Main

Motor cortex and its descending projections have expanded in certain mammalian lineages, seemingly because of the fitness conferred by the motor performance that they support1,2,3. Without normal motor cortical output, certain types of movement cannot be executed4,5,6. Many other types are slower, less agile and less effective, especially when dexterity is challenged or movements must adapt during execution7,8,9,[10](#ref-CR10 “Warren, R. A. et al. A rapid whisker-based decision underlying skilled locomotion in mice. eLife https://doi.org/10.7554/eLife.63596

(2021).“),11,12. Yet, when and how motor cortical output directly influences muscle activity through its descending projections to mediate this influence remains poorly resolved. The consequent ambiguity of direct motor cortical influence on muscles has stymied the development of more mechanistic models of descending motor control13.

Deficits from lesions and other inactivation of the motor cortex have not clearly resolved its involvement in movement execution. As the motor cortex is involved in motor learning14,15 and movement preparation or initiation16,17, deficits could reflect disturbance to these processes on which execution depends, rather than on the execution itself. Moreover, recent results indicate that motor cortical influence on muscle activity at the shortest latencies (10–20 ms in mice) differs from its influence on even slightly longer timescales (~50 ms)18.

During tasks requiring the motor cortex, existing results leave open several basic possibilities for the form that direct motor cortical influence on muscles could take. First, the motor cortex could drive the entirety of limb muscle activity patterns, with substantial compensation provided by other motor system regions after motor cortical disturbance. For example, when motor cortex needs to generate some muscle activity patterns that cannot be achieved by other regions19, it may assume all of the pattern-generating burden20. Second, the motor cortex could participate together with the rest of the motor system in generating motor output, without playing a role necessary to determining its pattern. Here loss of direct motor cortical influence on muscles would cause, at most, a nonspecific, fractional attenuation of motor output. Third, the motor cortex could selectively influence particular components of muscle activity, such that it informs (‘instructs’) ongoing muscle activity patterns and acts in a distinctly different way from the rest of the motor system. The loss of direct motor cortical influence would then cause changes in muscle activity that themselves vary as the state of muscle activity changes.

This ambiguity about the form of direct motor cortical influence on muscles has prevented resolution of other key issues related to the mechanisms of this influence. It remains unclear whether, on balance, motor cortical output only activates individual limb muscles or at times also suppresses their activity. The motor cortex is thought to drive online movement corrections and the adaptation of movements based on context9,21,22,23,[24](https://www.nature.com/articles/s41593-025-02093-z#ref-CR24 “Bollu, T. et al. Motor cortical inactivation impairs corrective submovements in mice performing a hold-still center-out reach task. J. Neurophysiol. https://doi.org/10.1152/jn.00241.2023

(2024).“); such a role could involve the activation and deactivation of individual muscles at different times to steer movement as the context requires.

It also remains unclear what components of motor cortical output drive muscle activity. Previous descriptions of motor cortical activity have focused on components that covary with limb muscle activity25,26 or movement parameters like joint angles or reach direction (kinematics)27,28. However, if motor cortical output does not contribute to all muscle activity patterns, but instead selectively alters them, we might expect that the components of motor cortical output driving muscle activity may not reflect muscle activity in total, but only some fraction of it. Moreover, motor cortical activity that covaries with muscle activity or kinematics in total may be a consequence of monitoring or predicting the body state29, perhaps to subserve aspects of motor control apart from directly driving muscle activation30. In line with this, muscle activity can be decoded from motor cortical activity during movements where this activity does not directly drive muscles18. Thus, the components of motor cortical activity responsible for its direct influence may differ from those to which functional roles have previously been attributed31,32,33,34,35.

Below we address these basic questions about direct motor cortical influence on limb muscles36 using mice. The three possible forms that direct motor cortical influence on muscles could take make different predictions about how the influence will vary across different muscle activity states during a given motor behavior. Thus, we measured this influence across muscle activity states during a behavior expected to depend on motor cortical output. Our characterization of this influence includes identification of states where motor cortical output activates and deactivates muscles. Finally, we describe components of motor cortical output that could be responsible for its influence on muscle activity.

A naturalistic climbing paradigm

As previous studies have implicated the motor cortex in adaptive limb movements in response to unpredictable sensory information[10](https://www.nature.com/articles/s41593-025-02093-z#ref-CR10 “Warren, R. A. et al. A rapid whisker-based decision underlying skilled locomotion in mice. eLife https://doi.org/10.7554/eLife.63596

(2021).“),13,37, we developed a behavioral paradigm that emphasizes such movements. Inspired by the natural movement repertoire of mice, we developed a paradigm in which head-fixed mice climbed across a series of handholds that extend radially from a wheel, thereby rotating the wheel (Fig. 1a–d, Extended Data Fig. 1 and Supplementary Video 1). After each handhold accessible to the right limbs has rotated 180° past the mouse, a linear actuator embedded within the wheel moves the handhold to a new, randomly chosen, mediolateral position; the left handholds remain fixed (Fig. 1e and Extended Data Fig. 1b–d). This ensures that the sequence of right handholds across which the mouse climbs is unpredictable (Fig. 1f), so sensory information must be used in real time to steer right limb movement. In this paradigm, water-restricted mice earn water rewards by climbing intermittently in bouts throughout hour-long daily sessions. The variation in the mediolateral position of the right handholds leads to a variation in the direction in which the right forelimb reaches (Supplementary Video 1). A broad range of body postures is expressed (Extended Data Fig. 1e).

Fig. 1: Head-fixed climbing paradigm.

a, Bird’s eye view of wheel apparatus for climbing. A shaft encoder measures the wheel’s angular position. Actuators randomize the position of each right handhold when they reach a point 180° away from the mouse. A ratchet ensures that the wheel rotates in only one direction. A slip ring commutes voltage signals to and from the actuators. b, A head-fixed mouse climbing in the paradigm. c, Frame of side-view video of a mouse climbing, with line plots connecting points tracked on the right forelimb and hindlimb from 50 sequential images (100 Hz) that have been overlaid. Line plot color reflects the time in the sequence. The points tracked were on the shoulder, elbow, wrist, last digit of the hand, hip, knee, ankle and edge of the foot. d, Same as c, but showing only the last frame in the sequence. e, Example sequence of right handhold positions over time, illustrating randomization. f, Autocorrelation of right handhold positions. g–i, Median (black dots, n = 9 mice) and first and third quartiles (whiskers) for the fraction of time spent climbing (g), median climbing velocity (h) and median climbing bout distance (i) across sessions. Gray lines in g–m are for individual mice. Session 1 indicates the first session after the mice had learned the pairing between climbing and reward, when reward dispensation switched from experimenter to computer control. j,k, Median (black dots, n = 9 mice) and first and third quartiles (whiskers) for the first (j) and second (k) principal angles between electromyographic time series collected during the twentieth climbing session and each of the first 20 climbing sessions. l,m, Median (black dots, n = 9 mice) and first and third quartiles (whiskers) for the sample entropy of muscle activity (l) and limb kinematics (m) time series across sessions. For each session, we took the mean across-sample entropy values for each muscle or for the xand y positions of each tracked limb point. The sample entropy measures the regularity in the time series40.

As it may be relevant to motor cortical involvement38, we assessed how the performance of climbing mice varied across daily sessions. To look for progressive improvement in a performance, we examined the measures of bout length and climbing speed, because the reward scheme depends on them. We found that, after mice are acclimated to head fixation (two sessions) and taught the pairing between climbing and reward (one to three sessions), there was little change, on average, in the time spent climbing (Fig. 1g), the velocity of climbing (Fig. 1h) and the distance of climbing bouts (Fig. 1i). To assess whether forelimb muscle activity patterns change progressively across sessions, we computed the principal angles between the first two principal components (PCs) for the activity of four muscles in the right forelimb39 during each session (two PCs by T time points; Extended Data Fig. 1f–h). Comparing each of the first 20 daily sessions to the twentieth session, we found that the first principal angle was generally low, averaging <2° (Fig. 1j,k). Although adjacent sessions appeared more similar (see the lower angles for session 19), there was little indication that increasingly distant sessions were increasingly more dissimilar, which would be expected for a progressive change in muscle activity. We also found that the stereotypy in both muscle activity and limb kinematics did not show clear signs of increasing across sessions40 (Fig. 1l,m and Extended Data Fig. 1i). Thus, after beginning to climb for rewards, mice do not appear to progressively develop climbing skills specific to our paradigm nor does muscle activity appear to change progressively across sessions. These results indicate that our climbing paradigm differs from those in which participants learn new tasks and become increasingly skillful and stereotyped with repeated training14,41.

Quantifying direct motor cortical influence during climbing

We next sought to quantify direct motor cortical influence on contralateral forelimb muscles across the range of muscle activity states expressed during climbing. Such an influence is not seen during treadmill walking18 and mice can still learn new stereotyped locomotor behaviors during split-belt treadmill adaptation after a bilateral motor cortical lesion42. However, lesion and pharmacological inactivation of the motor cortex does affect the execution of new locomotor adaptations in mice[10](https://www.nature.com/articles/s41593-025-02093-z#ref-CR10 “Warren, R. A. et al. A rapid whisker-based decision underlying skilled locomotion in mice. eLife https://doi.org/10.7554/eLife.63596

(2021).“),42,43. Given the different form and predictability of the movements elicited in our climbing paradigm, motor cortical influence was unclear a priori.

While mice (n = 8) were actively climbing, we sporadically and briefly inactivated the left caudal forelimb area (CFA, forelimb primary motor cortex + primary somatosensory cortex (M1 + S1)) at random. We used transgenic mice that express Channelrhodopsin-2 in all cortical inhibitory interneurons, applying occasional 25-ms blue light pulses that covered the surface of CFA (10 mW mm−2; Fig. 2a). This yields an ~50% activity reduction across cortical layers within 7 ms, which reaches 90–95% in <20 ms (refs. 18,44). Light pulses were always >4 s apart to allow recovery of neural activity between events; on average, ~100–200 trials were collected during each daily session (11–37 sessions per animal). Equivalent events without blue light were notated in recordings to serve as control trials. Random trial timing ensured broad coverage of the muscle activity states that each mouse expressed during climbing. We found that inactivation and control trial averages diverged ~10 ms after light onset, which reflects the shortest latency at which CFA output influences muscles18 (Fig. 2a–c and Extended Data Fig. 2a–d). We also found that inactivation effects were similar in form across mice (Fig. 2a and Extended Data Fig. 2b) and strikingly consistent both within and across sessions (Extended Data Fig. 2e–g). Thus, CFA directly influences muscle activity during climbing, as we previously observed in mice performing a trained forelimb reaching task18.

Fig. 2: Comprehensive assessment of CFA influence across muscle activity states.

a, Control (n = 1,671 trials) and inactivation (795 trials) trial averages (mean ± s.e.m.) for 4 muscles in 1 mouse. The inset showing the brain schematic is adapted from ref. 18. Vertical cyan bars in a–d, g and k indicate the 25-ms epoch of blue light applied to the CFA and gray dotted lines are 10 ms after light onset. As we z-scored muscle activity measurements using the mean and s.d. from each given session, here and throughout we express measurements in s.d. values of the recorded signal. b, Control (18,397 trials) and inactivation (9,029 trials) trial averages for all 8 mice. c, Mean absolute difference between inactivation and control trial averages across all four muscles. Light-gray lines are individual animals and the solid black line is the mean across animals. For baseline subtraction, control trials were resampled to estimate the baseline difference expected by chance. d, Example of muscle activities and their corresponding first derivatives surrounding trials that were used for creating muscle activity state maps. The weight epoch immediately precedes the start of effects. e, Example of muscle activity state map from one animal. Larger, connected dots show examples of states for sequential overlapping epochs from individual trials. Pairs were chosen based on their similarity during the weight epoch. f, Grid overlaying a map, including only points from the weight epochs used for weighting trials in grid point trial averages. g, Schematic of the calculation of the inactivation effect at each grid point from the control (black) and inactivation (cyan) trial-averaged muscle activity. ΔC and ΔL reflect the slopes of lines connecting the average activity just before to just after the inactivation effect begins, for control trials and light trials, respectively. h, Schematic illustration of the effect size on a plot of ΔL versus ΔC. Their difference is proportional to the distance from the identity line. i, Map in which each grid point colored by the mean distance, in the full 80 dimensional space, between all pairs of embedded state vectors, with each individual distance weighted by a Gaussian function of the pair’s mean distance from the grid point on the 2D map. The Gaussian function is the same as that used for inactivation maps. j, Inactivation effect maps for the four recorded muscles. The color scale maximum and minimum reflect the maximum and minimum effect sizes across all four muscles collectively. j–m, Representative results from one mouse. k, Grid point-averaged muscle activity from control (gray, mean ± s.e.m.) and inactivation (cyan, mean) trials, for three example grid points from the maps in j. l, Maps of P values computed for inactivation effects at each grid point. The q values (gray overlay) reflect the expected false discovery rate below the corresponding P value46. m, Maps showing the average activity for the four recorded muscles at each grid point. The color scale maximum and minimum reflect the maximum and minimum activity level for each muscle separately. The darker blue regions reflect states where the given muscle is inactive. The darker red regions reflect states where the given muscle is highly active, up to between 2.7 s.d. and 5.5 s.d. values above the mean. dist., distance; max., maximum; min., minimum.

To initially gauge whether direct motor cortical influence varies throughout climbing, we examined the effects of CFA inactivation during three stereotypical features of climbing: pulling a handhold down with the right forelimb, reaching the right forelimb up and palpation of the right handhold while grasping it (Extended Data Fig. 3). We assembled trial averages for muscle activity and limb kinematic time series aligned on trial onsets that occurred during each feature. The effect magnitude appeared to be vary across features. The effects were also more prominent in trial-averaged muscle activity than limb kinematics across all three features, which we explore further below.

We thus proceeded to more comprehensively assess CFA influence at different muscle activity states during climbing. We first sought a means for collecting together inactivation and control trials that began at similar muscle activity states, so that we could average across them. Plotting trials according to linear functions of muscle activity at trial onset led to an uneven distribution of trials across plots (Extended Data Fig. 3d,e). This was suboptimal for efficiently utilizing the statistical power afforded by our trials to differentiate CFA influence across states (Methods). We also suspected that this statistical power would be improved if trials were grouped together based on the time-varying pattern of muscle activity right before trial onset, because CFA influence could depend on this pattern.

We thus defined states using the activity of all four muscles over 50-ms epochs, rather than individual time points, and used Uniform Manifold Approximation and Projection (UMAP)45 to generate a two-dimensional (2D) map of the states expressed by each mouse, where similar states are close together. To ensure proximity on maps among states visited that are close in time, we defined epochs that overlapped in time. The muscle activity traces surrounding each trial, together with their corresponding first derivatives, were subsampled in 5-ms increments and divided into overlapping 50-ms epochs that began every 10 ms (Fig. 2d). For each 50-ms epoch, the muscle activity and first derivative trace segments were concatenated into a single vector (8 segments × 10 time bins = 80 elements). UMAP was then applied to map vectors onto two dimensions (Fig. 2e). On the resulting maps, embedded state vectors (points) from successive epochs form trajectories that reflect the sequence of states surrounding control and inactivation trials.

To measure direct CFA influence at different muscle activity states, we quantified the immediate inactivation effects for trials starting from states within local neighborhoods on the maps. We first defined a grid over each map (Fig. 2f). For each muscle, we computed its trial-averaged activity at each grid point, separately, for inactivation and control trials. For these averages, we used all trials, but we weighted each by a Gaussian function of the Euclidean distance between the given grid point and the point for the epoch just before an inactivation effect could begin on the given trial (−40 ms to +10 ms from trial onset: ‘weight epoch’; Fig. 2d). Trial weights are hence not influenced by inactivation effects. As a consequence, weight epoch states from control and inactivation trials are similarly distributed across maps (Fig. 2f). We set the Gaussian s.d. as roughly 10% of the map width, so trial averages are heavily weighted toward trials beginning at states close to the given grid point. Only grid points close to a substantial number of weight epoch states were subsequently considered (‘valid grid points’; Methods). For each muscle, we measured separately the size of the inactivation effect at each valid grid point as the difference between the rate of change in inactivation and the control trial averages from 0 ms to 20 ms after trial onset (Fig. 2g,h). We then plotted the resulting effect sizes at each valid grid point across the map, producing an ‘inactivation effect map’ (Fig. 2j and Extended Data Fig. 4a–d). Maps for different muscles in a given mouse show wide variation in the magnitude and sign of inactivation effects across grid points (Fig. 2j,k). We resampled from control trials to compute a P value for each grid point’s effect size (Fig. 2l and Extended Data Fig. 4a–c). The map structure was not strongly dependent on the choice of key parameters (Extended Data Fig. 4f–i).

As UMAP is nonlinear, it is not clear how the distance across maps will correspond to differences in muscle activity state. To address this and clarify muscle activity levels at different map locations, we plotted the average activity of each individual muscle at each grid point using the same Gaussian weighting method as above (Fig. 2m). These plots showed smooth and gradual variation across grid points. We also plotted the average similarity between nearby state vectors across maps (Fig. 2i and Extended Data Fig. 4e). The only prominent variation that we observed was a gradual increase in pairwise distance from the map center to the edges; there was no indication of abrupt changes in pairwise distance. Thus muscle activity state varies smoothly across state maps at the resolution of our inactivation effect maps.

The CFA acts primarily by selectively exciting physiological flexors

To distinguish among the three possible forms that direct motor cortical influence on muscles could take, we then analyzed the inactivation effect maps. We first generated histograms of the P values computed for effects on each muscle at all valid grid points. These histograms consistently showed a skew toward zero, reflecting a substantial fraction of grid points where the null hypothesis of no effect was false (Fig. 3a,b). From these distributions, we estimated the fraction of grid points showing effects46 (Fig. 3c). The mean estimated fractions were 0.62, 0.22, 0.73 and 0.37 for elbow flexor, elbow extensor, wrist extensor and wrist flexor muscles, respectively. These estimates were significantly above zero for all four muscles. Control maps generated from comparisons between separate sets of control trials yielded uniform distributions, as expected under the null hypothesis (Extended Data Fig. 5a). These results show that the direct influence of CFA on muscles is specific to a subset of muscle activity states. This is not consistent with either the CFA driving the entirety of limb muscle activity or the CFA having a nonspecific effect on muscles. Rather, CFA appears to selectively influence particular components of muscle activity.

Fig. 3: The CFA selectively excites physiological flexors.

a,b, Distributions of P values for inactivation effects on each muscle for all grid points in one mouse (a) and across all eight mice (b). Left: elbow flexor (first), elbow extensor (second). Right: wrist extensor (second) and wrist flexor (first). The error bars in b indicate the s.e.m. c, Estimated fraction of grid points at which the null hypothesis of no effect is false, calculated from distributions of P values for eight individual mice (black circles) and the mean across mice (red bars). Values were significantly above 0 for all muscles (P = 0.008 or P = 0.016, two-sided Wilcoxon’s signed-rank test). The estimated fraction of grid points yielding false null hypotheses was significantly greater for the elbow flexor (P = 0.007, two-sided Wilcoxon’s rank-sum test) and wrist extensor (P = 0.015), compared with their respective antagonists. d, For one mouse, the 2D autocorrelation (autocorr.) for inactivation effect maps and control maps generated with only control trials (top) and scatterplots of correlation values versus their spatial offset (lag from zero offset). pts., points. e, Difference between inactivation effect maps and control maps in their mean autocorrelation over spatial offsets from 0 grid points to 20 grid points for 8 individual mice (black circles) and the mean across mice (red bars). Differences were significantly >0 for all four muscles (P = 0.008, two-sided Wilcoxon’s signed-rank test). The magnitude of CFA influence across states differed significantly between muscles in six of eight mice (P < 0.004, P = 0.19 and P = 0.83 in two other mice, Kruskal–Wallis test). f, For one animal, scatterplots of inactivation effect size versus muscle activity at trial onset (averaged from −40 ms to +10 ms relative to onset). Each point reflects a different grid point. The R2 is for a linear fit (red). g, R2 for linear fits to scatterplots of inactivation effect size versus muscle activity at trial onset for eight individual mice (black circles) and the mean across mice (red bars). The residuals were significantly nonuniform (P < 10−10 for all mice, two-sided Kolmogorov–Smirnov test). h, Effect size distributions for all grid points across all eight mice, separately for inactivation effect maps and control maps. i,j, Effect size distributions for all significant grid points from all eight mice (i) and one mouse (j). Left: elbow; right: wrist. k, Same as i, but zoomed in to clarify rarer effects. l, Grid point-averaged muscle activity (mean ± s.e.m.) from control and inactivation trials, for four example grid points where inactivation significantly increased muscle activity in four different mice. The three on the left are for the elbow extensor and the one on the right is for the wrist flexor. m, The 2D correlation between inactivation effect maps for different muscles for one mouse (top) and the means across all eight mice (bottom).

To better characterize this selective influence, we next assessed whether CFA influence varies in magnitude across muscle activity states. If this were true, then the 2D autocorrelation of inactivation effect maps should be significantly above what is expected by chance. We computed the 2D autocorrelation of both inactivation effect maps and control maps, observing substantially heightened autocorrelation in the former (Fig. 3d). To assess whether these differences were significant, we computed the mean difference between inactivation effect map autocorrelation and that of control maps, averaged over spatial lags up to 20 grid points (Fig. 3e). These differences were significantly >0 for all four muscles. We also found that the magnitude of CFA influence across muscle activity states differed significantly between muscles in six out of eight mice. The magnitude of CFA influence was not simply proportional to the magnitude of the muscle activity; the coefficient of determination (R2) for linear fits to effect size versus muscle activity magnitude was low (Fig. 3f,g and Extended Data Fig. 5b–e) and the residuals were significantly nonuniform.

A number of previous observations indirectly suggest that primary motor cortex may preferentially control certain muscle groups more so than their antagonists47,48,49. We thus compared the distributions of effect sizes across grid points for each muscle. We found larger deviations from control effect sizes for the elbow flexor and wrist extensor (Fig. 3h). The estimated fraction of grid points showing effects (false null hypotheses, Fig. 3c) was significantly greater for the elbow flexor (61% higher) and wrist extensor (43% higher), compared with their respective antagonists. This indicates that CFA output preferentially influences elbow flexors and wrist extensors, which can be grouped together as physiological flexors because of their coactivation during both locomotion and the flexion reflex50.

We also assessed whether these differences in effects on muscles might extend to the direction of effects. Reduction or elevation of muscle activity after inactivation indicates that CFA output activates or suppresses muscle activity, respectively. We examined the effect sizes that were significantly different from zero, finding that, for the physiological flexors, effects were always a reduction in muscle activity (Fig. 3i–k). However, the elbow extensor exhibited both reduction and elevation and the wrist flexor showed elevation in a small fraction of states as well (Fig. 3k). This can be seen in the trial-averaged muscle activity for individual grid points from inactivation effect maps (Fig. 3l). We also observed that inactivation effect maps for the physiological flexors were more highly correlated compared to those for all other pairs of muscles (Fig. 3m), suggesting a greater degree of coordinated control of these muscles. Collectively, these results indicate that the CFA influences muscles to varying degrees and only at some muscle activity states (that is, the influence is selective). CFA’s influence is therefore primarily an activation of physiological flexors; only the elbow extensor, where influence was relatively infrequent, shows a balance of activation and suppression.

Weak covariation between CFA influence and gross kinematic state of the forelimb

A number of previous observations suggest that the motor cortex may, to some extent, control the limb via commands that dictate its kinematics rather than muscle activation51,52, although this remains controversial32,53. If the CFA were dictating contralateral forelimb kinematics, we reasoned that direct CFA influence on muscles should correlate with the orientation of the contralateral forelimb. We therefore probed for this correlation.

Here we mimicked the approach that we took to assess how CFA influence covaries with muscle activity state. We computed new state maps with UMAP using vectors composed of the horizontal and vertical positions of sites on the right forelimb tracked from a video (Fig. 4a,b and Extended Data Fig. 6a–c). Nearby points on these maps thus reflect 50-ms epochs of limb kinematics that are similar. The resulting 2D maps separated states that correspond to different limb orientations into different map regions (Fig. 4c), with the cyclic changes in limb orientation during iterative climbing ordered around the map. Using these maps, we quantified the effects of CFA inactivation on each forelimb muscle at grid points covering the maps as above. Histograms of the inactivation effect sizes across all grid points showed deviations from the controls (Fig. 4d). However, the P-value distributions for effects on each muscle showed very limited skew toward zero, indicating discernible effects on only a small fraction of grid points (Fig. 4e). Thus, CFA influence on muscles does not covary with forelimb orientation nearly as well as it does with muscle activity state.

Fig. 4: Gross forelimb kinematics capture CFA influence worse than muscle activity.

a, Time series of tracked forelimb sites and their corresponding first derivatives surrounding inactivation and control trials windowed into overlapping segments. The segments are used to create 2D embedding via UMAP. b, Image showing the locations of the eight sites tracked on the forelimb, according to the color code in a. c, Example map of forelimb orientation states from one mouse, along with the trial-averaged positions of the forelimb sites at selected grid points within the map (red circles). As the video was captured at 100 Hz, time series segments used here had 10-ms spacing between points instead of 5 ms, as in Fig. 2. d–g, Distributions of the sizes of inactivation effects on muscles (d), P-value distributions for inactivation effects on muscles (e), distributions of the sizes of inactivation effects on four main forelimb sites (f) and P-value distributions for inactivation effects on four main forelimb sites (g), calculated using forelimb orientation maps, across all grid points and all eight mice. The error bars in e and g indicate the s.e.m. h, Histograms of Pearson’s correlatio