dynamics. A, Mouse task probability density (color) and probability flux (white arrows) in the double-rotational component analysis (DRCA) plane. The annotation reports the fraction of total variance explained. B, Pre-cue (blue) and post-cue (orange) mouse trajectories aligned to cue onset. Translucent curves show trial-averaged trajectories for individual mice, thick curves show group averages, and arrowheads indicate temporal direction. C, Mouse response probability conditioned on the DRCA state 2.7 s pre-cue; gray bins were excluded because of insufficient occupancy. D, Receiver operating characteristic curves for classifiers evaluated on held-out sessions using the cue-onset RCA state (AUC = 0.638), pre-cue DRCA state (AUC = 0.612), or their combination (AUC = 0.677). State–response mappings and classifiers were fitted using separate training sessions. The diagonal denotes chance performance. E, Probability density and probability flux for the DRCA mode identified in the HCP-YA human fMRI social task. The annotation reports the fraction of total variance explained. F, Block-aligned trajectories during task (orange) and fixation (blue). Light curves show block-averaged trajectories for 10 randomly selected participants, thick curves show group averages, circles mark aligned time points, and arrowheads indicate temporal direction.

The preparatory branch of the double rotation carried information about the animal’s future response. Response probability depended strongly on the projected cortical state 80 frames, or 2.7 s, pre-cue (Figure 6C). Because cue timing was randomized, this dependence cannot be attributed to a deterministic countdown to a predictable cue time. Using separate training and testing data, the cue-onset state in the single-rotation RCA mode predicted response with an area under the receiver operating characteristic curve (AUC) of 0.638, whereas the pre-cue DRCA state achieved an AUC of 0.612. Combining the two states increased the AUC to 0.677 (Figure 6D), indicating that the preparatory DRCA mode contained information complementary to that provided by the RCA state at cue onset (see Methods for details). We also detected a double-rotation mode in fMRI recorded during the HCP-YA social task (Figure 6E). The paradigm alternated between task blocks, in which participants made a judgment after viewing a video, and fixation blocks. Block-aligned trajectories occupied different parts of the double rotation: fixation dynamics circulated around one center, whereas task dynamics circulated around the other (Figure 6F). Thus, double-rotation modes separated behaviorally defined epochs in both mouse cortex-wide calcium activity and human fMRI.

Discussion

Our results identify broken mirror symmetry as a useful principle for extracting behaviorally meaningful dynamics from brain-wide recordings. RCA and DRCA consistently recovered rotational modes across species and recording modalities. These rotations are defined in a latent phase space, complementing previous observations of brain-wide rotating waves in physical space [17, 18]. Although these rotational dynamics do not rely on external signals, their associations with behavior and learning, together with their alteration in early psychosis, suggest that rotational flow is a prevalent low-dimensional motif of brain-wide dynamics with functional and clinical relevance.

An important distinction is the space in which rotation is defined. Yet et al. [18] detected topographic waves in which oscillatory phase winds around a center in the physical cortical sheet, allowing the location, radius, direction and speed of individual wave events to be measured. By contrast, RCA detects directed circulation of the brain-wide activity trajectory in a learned latent state space whose axes are weighted combinations of recorded regions or sensors. It therefore does not require dense spatial sampling, a predefined frequency band or a spatial wave template, and can be applied across calcium imaging, fMRI and MEG datasets. A loop in this latent space does not by itself imply that activity propagates as a physical wave. The two forms of rotation are nevertheless compatible: a coherent rotating wave could generate a loop in population state space, whereas latent circulation could also arise from sequential interactions among distributed regions without forming a spatial vortex.

Why might rotation recur so broadly? Here we consider two possibilities. First, cortical circuits may reuse cyclic dynamical motifs that support central pattern generators and rhythmic movement [8, 33]. Such motifs could have been adapted to organize more complex computations, including discrete, non-rhythmic actions. Second, rotation is among the simplest flows capable of sustaining a nonzero steady-state current and breaking detailed balance. In a linear system, nonreciprocal coupling between only two macroscopic degrees of freedom can generate rotational drift, making rotation a natural outcome of high-dimensional recurrent dynamics. Consistent with this possibility, the trained RNN developed a spontaneous rotation even though no rotational structure was explicitly imposed.

The analysis of mouse cortical activity suggests how continuous rotational dynamics can shape a discrete, non-rhythmic behavioral response. Rotations in motor cortex are well established during individual reaches, where they have been interpreted as structured population dynamics for generating movement [8, 12, 13]. In our analysis, the rotation was already present in spontaneous cortex-wide activity before the cue. Random cue timing therefore sampled different states and directions of this ongoing flow. Because a dynamical state specifies not only the current activity pattern but also its direction of evolution, the cue-evoked drive could combine with the intrinsic flow to make a response-generating trajectory more or less accessible. According to this view, the cue gates an event-triggered readout of a continuously evolving latent process: neural activity may cycle, whereas behavior occurs only when the cue and ongoing flow jointly carry the system toward a response-producing regime. The rotational state thus provides a coordinate of dynamical readiness, converting structured ongoing dynamics into trial-to-trial variation in a discrete outcome without requiring rhythmic behavior. The organization of the cue-evoked drift field further suggests that the intrinsic rotation may constrain how the response is learned: reinforcement could preferentially strengthen cue-evoked drives that cooperate with the existing flow, rather than requiring the cortex to construct a new trajectory that opposes it. This interpretation is consistent with evidence that animals more readily learn activity patterns within existing neural manifolds and have difficulty producing trajectories that violate intrinsic dynamics, suggesting that learning preferentially follows existing neural structures [34, 35]. Our findings extend this idea from constraints on producing neural trajectories to constraints on learning an input–response mapping: a new behavior may be acquired by reshaping cue-evoked inputs within an existing dynamical scaffold rather than by rewriting the intrinsic dynamics.

The reaching benchmarks illustrate where RCA differs from conventional dimensionality-reduction methods (Figure 1). In the macaque data, coherent circulation was evident only in the RCA plane. Both RCA and jPCA revealed rotation in condition-averaged human activity, but RCA produced a more consistent individual-trial current. This advantage is expected when trial-specific or other rotation-independent variation obscures a low-variance rotational mode. Whereas PCA prioritizes variance and jPCA first restricts the analysis to leading principal components, RCA directly optimizes the strength and consistency of angular current. APR is related to the thermodynamic language of irreversibility but should not be equated with the entropy production rate (EPR). EPR quantifies all detectable time-reversal asymmetry. Current-based APR isolates the component expressed through a current that is odd under both time reversal and the chosen spatial transformation. This construction targets rotational handedness while retaining a quantitative connection to nonequilibrium probability currents [25, 26, 27].

Neuronal activity recorded immediately before trial onset (i.e., spontaneous) has been shown to nonetheless influence post-trial activity [24, 36], and even bias behavioral performance [37]. As discussed above, RCA is a dimensionality reduction technique that can identify within-trial rotational dynamics. Analysis of the mouse cortical activity during the tone-detection/lever-pulling task shows that the position of rotational dynamics before cue onset determined (in part) behavioral response (Figure 2F). The simple two-dimensional latent description that RCA provides greatly simplifies the connection from spontaneous to trial-evoked neural activity, compared to past analyses [24, 37].

Together, our results support a view in which ongoing brain-wide activity is not merely background variability around an evoked response. Its directed flow defines coordinates that shape how an input is converted into action, how rewarded dynamics are reinforced during learning and how preparatory and task-related states are organized. By optimizing for broken dynamical symmetry rather than variance, RCA and DRCA provide a route from high-dimensional recordings to low-dimensional flows with direct behavioral meaning. Although developed here for brain activity, the framework is broadly applicable to high-dimensional data and could uncover hidden nonreciprocal dynamics across diverse physical and biological systems [38, 39] such as epigenetic landscapes [40].

Online Methods

Study design

This study reanalyzed existing datasets comprising mouse calcium imaging, human fMRI, human MEG, human intracortical electrophysiology and monkey electrophysiology. No new animals or human participants were recruited. Sample sizes were determined by the available datasets and were not set by an a priori power calculation. Ethical approval and consent for the original experiments are described in the corresponding data publications and repository documentation.

Current-based APR and trajectory-level bounds

Let Γ={x(t):0≤t≤τ}\Gamma= \{x(t) : 0 \le t \le\tau\} denote a trajectory segment of duration τ\tau with path distribution PP. We seek symmetry breaking carried by temporal evolution rather than by the stationary distribution alone. We therefore choose a spatial transformation M\mathcal{M} according to the dynamical pattern of interest and ask whether trajectory currents distinguish PP from its transformed distribution.

To formalize this idea, let J[Γ]J[\Gamma] be a real-valued trajectory current. We require the current to be additive under concatenation, J[Γ1∘Γ2]=J[Γ1]+J[Γ2]J[\Gamma_1 \circ\Gamma_2] = J[\Gamma_1] + J[\Gamma_2], and to satisfy

J[MΓ]=−J[Γ],J[TΓ]=−J[Γ].(14)J[\mathcal{M}\Gamma] = -J[\Gamma], \qquad J[\mathcal{T}\Gamma] = -J[\Gamma]. \tag*{(14)}

The first condition makes the current sensitive to breaking of the chosen spatial symmetry. The second makes it odd under the time-reversal transformation T\mathcal{T}. We denote the class of additive currents satisfying these two parity conditions by CM\mathcal{C}_{\mathcal{M}}. These requirements define an admissible class but do not select a unique current.

Let pJ(j)p_J(j) denote the distribution of J[Γ]J[\Gamma] for trajectories drawn from PP. For each admissible current, we define a current-specific APR and then define the APR associated with M\mathcal{M} by optimizing over the admissible class:

σAPR(M,J)=1τDKL[pJ(j) ∥ pJ(−j)],\sigma_{\mathrm{APR}}^{(\mathcal{M},J)} = \frac{1}{\tau}D_{\mathrm{KL}}\left[p_J(j)\,\middle\|\,p_J(-j)\right],
σAPRM=sup⁡J∈CMσAPR(M,J).(15)\sigma_{\mathrm{APR}}^{\mathcal{M}} = \sup_{J\in\mathcal{C}_{\mathcal{M}}}\sigma_{\mathrm{APR}}^{(\mathcal{M},J)}. \tag*{(15)}

If the dynamics are invariant under M\mathcal{M}, the first parity condition requires pJ(j)=pJ(−j)p_J(j) = p_J(-j). The same equality follows from time-reversal symmetry and the second condition. A positive APR therefore requires both spatial-symmetry breaking and temporal irreversibility.

A well-chosen current can therefore provide a tight lower bound on the maximal APR. In practice, both the spatial transformation M\mathcal{M} and the current JJ should be chosen according to the dynamical pattern of interest; the construction is nontrivial only when CM\mathcal{C}_{\mathcal{M}} contains a nonzero current. Extended Data Figure 1 illustrates the two cases studied here. A single directed rotation breaks reflection symmetry about every axis through its center, making the angular current in (1) a natural choice because it changes sign under each such reflection. By contrast, an ideal pair of counter-rotating vortices can preserve the reflection that exchanges the two vortices while breaking half-turn symmetry about cc. The modified current J∗=(x1−c1)JJ^* = (x_1-c_1)J detects this pattern: it changes sign under the half-turn x↦2c−xx \mapsto2c-x but remains invariant under the reflection x1−c1↦−(x1−c1)x_1-c_1 \mapsto-(x_1-c_1).

The current-based APR is bounded by the corresponding asymmetries of the full trajectory distribution. Define the transformed distributions PM=M#PP^{\mathcal{M}}=\mathcal{M}_{\#}P and PT=T#PP^{\mathcal{T}}=\mathcal{T}_{\#}P. The data-processing inequality [41] and (14) give

0≤σAPR(M,J)≤σAPRM≤min⁡{DKL(P∥PM)τ,DKL(P∥PT)τ}.(16)0 \le\sigma_{\mathrm{APR}}^{(\mathcal{M},J)} \le\sigma_{\mathrm{APR}}^{\mathcal{M}} \le\min\left\{\frac{D_{\mathrm{KL}}(P\|P^{\mathcal{M}})}{\tau},\frac{D_{\mathrm{KL}}(P\|P^{\mathcal{T}})}{\tau}\right\}. \tag*{(16)}

The temporal path-space divergence is a standard expression for EPR [25, 26, 42]. It has also been used to quantify temporal irreversibility in large-scale brain dynamics [27, 43]. With complete physical variables and the appropriate stochastic thermodynamic assumptions, the temporal path-space divergence is the thermodynamic EPR. Neural recordings, however, provide coarse-grained observables, and current-specific APR retains only the temporal asymmetry expressed through the selected current. We therefore interpret APR as a statistical measure of dynamical symmetry breaking and do not assign it a thermodynamic meaning a priori.

For an admissible accumulated current Jτ\mathcal{J}_{\tau} whose distribution is Gaussian, with mean J‾τ\overline{J}_{\tau} and variance ΔJτ\Delta J_{\tau}, the current-specific APR is exactly

σAPR(M,J)=1τ2J‾τ 2ΔJτ≤σAPRM.(17)\sigma_{\mathrm{APR}}^{(\mathcal{M},\mathcal{J})} = \frac{1}{\tau}\frac{2\overline{J}_{\tau}^{\,2}}{\Delta J_{\tau}} \leq\sigma_{\mathrm{APR}}^{\mathcal{M}}. \tag*{(17)}

The moment ratio in Eq. (17) is the objective σ^\hat{\sigma} used in Eq. (2). Thus, under the Gaussian assumption, RCA maximizes a current-specific lower bound on the transformation-specific APR. Its mean-current-to-variance form parallels thermodynamic uncertainty relations [28, 44]. For a general, non-Gaussian current distribution, the involution uncertainty relation gives

σAPRM≥σAPR(M,J)≥1τB(u),u=ΔJτJτ‾2,\sigma_{\mathrm{APR}}^{M} \ge\sigma_{\mathrm{APR}}^{(M,J)} \ge\frac{1}{\tau}B(u), \qquad u = \frac{\Delta J_{\tau}}{\overline{J_{\tau}}^{2}},
B(u)=21+uartanh⁡(11+u).(18)B(u) = \frac{2}{\sqrt{1+u}}\operatorname{artanh}\left(\frac{1}{\sqrt{1+u}}\right). \tag*{(18)}

For weak asymmetry, u≫1u \gg1 and B(u)=2/u−4/(3u2)+O(u−3)B(u) = 2/u - 4/(3u^{2}) + O(u^{-3}). Thus, σ^\hat{\sigma} is the leading quadratic approximation to this nonlinear lower bound.

For empirical estimation, we used one-step trajectory segments after all temporal binning or downsampling. Thus, for adjacent samples xkx_k and xk+1x_{k+1} separated by τ=Δt\tau= \Delta t, the accumulated angular current was discretized as

JΔt,k=(x1,k−c1)(x2,k+1−x2,k)−(x2,k−c2)(x1,k+1−x1,k).(19)J_{\Delta t,k} = (x_{1,k} - c_1)(x_{2,k+1} - x_{2,k}) - (x_{2,k} - c_2)(x_{1,k+1} - x_{1,k}). \tag*{(19)}

The mean and variance in Eq. (2) were computed over all valid within-trajectory transitions used to fit a given projection; transitions were never formed across trial, epoch, run or session boundaries. The effective values of Δt\Delta t were 1/30 s for mouse resting-state calcium imaging and human MEG, 0.72 s for human fMRI, 50 ms for human reaching and 10 ms for macaque reaching. Mouse task data used three-frame temporal coarse-graining for DRCA, giving Δt=0.10\Delta t = 0.10 s. The prefactor 1/τ1/\tau expresses the segment-level statistic as a rate. Estimates obtained at different temporal resolutions need not be numerically identical because temporal coarse-graining changes the current distribution.

To test whether the current-based RCA objective could robustly detect hidden rotational degrees of freedom, we applied RCA to simulated traveling patterns and experimental ciliary beating data from Geyer et al. [45] (Extended Data Figure 5). We generated the simulated pattern from two coupled complex amplitudes, A1(t)A_1(t) and A2(t)A_2(t), obeying

A˙1=μ1A1−A2−α∣A1∣2A1−β∣A2∣2A1,\dot{A}_1 = \mu_1 A_1 - A_2 - \alpha\lvert A_1\rvert^2 A_1 - \beta\lvert A_2\rvert^2 A_1,
A˙2=μ2A2+ϵA1−γ∣A1∣2A2−δ∣A2∣2A2.(20)\dot{A}_2 = \mu_2 A_2 + \epsilon A_1 - \gamma\lvert A_1\rvert^2 A_2 - \delta\lvert A_2\rvert^2 A_2. \tag*{(20)}

with μ1=μ2=0.1\mu_1 = \mu_2 = 0.1 and α=β=γ=δ=ϵ=1\alpha= \beta= \gamma= \delta= \epsilon= 1. The observed spatial field was

u(x,t)=Re⁡[A1(t)eikx+A2(t)e2ikx]+ζ(x,t),(21)u(x,t) = \operatorname{Re}\left[A_1(t)e^{\mathrm{i}kx} + A_2(t)e^{2\mathrm{i}kx}\right] + \zeta(x,t), \tag*{(21)}

where k=2πk = 2\pi and ζ(x,t)\zeta(x,t) was an Ornstein–Uhlenbeck noise field,

∂tζ(x,t)=−ζ(x,t)+ξ(x,t),\partial_t\zeta(x,t) = -\zeta(x,t) + \xi(x,t),
⟨ζ(x,t)ζ(x′,t′)⟩=2Dδ(x−x′)δ(t−t′).(22)\langle\zeta(x,t)\zeta(x',t')\rangle= 2D\delta(x-x')\delta(t-t'). \tag*{(22)}

An interactive browser-based simulation, with adjustable model parameters and noise strength, is available online [46]. In steady state, the model produces a limit-cycle traveling pattern that can be described by two spatial components whose amplitudes circulate in a reduced plane [38]. For the simulated pattern, RCA recovered these spatial components and the associated rotational probability current when the wave was readily visible (Extended Data Figure 5A–D). After the noise strength was increased such that the traveling pattern was obscured in both the instantaneous spatial profile and the space-time map, RCA still identified structured projection weights and coherent circulation (Extended Data Figure 5E–H). Applied to measurements of ciliary curvature, RCA likewise recovered two spatial modes and a circulating current associated with the propagating bend along the cilium (Extended Data Figure 5I–L). These examples show that the APR objective can recover latent rotational coordinates generated by propagation even when rotation is not directly visible in the high-dimensional observations.

Because angular current is also odd under time reversal, the same data-processing argument relates its directional asymmetry to EPR; its population mean vanishes for a time-reversal-symmetric process. With finite data, however, sampling fluctuations and optimization over projection planes can produce an apparent nonzero rotation. To test whether the observed rotations depended on temporal ordering, we independently permuted the time points within every trajectory in the mouse calcium, human fMRI and human MEG datasets. This procedure preserved the observed multivariate states while destroying the transitions between successive states; we then relearned the RCA plane from each shuffled dataset.

The original data showed coherent circulation in all three modalities, whereas the shuffled controls contained only much weaker, disorganized currents (Extended Data Figure 6). Thus, the rotations in the original data depend on ordered temporal transitions rather than only on the distribution of instantaneous states.

Extended Data Figure 6. Temporal shuffling strongly suppresses rotational probability currents. Probability density (color) and probability flux (white arrows) are shown in RCA planes for mouse resting-state calcium imaging (top), HCP-YA resting-state fMRI (middle) and HCP-YA resting-state MEG (bottom). The left column shows the original data. For the control in the right column, time points were independently permuted within each trajectory and RCA was relearned from the shuffled data. This procedure preserves the set of instantaneous multivariate states within every trajectory while destroying their temporal order. Coherent circulation was present in each original dataset but was strongly attenuated after shuffling, indicating that the identified rotational currents reflect ordered temporal transitions rather than the stationary distribution of states or optimization over the projection plane. Scale arrows indicate the probability-flux scale.

Normalization and orientation of two-dimensional modes

Across all datasets, we treated a two-dimensional linear mode first as a subspace, independent of the scaling of its two weight vectors. Let V=[v1,v2]∈RN×2V = [v_1,v_2] \in\mathbb{R}^{N \times2} contain the vectors returned by a dimensionality-reduction method. We computed the reduced QRQR decomposition V=QRV = QR, where QTQ=I2Q^{T}Q = I_2. The columns of QQ form an orthonormal basis for the same plane as the original vectors, thereby removing arbitrary differences in their lengths and mutual angle. This orthonormalization was performed in the feature space to which the mode was applied, after the dataset-specific preprocessing described below.

We used the orthonormal basis to calculate the fraction of variance explained by the plane,

tr⁡(QTCQ)tr⁡(C),(23)\frac{\operatorname{tr}(Q^{T}CQ)}{\operatorname{tr}(C)}, \tag*{(23)}

where CC is the covariance matrix of the preprocessed population activity. This quantity depends only on the plane and not on the choice of axes within it. The within-plane basis is otherwise non-unique and was oriented according to the mode being analyzed.

For an RCA mode, we diagonalized the projected covariance QTCQQ^{T}CQ and chose the first axis along the eigenvector with the larger eigenvalue. For a DRCA mode, whose objective is not invariant under arbitrary within-plane rotations, the first axis was chosen to maximize Eq. (13).

Finally, we standardized the scale of the two oriented coordinates. If qiq_i denotes an oriented unit vector and si2=qiTCqis_i^2 = q_i^{T}Cq_i is the variance of its projection, the weight used to report the mode was

wi=qisi,xi=wiTy,(24)w_i = \frac{q_i}{s_i}, \qquad x_i = w_i^{T}y, \tag*{(24)}

where yy is the preprocessed population state, so that Var⁡(xi)=1\operatorname{Var}(x_i) = 1 for both coordinates. Thus, QQ defines the orthonormal plane and its explained variance, whereas w1w_1 and w2w_2 define the unit-variance coordinates used in subsequent analyses and figures. This convention placed modes from different methods and datasets on a common scale, including the response-gradient comparison in Figure 2H.

Mouse wide-field calcium imaging

We used the processed BrainDyn-BC dataset [30]. The dataset contains recordings from 25 head-fixed VGlut1-Cre ×\times Ail62 mice expressing GCaMP6s in excitatory cortical neurons, obtained during rest and across 16 daily sessions of a cued lever-pull task. A 10-kHz tone initiated a 1-s response window, and a successful pull triggered delivery of a 4-μ\mul water reward. Cue timing included a random pre-cue delay. We analyzed the processed signals averaged within 44 cortical regions registered to the Allen Common Coordinate Framework (22 regions per hemisphere). We used these released regional traces directly and did not reprocess the raw wide-field movies. ROI ordering was required to be identical across files, and timestamps were retained to align neural activity to behavioral events. The effective sampling rate was approximately 30 Hz.

The resting-state analysis included 98 sessions from all 25 mice. Each session contained 18,000 samples from 44 regions, corresponding to approximately 600 s of recording. All 98 traces had the complete expected duration and dimensionality, so no resting-state session was excluded.

cluded. No temporal coarse-graining was applied in the analyses reported here.

The task dataset contained 364 sessions from all 25 mice, spanning task days 1–15 (13–15 sessions per mouse). Each session contained 54,000 samples, corresponding to approximately 1,800 s. The behavioral table contained 48,549 trials with finite cue times. For cue-aligned analyses, we extracted 240-frame windows comprising 120 pre-cue and 120 post-cue frames (approximately 4 s on either side of cue onset). A trial was retained only if the complete 240 × 44 window lay within the valid trace and all values were finite. This procedure retained 48,342 trials from all 364 sessions and all 25 mice: 28,760 trials with a recorded pull onset and 19,582 without a recorded pull. We excluded 207 trials because the requested window crossed a recording boundary (107 before the start and 100 after the end); no trials were excluded because of nonfinite values, an unexpected window shape or a missing processed session.

Anatomical representation of the mouse RCA mode

To visualize the anatomical organization of the resting-state rotational mode, we examined the two RCA loading vectors, w1,w2∈R44w_1,w_2 \in\mathbb{R}^{44}, that define the latent coordinates x1=w1Tyx_1 = w_1^{\mathsf{T}}y and x2=w2Tyx_2 = w_2^{\mathsf{T}}y from the regional activity vector yy. The signed loading of each Allen Common Coordinate Framework region on the two axes is shown in Extended Data Figure 2A. Because the orientation of the coordinate axes within an identified plane is arbitrary, the two loading vectors should be interpreted jointly as specifying a rotational subspace rather than as two independently meaningful anatomical components.

We summarized the directed regional organization implied by this plane using a rank-two antisymmetric matrix. Let ss denote the rotation direction such that s=1s = 1 for clockwise rotation and s=−1s = -1 for counterclockwise rotation. We defined the corresponding unit-magnitude generator in regional space as

A=2s(w1w2T−w2w1T)∥w1∥2+∥w2∥2.(25)A = \frac{2s\left(w_1w_2^{\mathsf{T}} - w_2w_1^{\mathsf{T}}\right)}{\lVert w_1\rVert^2 + \lVert w_2\rVert^2}. \tag*{(25)}

This construction embeds a unit rotational generator in the subspace spanned by the two RCA loading vectors while introducing no coupling in the orthogonal complement. It is antisymmetric, Aij=−AjiA_{ij} = -A_{ji}, and therefore assigns an orientation to each pairwise contribution without introducing a reciprocal component. For each region, we defined node strength as si=∑j∣Aij∣s_i = \sum_j |A_{ij}|. In Extended Data Figure 2B, region color represents sis_i, and the 20 largest off-diagonal elements by absolute magnitude are displayed as arrows, oriented from region ii to region jj when Aij>0A_{ij} > 0. Region locations were taken from the Allen cortical map and displayed separately for the two hemispheres. Because AA is constructed solely from the RCA projection weights, it visualizes the directed interregional organization implied by the latent rotational mode; it should not be interpreted as a measurement of anatomical connectivity, synaptic coupling or causal influence between regions.

Inspection of the ranked matrix elements showed that the strongest antisymmetric coupling linked right primary motor cortex (3) to left lateral agranular retrosplenial cortex (21). Four of the seven strongest couplings formed the closed sequence right primary motor cortex (3) →\to left lateral agranular retrosplenial cortex (21) →\to left supplemental somatosensory cortex (11) →\to left dorsal retrosplenial cortex (22) →\to right primary motor cortex (3). The displayed structure is therefore dominated by a long-range, interhemispheric loop spanning sensorimotor and retrosplenial regions, rather than by interactions restricted to neighboring cortical areas. This organization provides a possible dynamical scaffold linking ongoing cortical state to subsequent sensorimotor processing.

Human fMRI

We analyzed cortical fMRI data from the HCP Young Adult S1200 release using the cohort and preprocessed time series described previously [27, 47]. Cortical data were registered to fsLR-32k space using MSM-All and parcellated into 100 regions. Five component-based noise-correction (CompCor) components derived from ventricular and white-matter masks and 12 detrended HCP motion regressors were removed. The mean global signal was regressed out, after which the time series were band-pass filtered between 0.009 and 0.08 Hz. Frames with framewise displacement greater than 0.2 mm or DVARS greater than 75 were treated as outliers, and sessions in which more than 50% of frames were outliers were excluded. Participants were retained only if all selected scans survived quality control, yielding 590 healthy adults.

For the resting-state analysis, we used the REST1 LR and RL runs (repetition time, 0.72 s) and divided each run into nonoverlapping 50-frame trajectories. For the social task analysis, we used the complete LR and RL runs (274 frames per run), yielding 1,180 participant-run trajectories from 590 participants for mode estimation. HCP task labels were not used to identify the projection and were introduced only afterward to align and average task and fixation trajectories. For this block-aligned analysis, participants whose block order differed from the majority order were excluded separately for each phase-encoding direction (3 from LR and 2 from RL), leaving 587 LR runs and 588 RL runs from 589 unique participants.

Human MEG

We analyzed resting-state MEG recordings from the HCP Young Adult S1200 release [48]. Data were acquired with a whole-head MAGNES 3600 system containing 248 magnetometer channels, at a sampling rate of 2,034.51 Hz, in up to three approximately 6-min runs per participant. We used the released channel-level rmegpreproc outputs. In the HCP pipeline, bad channels and time segments were removed, independent components classified as artifacts, including ocular and cardiac components, were projected out, and the cleaned data were downsampled to approximately 508.63 Hz and segmented into artifact-clean 2-s epochs.

We used 267 preprocessed runs and constructed a common sensor space by selecting the 238 channels present in at least 95% of runs (at least 254 of 267). We then retained only runs containing all 238 selected channels, without channel imputation. This procedure yielded 140 runs from 71 participants, comprising 19,960 artifact-clean epochs. Within each epoch, we averaged neighboring samples in consecutive 1/30-s bins, producing an independent 60 × 238 trajectory sampled at 30 Hz; no transitions were formed across epoch boundaries. Only trajectories with finite values and the common channel ordering were analyzed.

HCP Early Psychosis fMRI

We analyzed the derived AAL116-parcellated resting-state fMRI time series distributed with the HCP Early Psychosis (HCP-EP) study [32]. The dataset comprised 53 healthy controls and 82 participants with non-affective psychosis. Each participant completed four resting-state runs over two consecutive days, with 410 frames per run and a repetition time of 0.72 s. The source images were processed using the HCP minimal preprocessing pipeline and denoised using ICA-FIX; regional time series were band-pass filtered between 0.01 and 0.08 Hz. We retained all four runs, yielding 212 control runs and 328 non-affective-psychosis runs.

In the primary analysis, RCA and PCA projections were learned exclusively from healthy-control data. The projections were fitted using all 53 controls, and participant-level APR and PCA variability were calculated by combining the four runs from each participant; PCA variability was defined as the standard deviation along the leading principal component. In the control-trained split-sample analysis, 26 randomly selected controls were used for projection learning, while the remaining 27 controls and all 82 participants with non-affective psychosis were held out for evaluation. Group differences were assessed using two-sided Welch’s tt-tests, with participants as the independent analysis units. In this control-trained analysis, PCA variability did not differ between groups (P=0.592P = 0.592), whereas APR remained lower in participants with non-affective psychosis (P=2.52×10−4P = 2.52 \times10^{-4}; Extended Data Figure 7A,B).

We performed a complementary disease-trained analysis to test whether this result depended on learning the projections from control data. We randomly divided the 82 participants with non-affective psychosis into two equal groups, learned the RCA and PCA projections from 41 participants, and evaluated them in the remaining 41 participants and all 53 controls. In this analysis, PCA variability was only modestly lower in the held-out psychosis group (P=0.0275P = 0.0275; one asterisk), whereas APR showed a more significant reduction (P=8.29×10−5P = 8.29 \times10^{-5}; Extended Data Figure 7C,D). Thus, reduced APR was detected irrespective of which group supplied the data used to learn the rotational plane.

Reaching-data benchmark

We benchmarked RCA against jPCA using the macaque and human motor-cortical recordings shown in Figure 1C–N. We performed two complementary analyses. First, following the conventional jPCA analysis, we fitted each method to activity averaged across repeated trials within each reach condition [8]. Second, we fitted each method directly to the same individual-trial activity pooled across reach conditions, without averaging trials before dimensionality reduction. The first analysis tested whether RCA recovered the established condition-averaged rotational structure; the second tested whether each method could isolate circulation in the presence of trial-specific neural variability.

For the human analysis, we used motor-cortical activity recorded from participant C1 during an active center-out reaching task [49]. C1 was male, 55–60 years old at implantation, and had a chronic C4-level, ASIA D spinal cord injury. The original study was conducted under a US Food and Drug Administration Investigational Device Exemption, was approved by the Institutional Review Boards of the Universities of Pittsburgh and Chicago, and was registered at ClinicalTrials.gov (NCT01894802). The participant provided informed consent before study procedures.

Neural activity was recorded from two 4×44 \times4-mm microelectrode arrays implanted in motor cortex, each with 96 active electrode shanks. Voltages were sampled at 30 kHz, high-pass filtered at 750 Hz, thresholded at −4.5-4.5 times the root-mean-square voltage and sorted offline. During the task, the participant’s arm rested on a low-friction forearm support, and eight targets were arranged around a circle 30 cm in diameter. Each trial comprised five 1-s phases indicated by auditory tones: target cue, movement, hold, return and a final wait. The target direction was presented as text at cue onset, and the participant moved the arm toward the target after the movement tone. Participant C1 completed 11 blocks, with ten repetitions of each target per block, for a total of 880 reaches.

For our analysis, spike trains were aligned to the first auditory tone (cue onset), binned at 50 ms across the full 0–5-s trial and converted to firing rates. The firing-rate traces were smoothed with a Gaussian kernel with a standard deviation of 100 ms. For the RCA–jPCA comparison, we retained activity from the second tone (movement instruction; median latency from cue onset, 1.01 s) to the end of the trial. This interval contained 80 bins, with centers spanning 1.025–4.975 s relative to cue onset.

For the macaque analysis, we used the publicly available motor-cortical activity of monkey Jenkins from the delayed-reaching dataset of Churchland and colleagues [8]. Spike counts from 137 units were aligned to movement onset, binned at 10 ms from −250-250 to 450 ms and smoothed with a Gaussian kernel with a standard deviation of 20 ms. The 2,295 analysis trials were grouped into 36 maze conditions, each containing at least 20 trials. Projection planes were fitted using the full analysis interval.

For both datasets, we first calculated the mean firing-rate trajectory for each reach condition. For each unit, we calculated the range of its condition-averaged activity across conditions and time. Both the condition averages and every individual-trial firing-rate trace were divided by this range plus a 5-Hz soft-normalization constant. At each time point, we then subtracted the mean normalized activity across conditions. This common transformation ensured that the condition-average and individual-trial fits used the same unit scaling and time-dependent baseline.

For the condition-average analysis, RCA and jPCA were fitted to the resulting condition-by-time-by-unit tensor. RCA learned two projection vectors by maximizing the angular-current objective in Eq. (2). For jPCA, the same tensor was first reduced to six principal components, a skew-symmetric linear dynamical system was fitted to the temporal derivative and the complex eigenplane with the highest rotational frequency was retained [8]. The fitted condition-average trajectories are shown for macaque (Figure 1C,D) and human (Figure 1I,J) reaching.

For the individual-trial analysis, RCA and jPCA were refitted from scratch using all preprocessed individual trials pooled across reach conditions. The methods received identical trial-level inputs but selected planes using their respective angular-current and skew-symmetric-dynamics objectives. For visualization, we projected each trial into the corresponding fitted plane and then averaged the projected trajectories within each condition. These post-fit condition averages are shown for macaque (Figure 1E,F) and human (Figure 1K,L) reaching; colors distinguish conditions, and dots mark trajectory onset.

We quantified trial-level circulation in the individual-trial-fit planes for macaque maze condition 1 (Figure 1G,H) and the representative human bottom-target condition (Figure 1M,N). For each method and dataset, the fitted plane was expressed as an orthonormal basis before projection. We estimated the empirical center and mean squared radius from all projected samples in the selected condition. The normalized cumulative angular current was then calculated from finite differences using Eq. (3). Lines show the across-trial mean, and shading denotes the standard error of the mean, SDi[Ji(t)]/N\mathrm{SD}_{i}[J_{i}(t)]/\sqrt{N}, where NN is the number of trials. The reported SNR was calculated from the final current using Eq. (4).

Comparison with alternative dimensionality-reduction methods

We evaluated the detected rotational modes by comparing the current-based scores obtained with RCA, DyCA, jPCA, an unconstrained linear dynamical system (LDS), PCA and UMAP across the datasets analyzed in this study (Extended Data Figure 8). For the six single-rotation analyses, each method was evaluated using the surrogate APR σ^\hat{\sigma} in Eq. (2); for the mouse task and HCP-YA social task, the corresponding projection was evaluated using the double-rotation score σ^double\hat{\sigma}_{\mathrm{double}} in Eq. (13). RCA yielded the largest single-rotation score in every dataset, with DyCA providing the closest alternative in several cases. This similarity is expected because DyCA favors components whose derivatives are captured by low-dimensional linear dynamics, and asymmetric linear coupling between two variables can generate rotation. Its performance was much weaker for macaque reaching, however, where the rotational trajectories may not be adequately described by a single linear flow. RCA showed a particularly large advantage in this dataset, for which the other projections produced little net angular current. DRCA likewise produced substantially larger double-rotation scores than the comparison methods in both mouse and human task data.

Extended Data Figure 8. Comparison of current-based rotation scores across dimensionality-reduction methods and datasets. Bars show the surrogate APR σ^\hat{\sigma} for single-rotation analyses (first three rows) and the double-rotation score σ^double\hat{\sigma}_{\mathrm{double}} for double-rotation analyses (bottom row). From top left to bottom right, the datasets are mouse resting-state calcium imaging, HCP-YA resting-state fMRI, HCP-EP resting-state fMRI, HCP-YA resting-state MEG, human intracortical activity during reaching, macaque motor-cortical activity during reaching, mouse calcium imaging during the lever-pull task and HCP-YA fMRI during the social task. Single-rotation projections were obtained using rotational component analysis (RCA), dynamical component analysis (DyCA), jPCA, an unconstrained linear dynamical system (LDS), principal component analysis (PCA) and UMAP. For the double-rotation analyses, double-rotational component analysis (DRCA) was compared with the same alternative methods.

Because RCA and DRCA directly optimize the respective scores, these comparisons do not provide an independent test of predictive performance. Nevertheless, σ^\hat{\sigma} is a physically motivated, current-based measure of the strength and consistency of directed circulation; hence, the comparison shows that RCA and DRCA detect stronger and more consistent rotational dynamics.

Mouse response-drift model

Mouse activity was projected onto the RCA plane identified from resting-state data and aligned to cue onset. We pooled response and nonresponse trials and represented the instantaneous velocity by the displacement between successive samples. The plane was divided into a 10×1010 \times10 grid spanning −2.5≤x1,x2≤2.5-2.5 \leq x_{1}, x_{2} \leq2.5. Within each grid cell, the pre-cue drift fpre\mathbf{f}^{\mathrm{pre}} was estimated by averaging the displacements from the ten frames immediately preceding cue onset, whereas the post-cue drift fpost\mathbf{f}^{\mathrm{post}} was estimated from the first five post-cue frames.

Only cells containing at least 1,000 state–displacement pairs in the corresponding window were retained. Because the calcium signals were sampled at approximately 30 Hz, the pre-cue and post-cue windows spanned approximately 0.33 s and 0.17 s, respectively.

We fitted the pre-cue field with the linear model

fpre(x)=Ax+b\mathbf{f}^{\mathrm{pre}}(\mathbf{x}) = A\mathbf{x} + \mathbf{b}

by weighted least squares, using the number of samples in each cell as the weight. The empirical cue-evoked response field was frsp=fpost−fpre\mathbf{f}^{\mathrm{rsp}} = \mathbf{f}^{\mathrm{post}} - \mathbf{f}^{\mathrm{pre}} in cells for which both drift fields were available. We defined the response direction n^∥\hat{\mathbf{n}}_{\parallel} as the direction of the mean response-drift vector across these cells. The response-drift magnitude was then fitted with Eq. (8).

For each trial, we evaluated the modeled post-cue field at the cue-onset state and defined the scalar drift score f∥post=∣fpost(xcue)⋅n^∥∣f_{\parallel}^{\mathrm{post}} = \lvert\mathbf{f}^{\mathrm{post}}(\mathbf{x}^{\mathrm{cue}}) \cdot\hat{\mathbf{n}}_{\parallel} \rvert. In Figure 3, data from all fifteen task days were used for fitting. In Figure 4, to account for changes in the response field during learning, the parameters of frsp\mathbf{f}^{\mathrm{rsp}} were estimated separately in five nonoverlapping three-day training windows (Extended Data Figure 4); the pre-cue field and n^∥\hat{\mathbf{n}}_{\parallel} were held fixed. For the function Prsp(f∥post)P^{\mathrm{rsp}}(f_{\parallel}^{\mathrm{post}}), scores from all trials were divided into 30 equally spaced bins over their observed range, and bins containing fewer than 50 trials were excluded. The response probability in each retained bin was the fraction of trials followed by a lever pull, with binomial standard errors. We fitted Eq. (9) to these probabilities by nonlinear least squares, weighting each residual by the square root of the number of trials in that bin.

Recurrent neural network model

We implemented a discrete-time leaky RNN with one scalar input, N=64N = 64 hidden units and one scalar output. Its hidden state followed Eq. (10), and the output was

yt=Woutht.(26)y_{t} = \mathbf{W}_{\mathrm{out}}\mathbf{h}_{t}. \tag*{(26)}

We used α=0.2\alpha= 0.2 and σ=0.5\sigma= 0.5, initialized the hidden state to zero at the beginning of every trial and drew an independent Gaussian noise vector at every time step. The elements of Win\mathbf{W}_{\mathrm{in}} were drawn from N(0,1)\mathcal{N}(0,1), whereas those of Wout\mathbf{W}_{\mathrm{out}} and Wrec\mathbf{W}_{\mathrm{rec}} were drawn from N(0,1/N)\mathcal{N}(0,1/N). The recurrent bias was initialized from the PyTorch default uniform distribution U(−N−1/2,N−1/2)\mathcal{U}(-N^{-1/2},N^{-1/2}). The input and output weights were fixed after initialization; only Wrec\mathbf{W}_{\mathrm{rec}} and b\mathbf{b} were trained.

Each training trial comprised 100 time steps. A unit-amplitude cue was presented for one time step at an onset sampled uniformly from steps 10–79. The response window comprised the following W=10W = 10 steps, with no intervening delay. Rather than training the network to match a prescribed output trajectory, we trained it to maximize the net output displacement over this window. The sum of the one-step output increments telescopes, so for a batch of BB trials with cue onsets tit_i, the loss can be written in endpoint form as

L=−1BW∥Wout∥∑i=1B(yti+W(i)−yti(i)).(27)\mathcal{L}=-\frac{1}{BW\lVert W_{\mathrm{out}}\rVert}\sum_{i=1}^{B}\left(y_{t_i+W}^{(i)}-y_{t_i}^{(i)}\right). \tag*{(27)}

The normalization by WW expresses the endpoint displacement per time step. Because WW was fixed, however, the loss rewarded net displacement at the end of the response window and did not independently constrain response speed or the shape of the output trajectory within the window. We minimized this loss by backpropagation through time using Adam with a learning rate of 10−310^{-3}, a batch size of 64 and 1,000 parameter-update iterations.

For subsequent analyses, the response strength of trial ii was defined as virsp=(yti+W(i)−yti(i))/[W∥Wout∥]v_i^{\mathrm{rsp}}=(y_{t_i+W}^{(i)}-y_{t_i}^{(i)})/[W\lVert W_{\mathrm{out}}\rVert], corresponding to the single-trial contribution to (27) without the leading minus sign.

To identify spontaneous rotational dynamics, we simulated the trained network with zero input. We generated 200 trajectories, discarded the first 100 steps of each as burn-in and retained the following 500 steps. RCA was fitted to these zero-input trajectories without access to the cue or response strength.

For the drift fields in Figure 3H,I, we simulated 100,000 noisy trials after a 150-step burn-in and sampled cue onset uniformly over the following 60 steps. We estimated the pre-cue and post-cue drifts from mean one-step displacements within five-step pre-cue and post-cue windows, respectively. The response direction n^∥\hat{n}_{\parallel} was defined as the normalized spatial mean of fpost−fpref_{\mathrm{post}}-f_{\mathrm{pre}} over cells in which both fields were available.

The response-strength map in Figure 3J was estimated from a separate set of 6,000 noisy random-cue trials. Cue-onset state was defined as the state one step before the cue, and grid cells containing fewer than 15 trials were masked. For Figure 3K, we simulated 3,000 trials with a fixed cue after burn-in, projected displacement from the cue-onset state onto n^∥\hat{n}_{\parallel} and divided trials at the median response strength. For Figure 3L, each trial was assigned the value of f∥post=fpost⋅n^∥f_{\parallel}^{\mathrm{post}} = f^{\mathrm{post}} \cdot\hat{n}_{\parallel} in its cue-onset grid cell.

To construct the RNN analogue of Figure 2H, we simulated an independent set of 20,000 random-cue trials and retained the 64-dimensional hidden state one step before cue onset together with the response strength vrspv^{\mathrm{rsp}}. The RCA, PCA and jPCA planes were learned from the same zero-input trajectories used to characterize spontaneous network activity; jPCA was fitted in the leading six-PC subspace. We additionally generated 1,000 random orthonormal planes by QR decomposition of Gaussian random matrices. For every plane, cue-onset states were centered, oriented with the first coordinate along the leading within-plane covariance axis and scaled to unit variance along each coordinate. We divided the resulting plane into a 15×1515 \times15 grid spanning −2.5≤x1,x2≤2.5-2.5 \le x_1, x_2 \le2.5 and excluded cells with fewer than 50 trials. The mean response-strength map was fitted with the same radial-plus-directional form as Eq. (5), replacing response probability with vrspv^{\mathrm{rsp}}.

The RCA plane produced a gradient of 0.268 and exceeded all 1,000 random-plane values (Extended Data Figure 9). However, it is only slightly larger than those of PCA (0.254) and jPCA (0.255). This is probably because the RNN is itself low-dimensional as it is trained from a simple task. Therefore, different dimensional reduction techniques may converge on the same plane.

Extended Data Figure 9. The spontaneous RNN rotational plane is strongly aligned with the cue-evoked response. Distribution of response-gradient magnitudes obtained from 1,000 random orthonormal planes in the RNN hidden state. Vertical lines indicate the gradients obtained in the RCA (magenta; 0.268), PCA (blue; 0.254) and jPCA (green; 0.255) planes. Projection planes were learned from zero-input network activity, without access to the cue or response strength, and all projected coordinates were scaled to unit variance before fitting. The RCA gradient exceeded the gradients from all random planes.

Held-out prediction from rotational states and pupil-indexed arousal

For the receiver operating characteristic analysis in Figure 6D, we randomly divided recording sessions into training (75%) and test (25%) sets. The RCA and DRCA projection vectors were fixed before this supervised analysis. Using only the training sessions, we separately fitted the response-probability model P(r,θ)=P0+[f0+fsin⁡(θ−θ0)]P(r,\theta) = P_0 + [f_0 + f\sin(\theta-\theta_0)] to the cue-onset RCA state and to the DRCA state 80 frames (2.7 s) pre-cue. Each fitted function assigned a scalar response score to every trial.

We then fitted three logistic-regression classifiers using the training sessions: one based on the RCA score, one based on the DRCA score and one based on both scores. Predictors were standardized using the training data, and the classifiers used L2L_2 regularization with C=1C = 1. The response-probability fits, feature scaling and classifier parameters were applied without refitting to trials from the held-out test sessions. The receiver operating characteristic curves and AUC values in Figure 6D were calculated exclusively from these held-out predictions.

For the comparison with pupil-indexed arousal (Extended Data Figure 3), we used the continuous right-eye pupil-diameter trace d(t)d(t), measured in pixels. Pupil diameter was converted separately for each session to a robust arousal score,

zarousal(t)=d(t)−median⁡[d]1.4826 median⁡[∣d−median⁡[d]∣],(28)z_{\mathrm{arousal}}(t) = \frac{d(t) - \operatorname{median}[d]}{1.4826\,\operatorname{median}[|d-\operatorname{median}[d]|]}, \tag*{(28)}

where the median and median absolute deviation were calculated over all finite, positive pupil-diameter samples in that session; the factor 1.4826 makes the scaled median absolute deviation consistent with the standard deviation for Gaussian data [50]. Nonfinite or nonpositive samples, as well as samples with ∣zarousal∣>20|z_{\mathrm{arousal}}| > 20, were treated as tracking artifacts and set to missing. For each trial, we used the arousal score from the nearest time point with a finite value, without temporal interpolation. We retained trials with finite arousal values and finite cue-onset RCA coordinates and applied the same 75:25 session-level splitting procedure described above. This yielded 35,547 training trials from 267 sessions and 12,272 test trials from 89 held-out sessions. Using only the training sessions, we fitted the RCA response score with the radial-sine model above. We also divided pupil arousal into 20 equally spaced bins, retained bins containing at least 25 trials and fitted the binned response probabilities with a cubic polynomial weighted by the square root of the bin count. Predicted probabilities were clipped to the interval [0,1][0,1].

The RCA-only model was an L2L_2-regularized logistic regression using the fitted response score. The arousal-only model used the cubic fit directly, and the combined model was an L2L_2-regularized logistic regression using both the RCA score and the cubic arousal prediction. Logistic-regression predictors were standardized using the training sessions only. The receiver operating characteristic curves and AUCs in Extended Data Figure 3B were evaluated exclusively on the held-out sessions. Extended Data Figure 3A is a descriptive visualization: the arousal bins and cubic curve shown there were refitted using all matched trials and were not used to calculate the held-out AUCs.

Simulation of the learning model

The phenomenological learning simulation used the affine pre-cue drift fpref^{\mathrm{pre}}, global response direction n^∥\hat{n}_{\parallel} and response-probability function in Eq. (9) inferred from the mouse data. We approximated the distribution of observed cue-onset states by fitting an isotropic Gaussian N(μ,σx2I)\mathcal{N}(\boldsymbol{\mu},\sigma_x^2 I). For the illustration in Figure 4F, we drew 150 states from this distribution, evaluated their intrinsic drift and assigned each state a binary response outcome with probability Prsp(fpre(x)⋅n^∥)P^{\mathrm{rsp}}(f^{\mathrm{pre}}(x)\cdot\hat{n}_{\parallel}). The displayed response and nonresponse groups were randomly subsampled to equal sizes, whereas the Fisher linear discriminant shown in the panel was fitted using all 150 samples.

To simulate learning, we drew a fixed ensemble of M=200M=200 cue-onset states from the same Gaussian distribution and initialized f0rsp=0f^{\mathrm{rsp}}_0=0 and k=0k=0. At each learning iteration, we evaluated Eq. (11) at every sampled state xix_i and computed

fipost=fpre(xi)+frsp(xi),pi=Prsp(fipost⋅n^∥).(29)f_i^{\mathrm{post}}=f^{\mathrm{pre}}(x_i)+f^{\mathrm{rsp}}(x_i),\qquad p_i=P^{\mathrm{rsp}}\left(f_i^{\mathrm{post}}\cdot\hat{n}_{\parallel}\right). \tag*{(29)}

Rather than drawing a single response outcome at every iteration, we used the corresponding mean-field update:

f0rsp←(1−a1)f0rsp+b1⟨fpost⟩rsp,k←(1−a2)k+b2⟨x⟩rsp.(30)\begin{aligned} f^{\mathrm{rsp}}_0 &\leftarrow(1-a_1)f^{\mathrm{rsp}}_0+b_1\langle f^{\mathrm{post}}\rangle_{\mathrm{rsp}},\\ k &\leftarrow(1-a_2)k+b_2\langle x\rangle_{\mathrm{rsp}}. \tag*{(30)} \end{aligned}

where ⟨⋅⟩\langle\cdot\rangle indicates an ensemble average. Thus, states and drifts were reinforced in proportion to their probability of producing a response.

We ran 50,000 learning iterations with a1=0.0014a_1=0.0014, a2=0.018a_2=0.018, b1=0.001b_1=0.001 and b2=0.01b_2=0.01. At the end of the simulation, f0rsp=(0.0510,−0.0444)f^{\mathrm{rsp}}_0=(0.0510,-0.0444) and k=(0.214,0.235)k=(0.214,0.235), and the mean modeled response probability across the sampled states was 0.584. The resulting response-drift field is shown in Figure 4G.

Data availability

The BraiDyn-BC calcium-imaging dataset is described by Kondo et al. [30]. HCP-YA fMRI and MEG data were obtained from the Human Connectome Project, subject to its data-use terms. The derived AAL116-parcellated HCP-EP time series were obtained from the archive accompanying Hancock et al. [32]. The human intracortical reaching data are described by Tian et al. [49]; the data will be released upon publication. The monkey reaching dataset is available with the original study [8].

Code availability

Code will be available upon publication.

References

Acknowledgements

We thank Luca Mazzucato for fruitful discussions. D.E.G. acknowledges support by the NSF-Simons National Institute for Theory and Mathematics in Biology (NITMB) Fellowship supported via grants from the NSF (DMS-2235451) and Simons Foundation (MPS-NITMB-00005320). V.V. acknowledges partial support from the Army Research Office under grants W911NF-22-2-0109 and W911NF-23-1-0212. B.D. is supported by the National Institutes of Health (NIH) grants R01EY037119, R01NS133598, T90DA059109; National Science Foundation (NSF) grant 2605555, and the Simons Foundation Collaboration on the Global Brain. Y.X. is supported by a NeuroData Discovery Grant from the Kavli Foundation. M.F. and V.V. acknowledge partial support from the France Chicago center through a FACCTS grant.

We thank Charles Greenspon and John Downey for access to the human participant datasets. We thank Michael Boninger, MD, for serving as the sponsor-investigator of the BCI clinical trial conducted under an FDA Investigational Device Exemption.

This research was partly supported by the National Science Foundation through the Center for Living Systems (grant no. 2317138), the National Institute for Theory and Mathematics in Biology, the Simons Foundation and the Chan Zuckerberg Foundation. This work was completed in part with resources provided by the University of Chicago’s Research Computing Center. We acknowledge the use of ChatGPT (OpenAI) and Claude (Anthropic) for assistance with code development, data-analysis workflows, and language editing. The authors reviewed and verified all AI-assisted outputs and take full responsibility for the content of the manuscript.

References

  1. [1]M. G. Perich, D. Narain, and J. A. Gallego, Nature Neuroscience 28, 1582–1597 (2025).
  2. [2]B. M. Yu, J. P. Cunningham, G. Santhanam, S. I. Ryu, K. V. Shenoy, and M. Sahani, Journal of Neurophysiology 102, 614 (2009).
  3. [3]V. Mante, D. Sussillo, K. V. Shenoy, and W. T. Newsome, Nature 503, 78 (2013).
  4. [4]C. D. Harvey, P. Coen, and D. W. Tank, Nature 484, 62 (2012).
  5. [5]H. Benisty et al., Nature Neuroscience 27, 148 (2024).
  6. [6]C. J. MacDowell, A. Libby, C. I. Jahn, S. Tafazoli, A. Ardalan, and T. J. Buschman, Nature Communications 16, 3359 (2025).
  7. [7]E. Marder and R. L. Calabrese, Physiological Reviews 76, 687 (1996).
  8. [8]C. M. Churchland, J. P. Cunningham, M. T. Kaufman, J. D. Foster, P. Nuyujukian, S. I. Ryu, and K. V. Shenoy, Nature 487, 51 (2012).
  9. [9]C. Pandarinath, V. Gilja, C. H. Blabe, P. Nuyujukian, A. A. Sarmash, B. L. Sorice, E. N. Eskandar, L. R. Hochberg, J. M. Henderson, and K. V. Shenoy, eLife 4, e07436 (2015).
  10. [10]J. A. Michaels, B. Dann, and H. Scherberger, PLOS Computational Biology 12, e1005175 (2016).
  11. [11]A. H. Lara, J. P. Cunningham, and M. M. Churchland, Nature Communications 9, 2754 (2018).
  12. [12]A. A. Russo, S. R. Bittner, S. M. Perkins, J. S. Seely, B. M. London, A. H. Lara, A. Miri, N. J. Marshall, A. Kohn, T. M. Jessell, L. F. Abbott, J. P. Cunningham, and M. M. Churchland, Neuron 97, 953 (2018).
  13. [13]D. A. Sabatini and M. T. Kaufman, Nature Communications 15, 7007 (2024).
  14. [14]L. M. Ward, Trends in Cognitive Sciences 7, 553–559 (2003).
  15. [15]E. Başar, C. Başar-Eroğlu, S. Karakaş, and M. Schürmann, International Journal of Psychophysiology 39, 241–248 (2001).
  16. [16]A. Schnitzler and J. Gross, Nature Reviews Neuroscience 6, 285–296 (2005).
  17. [17]Y. Xu, X. Long, J. Feng, P. Gong, et al., Nature Human Behaviour 7, 1196 (2023).
  18. [18]Z. Ye, A. E. Ladd, N. MacKenzie, L. Kolich, A. J. Li, D. Birman, M. S. Bull, T. L. Daigle, B. Tasic, H. Zeng, and N. A. Steinmetz, Science 392, eadx1369 (2026).
  19. [19]Y. Liu, M. M. Nour, N. W. Schuck, T. E. J. Behrens, and R. J. Dolan, Nature Reviews Neuroscience 23, 204 (2022).
  20. [20]M. Tsodyks, T. Kenet, A. Grinvald, and A. Arieli, Science 286, 1943 (1999).
  21. [21]R. Srinath, Y. Xu, D. A. Ruff, A. M. Ni, B. Doiron, and M. R. Cohen, Proceedings of the National Academy of Sciences 123, e2523217123 (2026).
  22. [22]P. Berkes, G. Orbán, M. Lengyel, and J. Fiser, Science 331, 83 (2011).
  23. [23]J. A. Goldberg, U. Rokni, and H. Sompolinsky, Neuron 42, 489 (2004).
  24. [24]A. Arieli, A. Sterkin, A. Grinvald, and A. Aertsen, Science 273, 1868 (1996).
  25. [25]G. E. Crooks, Physical Review E 60, 2721 (1999).
  26. [26]U. Seifert, European Physical Journal B 64, 423 (2008).
  27. [27]C. W. Lynn, E. J. Cornblath, L. Papadopoulos, M. A. Bertolero, and D. S. Bassett, Proceedings of the National Academy of Sciences USA 118, e2109889118 (2021).
  28. [28]T. R. Gingrich, J. M. Horowitz, N. Perunov, and J. L. England, Physical Review Letters 116, 120601 (2016).
  29. [29]J. Li, J. M. Horowitz, T. R. Gingrich, and N. Fakhri, Nature Communications 10, 1666 (2019).
  30. [30]M. Kondo, K. Sahara, R. Harukuni, R. Aoki, S. Sugimoto, Y. R. Tanaka, M. Matsuzaki, and K. Nakae, Scientific Data 12, 1264 (2025).
  31. [31]D. C. Van Essen, S. M. Smith, D. M. Barch, T. E. J. Behrens, E. Yacoub, K. Ugurbil, and WU-Minn HCP Consortium, NeuroImage 80, 62 (2013).
  32. [32]F. Hancock, F. E. Rosas, R. A. McCutcheon, J. Cabral, O. Dipasquale, and F. E. Turkheimer, PLOS ONE 18, e0282707 (2023).
  33. [33]E. Marder and D. Bucher, Current Biology 11, R986 (2001).
  34. [34]P. T. Sadtler, K. M. Quick, M. D. Golub, S. M. Chase, S. I. Ryu, E. C. Tyler-Kabara, B. M. Yu, and A. P. Batista, Nature 512, 423 (2014).
  35. [35]E. R. Oby, A. D. Degenhart, E. M. Grigsby, A. Motiwala, N. T. McClain, P. J. Marino, B. M. Yu, and A. P. Batista, Nature Neuroscience 28, 383 (2025).
  36. [36]A. Luczak, P. Barthó, and K. D. Harris, Neuron 62, 413 (2009).
  37. [37]Z. W. Davis, J. Muller, J. Martinez-Trujillo, T. Sejnowski, and J. H. Reynolds, Nature 587, 432 (2020).
  38. [38]M. Fruchart, R. Hanai, P. B. Littlewood, and V. Vitelli, Nature 592, 363 (2021).
  39. [39]M. Fruchart and V. Vitelli, Nonreciprocal many-body physics (2026), arXiv:2602.11111, arXiv:2602.11111 [cond-mat.stat-mech].arxiv.org/abs/2602.11111
  40. [40]V. Mochulska and P. François, Proceedings of the National Academy of Sciences USA 122, e2514508122 (2025).
  41. [41]T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed. (Wiley-Interscience, Hoboken, New Jersey, 2006).
  42. [42]R. Kawai, J. M. R. Parrondo, and C. Van den Broeck, Physical Review Letters 98, 080602 (2007).
  43. [43]J. Cruzat, R. Herzog, P. Prado, Y. Sanz-Perl, R. Gonzalez-Gomez, S. Moguilliner, M. L. Kringelbach, G. Deco, E. Tagliazucchi, and A. Ibañez, The Journal of Neuroscience 43, 1643 (2023).
  44. [44]A. C. Barato and U. Seifert, Physical Review Letters 114, 158101 (2015).
  45. [45]V. F. Geyer, J. Howard, and P. Sartori, Nature Physics 18, 332 (2022).
  46. [46]S. Chen, Hidden rotation in a noisy traveling pattern: Interactive simulation, https://sihanchenphysics.github.io/RCA/ (2026), accessed 8 October 2026.
  47. [47]D. M. Barch, G. C. Burgess, M. P. Harms, S. E. Petersen, B. L. Schlaggar, M. Corbetta, M. F. Glasser, S. Curtiss, S. Dixit, C. Feldt, D. Nolan, E. Bryant, T. Hartley, O. Footer, J. M. Bjork, R. Poldrack, S. Smith, H. Johansen-Berg, A. Z. Snyder, and D. C. Van Essen, NeuroImage 80, 169 (2013).
  48. [48]L. J. Larson-Prior, R. Oostenveld, S. Della Penna, G. Michalareas, F. Prior, A. Babajani-Feremi, J.-M. Schoffelen, L. Marzetti, F. de Pasquale, F. Di Pompeo, J. Stout, M. Woolrich, Q. Luo, R. Bucholz, P. Fries, V. Pizzella, G. L. Romani, M. Corbetta, A. Z. Snyder, and WU-Minn HCP Consortium, NeuroImage 80, 190 (2013).
  49. [49]G. J. Tian, O. Zhu, V. Shirhatti, C. M. Greenspon, J. E. Downey, D. J. Freedman, and B. Doiron, bioRxiv, 2024.08.30.610535 (2026), version posted 8 June 2026.
  50. [50]P. J. Rousseeuw and C. Croux, Journal of the American Statistical Association 88, 1273 (1993).

Paper details

vlnc:W_h4qqu8q3
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Agricultural and Biological Sciences

ChatGPT, Claude

View statement
Looking for related papers…
Contents