Neuron

Article

A Common Structure Underlies Low-Frequency

Cortical Dynamics in Movement, Sleep, and Sedation
Thomas M. Hall,1 Felipe de Carvalho,1 and Andrew Jackson1,*
1Institute of Neuroscience, Newcastle University, Framlington Place, Newcastle NE2 4HH, UK

*Correspondence: [email protected]
http://dx.doi.org/10.1016/j.neuron.2014.07.022
This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/3.0/).

SUMMARY Brain-machine interface (BMI) studies have found low-frequency

bands to be particularly informative for decoding direction from
Upper-limb movements are often composed of local field potentials (LFPs; Rickert et al., 2005; Bansal et al.,
regular submovements, and neural correlates of 2011), MEGs (Waldert et al., 2008), and EEGs (Waldert et al.,
submovement frequencies between 1 and 4 Hz 2008; Bradberry et al., 2010; but see Antelis et al., 2013).
have been found in the motor cortex. The temporal It has been argued that submovements reflect intermittent
profile of movements is usually assumed to be deter- corrections driven by visual feedback of errors (Craik, 1947; Miall
et al., 1986, 1993) and that their frequency should therefore
mined by extrinsic factors such as limb biome-
be determined by extrinsic factors such as feedback loop de-
chanics and feedback delays, but another possibility
lays. In support of this ‘‘extrinsic hypothesis,’’ submovements
is that an intrinsic rhythmicity contributes to low are locked to eye movements (McAuley et al., 1999) and often
frequencies in behavior. We used multielectrode re- disappear in the absence of vision (Miall et al., 1993; McAuley
cordings in monkeys performing an isometric move- et al., 1999; but see Doeringer and Hogan, 1998), whereas the
ment task to reveal cyclic activity in primary motor introduction of artificial feedback delays alters their frequency
cortex locked to submovements, and a distinct oscil- (Miall et al., 1986; Miall and Jackson, 2006). Nevertheless, sub-
lation in premotor cortex. During ketamine sedation movements are not restricted to tracking tasks, and a natural
and natural sleep, cortical activity traversed similar rhythmicity is observed across diverse upper-limb behaviors
cycles and became synchronized across areas. (Kunesch et al., 1989) including self-paced isometric drawing
Because the same cortical dynamics are coupled to (Massey et al., 1992) and finger tapping (Schöner and Kelso,
1988). Moreover, low-frequency cortical oscillations have long
submovements and also observed in the absence
been associated with slow-wave sleep, when large K complex
of behavior, we conclude that the motor networks
potentials signifying transitions from down to up states of the
controlling the upper limb exhibit an intrinsic period- cortex (Colrain, 2005; Cash et al., 2009) are accompanied by
icity at submovement frequencies that is reflected in bursts of activity in the delta (1–4 Hz)-frequency range (Amzica
the speed profile of movements. and Steriade, 1997). At least two mechanisms contribute to
these delta oscillations: intrinsic currents that cause bursting
patterns in thalamic relay cells (Amzica et al., 1992; Destexhe
INTRODUCTION and Sejnowski, 2003) and a second, purely cortical circuit (Amz-
ica and Steriade, 1998; Carracedo et al., 2013).
Movement requires coordinating dynamic patterns of activity Therefore, it remains possible that oscillatory properties of
across multiple muscles. Many simple rhythmic behaviors are cortical (and perhaps thalamic) circuits contribute to low-fre-
controlled by specialized central pattern generator (CPG) net- quency rhythms in movement, functioning much like a CPG.
works in the spinal cord or brainstem with intrinsic oscillatory Recently, this ‘‘intrinsic hypothesis’’ has been proposed to
characteristics (Grillner, 2006; Kiehn, 2006). However, goal- explain the complex, multiphasic profiles of motor cortical firing
directed upper-limb movements under cortical control can also rates observed during reaching (Churchland et al., 2012; Shenoy
exhibit rhythmicity. When tracking a moving target, trajectories et al., 2013). The high-dimensional neural state was projected
are composed of multiple submovements (Craik, 1947) at a fre- onto a plane revealing low-frequency cycles, even though move-
quency (usually one to four per second) that is largely indepen- ments in this case were not overtly rhythmic. It was proposed
dent of movement speed (Miall et al., 1986; Roitman et al., that this dynamical structure reflects ‘‘an engine of movement’’
2004; Pasalar et al., 2005; Selen et al., 2006). An oscillation at (Churchland et al., 2012), and could be reproduced by a recur-
around 3 Hz has also been reported in the kinematics of finger- rent neural network model trained to generate muscle patterns
tracking movements (McAuley et al., 1999). The motor cortical given static initial inputs and no sensory feedback (Shenoy
electroencephalogram (EEG) is phase locked to submovements et al., 2013).
(Dipietro et al., 2011), and coherence spectra between the mag- The intrinsic hypothesis suggests that low-frequency cortical
netoencephalogram (MEG) and movement speed show peaks dynamics may be preserved across different movements
around 3 Hz during visuomotor tracking (Jerbi et al., 2007). and resemble spontaneous delta oscillations during sleep.

Neuron 83, 1185–1199, September 3, 2014 ª2014 The Authors 1185

 Neuron
Low-Frequency Motor Cortical Dynamics

A LFPs and
Figure 1. Low-Frequency Cortical Dynamics during Isometric Task
neurons Performance
(A) Schematic of isometric wrist-torque task.
(B) Top: position of cursor (solid lines) and target (dashed lines) during a typical
trial. Middle: cursor speed (time derivative of radial position from the origin).
Bottom: raw (black) and rectified, smoothed (red; not to scale) EMG from a
wrist extensor muscle, extensor carpi ulnaris (ECU). Five submovements
occurred during this trial, indicated by arrowheads.
(C) Top: unfiltered, surface-referenced LFP from a representative electrode in
M1 during task performance. Middle: low-pass-filtered, mean-subtracted LFP
from all electrodes in the M1 array, ordered and color coded according to
2D torque EMGs phase relative to submovements. The bottom trace (blue) corresponds to the
unfiltered signal shown at the top (black). Bottom: spike rasters for seven M1
B Recording during isometric task neurons.
(D) Top: first two principal components (LFP-PCs) calculated from low-pass-
100 Position (%) Target
Flexion-extension filtered, mean-subtracted M1 LFP. Bottom: second LFP-PC and speed profile
0 overlaid.
Radial-ulnar Data are from monkey D. See also Movie S1.
−100

500 Speed (%/s) Therefore, we compared motor cortex activity in monkeys

0 during an isometric movement task, while retrieving food from
a Klüver board, during natural sleep, and under ketamine seda-
−500
tion. We found clear evidence for a common LFP correlation
1 ECU-EMG (mV) structure that could be explained by a single model of 3 Hz oscil-
latory dynamics underlying all behavioral states. Our results
0
thereby unify two previously unrelated phenomena: low-fre-
−1 quency structure in movement kinematics and low-frequency
1 sec
oscillations during slow-wave sleep, providing a new insight
C Unfiltered M1 LFP
into how the dynamics of cortical networks influence complex
upper-limb behaviors.
200 µV

RESULTS

Isometric Center-Out Wrist Movements Are Composed

Low-pass filtered, mean-subtracted M1 LFPs of Rhythmic Submovements
Three monkeys controlled the 2D position of a cursor with iso-
metric wrist torque to acquire targets in a center-out fashion
(Figure 1A). Similar to isometric trajectories in humans (Massey
200 µV

et al., 1992), the movements made by the monkeys were often

composed of multiple, regular submovements. The representa-
tive single trial shown in Figure 1B shows five submovements
between the go cue and successful acquisition of the target, ap-
M1 Neurons
pearing as peaks in the radial speed of the cursor. The distribu-
tion of intersubmovement intervals (Figure 2A), as well as their
autocorrelation structure (Figure 2B), revealed a tendency for
submovements to occur rhythmically at a frequency of around
3 Hz in all three animals. Movement intermittency was also
evident in the electromyogram (EMG) from wrist muscles
1 sec involved in the task, with a peak in coherence between radial po-
sition and rectified EMG at 3 Hz (Figure 2C).
D LFP-PC1
LFP-PC2
Low-Frequency LFP Oscillations Are Phase Locked
to Submovements
LFPs from multiple electrodes in primary motor cortex (M1) were
Speed low-pass filtered and mean referenced, revealing slow oscilla-
LFP-PC2 tions with a phase that varied across electrodes (Figure 1C). Prin-
cipal component analysis (PCA) reduced the LFP to two principal
components (LFP-PCs; Figure 1D) capturing orthogonal projec-
1 sec tions of the dominant oscillatory mode. There was a striking cor-
relation between LFP-PCs and submovements, evident even in

1186 Neuron 83, 1185–1199, September 3, 2014 ª2014 The Authors

Neuron
Low-Frequency Motor Cortical Dynamics

A B Figure 2. Relationship between LFP-PCs

(relative to uniform distribution)

Inter-submovement interval histogram Submovement autocorrelation function

Probability of submovement
20 3 and Movement Kinematics
Monkey D Monkey D (A) Intersubmovement interval histograms for
Count (% of total)

Monkey R Monkey R representative sessions in all three animals.

Monkey S 2 Monkey S (B) Autocorrelation histogram of intervals between
10
all pairs of submovements in the same trial (be-
1 tween go cue and end of successful hold). Histo-
grams are normalized by the interval distribution
0 0 expected for a uniform (Poisson) distribution with
0 0.5 1 1.5 2 0 0.5 1 1.5 2 the same rate. The peak at around 300 ms reveals
Interval (s) Interval (s) the underlying rhythmicity of submovements.
(C) Coherence spectra between radial cursor po-
C Monkey D example session D E sition and rectified EMG from four wrist muscles:
0.4 0.4 extensor carpi ulnaris (ECU), extensor carpi radi-
ECU-Position 150
ECR-Position PC1-Speed alis (ECR), flexor carpi ulnaris (FCU), and flexor
FCU-Position PC2-Speed carpi radialis (FCR). Data are from monkey D

LFP-PC2 (µV)
Coherence

Coherence

FCR-Position comprising 320 trials with 2,063 submovements.

0.2 Mean 0.2 0 (D) Coherence spectra between LFP-PCs and
radial cursor speed across the same session.
(E) LFP-PC trajectory for 2 s of the representative
0 0 trial shown in Figure 1. Circles indicate times of
−150
5 10 0 5 10 −150 0 150 peak cursor speed.
Frequency (Hz) Frequency (Hz) LFP-PC1 (µV)
(F) LFP-PC trajectories aligned to peak speed of
F G H submovements and averaged across nine equal-
8 0.2
Rotation frequency (Hz)

100 sized groups sorted by peak cursor speed. Tra-

Areal velocity (mV2/s)

jectories are plotted for 200 ms on either side of

LFP-PC2 (µV)

time of peak speed (indicated by circles) and color

R = 0.96
4 0.1 coded from black to red to yellow according to
0
R = 0.24 cursor speed.
(G) Rotational frequency of average trajectories for
different submovement speeds, calculated at the
−100 0 0
time of peak cursor speed.
−100 0 100 0 300 600 0 300 600
(H) Areal velocity (area swept out per unit time) of
LFP-PC1 (µV) Cursor speed (%/s) Cursor speed (%/s)
average trajectories for different submovement
I Monkey R example session J K speeds, calculated at the time of peak cursor
8 0.3 speed.
Rotation frequency (Hz)

150
Areal velocity (mV2/s)

(I–K) Equivalent analysis of a representative ses-

LFP-PC2 (µV)

0.2 sion with monkey R comprising 150 trials with 823

R = 0.98 submovements.
0 4
R = 0.15
(L) Average 2D LFP-PC trajectories for submove-
0.1
ments, binned and color coded according to the
direction of cursor movement. Arrows in the inset
−150 0 0 indicate the central direction of movement for
−150 0 150 0 250 500 0 250 500
each bin.
LFP-PC1 (µV) Cursor speed (%/s) Cursor speed (%/s) (M) Average LFP-PC trajectories in the plane of
L Monkey D example session M N 0.4 All sessions PC2 and PC3. The trajectories revolve around
100 100 slightly different angular velocity vectors, indicated
by arrows.
Angular CoD
LFP-PC3 (µV)
LFP-PC2 (µV)

(N) Angular coefficient of determination for sub-

Extension 0.2 movement direction, decoded from the orientation
0 0
Ulnar Radial of LFP-PC angular velocity. Leave-one-out cross-
P=0.05
validation was performed over every submove-
Flexion
ment in each data set. The plot shows the average
−100 −100 0
−100 0 100 Slow Fast CoD for validation submovements in each speed
−100 0 100
LFP-PC2 (µV) Cursor speed group, based on data from 13 sessions in three
LFP-PC1 (µV)
monkeys. Also shown is the average 95%
percentile from shuffled data.
See also Movies S2 and S3.

single trials (Figure 1D, bottom). Coherence analysis over the time onto the PC plane (Movie S1 available online). Due to
whole session confirmed strong correlation between LFP-PCs the 90
phase difference between components, this trajectory
and cursor speed at the 3 Hz frequency of submovements was cyclic with constant direction of rotation. Each submove-
(Figure 2D). ment was associated with a single LFP cycle, and the peak
The relationship between LFP oscillations and submove- cursor speed occurred at a similar phase within each cycle
ments was visualized by projecting the LFP trajectory over (Figure 2E).

Neuron 83, 1185–1199, September 3, 2014 ª2014 The Authors 1187

 Neuron
Low-Frequency Motor Cortical Dynamics

Submovement Kinematics Can Be Decoded from Areal Phase of LFP Oscillations Relative to Submovements
Velocity of LFP Trajectories and K Complexes
To examine how LFP-PC trajectories were related to submove- Figure 4 compares cortical activity aligned to the peak speed
ment kinematics, we binned submovements into nine equal of submovements (during task performance) and aligned to the
groups according to peak cursor speed. Average LFP-PC trajec- peak of K complexes (under ketamine sedation). Both events
tories for each group (Figures 2F and 2I; Movie S2) had constant were associated with phasic bursts of neural activity (Figure 4B)
direction and frequency of rotation (Figures 2G and 2J) but an and phase-locked low-frequency LFP oscillations in M1 (Fig-
areal velocity (area swept out per unit time about the origin; ure 4C), which were an order of magnitude larger in the sedated
see Equation 2 in Experimental Procedures) that increased state. Note that the same color scheme is used throughout Fig-
with cursor speed (Figures 2H and 2K). When binned according ures 1 and 4C to represent the LFP phase relative to submove-
to the direction of submovements, average LFP trajectories ments during task performance but each electrode shows a
in the PC plane appeared similar (Figure 2L). However, plotting similar phase relative to the K complex under sedation. As a
the trajectories in 3D PC space revealed a subtle variation in result, average LFP-PC trajectories followed similar rotational
the axis of rotation for different submovement directions (Fig- cycles aligned to both submovements and K complexes (Fig-
ure 2M; Movie S3). ure 4D). We calculated the circular-circular correlation coeffi-
In three dimensions, areal velocity is conveniently represented cient (rCC) across electrodes of phase relative to submovements
by a vector aligned to the axis of rotation. We hypothesized that (during task performance) against phase relative to K complexes
its magnitude should encode information about submovement (under sedation). For the sample pair of movement and sedation
speed, whereas its orientation might be informative of submove- sessions shown in Figure 4C, these phases were highly corre-
ment direction. We tested this directly by decoding speed (or lated (n = 10, rCC = 0.81, p = 0.025; Figure 4E), and across all
direction) from the magnitude (or orientation) of areal velocity the data sets this correlation was significant (p < 0.05) in 11/13
vectors associated with individual submovements (see Experi- pairs of sessions in three monkeys (mean ± SD; rCC = 0.75 ±
mental Procedures). Decoding performance for 13 data sets 0.25; Table S1). Pooling all the sessions for each animal also
across three monkeys is summarized in Table S1, using a coef- yielded significant correlation (monkey D: n = 52, rCC = 0.63, p =
ficient of determination (CoD) between zero (chance decoding) 2 3 105; monkey R: n = 36, rCC = 0.49, p = 0.002; monkey S:
and one (perfect decoding). In every case, we obtained signifi- n = 45, rCC = 0.81, p = 4 3 106; Figure 4F). Moreover, advancing
cant decoding of the speed from areal velocity magnitude, with several microwires from the most superficial depth down through
mean (±SD) CoD = 0.30 ± 0.13. Decoding submovement direc- the gray matter revealed that both submovement- and K com-
tion from areal velocity orientation was significant in 12/13 plex-related potentials underwent polarity reversals at the same
sessions, with mean CoD = 0.15 ± 0.07. As might be expected, depth, indicating a common cortical source (Figure S1).
direction decoding was better for faster submovements (Fig- In summary, the same patterns of cortical activity seen during
ure 2N), with a CoD of 0.26 ± 0.17 for the fastest submovements. isometric movements (and related to submovement kinematics)
Although statistically significant, this nevertheless corresponds also arose endogenously in the absence of behavior, suggesting
to an average decoding error of approximately 75
, only slightly that intrinsic circuits rather than extrinsic sensorimotor feedback
better than chance (90
average decoding error). loops impose this dynamical structure on low-frequency cortical
In summary, the rhythmic structure of submovements is re- activity. Because isometric torque tracking and ketamine seda-
flected in low-frequency M1 LFPs and can be revealed using tion are somewhat unnatural experimental conditions, we pro-
PCA to reduce the dimensionality of the multichannel data. In ceeded to examine whether the same low-frequency dynamics
the space defined by the first three PCs, each submovement is were also present in brain activity during more naturalistic condi-
associated with a cyclic LFP trajectory, and the peak of the sub- tions including sleep and unrestrained reach-to-grasp.
movement occurs at a consistent phase of the cycle. The areal
velocity of the trajectory is proportional to the speed of the sub- Low-Frequency LFP Oscillations during Natural Sleep
movement, whereas the axis of rotation provides statistically sig- and Free Reaching Movements
nificant (albeit modest) information about the direction of Two of our subjects often fell asleep at the end of recording
movement. sessions, providing an opportunity to examine low-frequency
activity during natural sleep. As the eyes closed, large-amplitude
Low-Frequency LFP Oscillations during Ketamine slow waves were observed in the LFP (Figure 3B), of a compara-
Sedation ble amplitude to sedation recordings and approximately an order
On separate days, we recorded from the same electrodes during of magnitude greater than that seen in the awake state. However,
ketamine sedation (Figure 3A). M1 LFPs exhibited typical signa- cortical activity appeared disorganized and lacked clear up-/
tures of slow-wave sleep including spindles and large K complex downstate transitions or K complexes, which is consistent with
potentials, thought to reflect transitions between cortical down stage 1 sleep.
and up states. Consistent with this, most neurons had low firing In addition, we collected data when the same animals
rates prior to each K complex, and fired particularly strongly retrieved food from wells in a Klüver board with the arm unre-
during its rising phase and peak. Each K complex was associated strained. In general, LFPs showed less rhythmicity than during
with large-amplitude delta activity in the low-pass-filtered, mean- isometric tracking, but there were nevertheless periods of pro-
subtracted LFP, comprising either a single cycle or an extended nounced low-frequency oscillation in M1 with a phase that varied
burst of two or more cycles of low-frequency oscillation. systematically across electrodes (Figure 3C). We did not

1188 Neuron 83, 1185–1199, September 3, 2014 ª2014 The Authors

Neuron
Low-Frequency Motor Cortical Dynamics

A Recording during ketamine sedation C Recording during free reaching

EMGs

ECU-EMG ECU
ECR
Unfiltered M1 LFP FCU
FCR
* FDS
1000 µV

FDP
1DI

Low-pass filtered, mean-subtracted M1 LFPs Unfiltered M1 LFP

200 µV
1000 µV

Low-pass filtered, mean-subtracted M1 LFPs

M1 Neurons

200 µV

M1 Neurons
1 sec

B Recording during natural sleep

ECU-EMG

1 sec

Unfiltered M1 LFP D Wrist muscle EMG vs. LFP-PC coherence

0.4 0.4
Coherence with PC1

Coherence with PC2

ECU ECU
1000 µV

ECR ECR
FCU FCU
FCR FCR
0.2 0.2
Low-pass filtered, mean-subtracted M1 LFPs

0 0
1000 µV

0 5 10 0 5 10
Frequency (Hz) Frequency (Hz)

Finger muscle EMG vs. LFP-PC coherence

M1 Neurons 0.4 0.4
Coherence with PC1

Coherence with PC2

FDS FDS
FDP FDP
1DI 1DI
0.2 0.2

0 0
0 5 10 0 5 10
1 sec Frequency (Hz) Frequency (Hz)

Figure 3. Low-Frequency Cortical Dynamics during Ketamine Sedation, Natural Sleep, and Free Reaching
(A) EMG, unfiltered LFP, processed LFP, and spike rasters during ketamine sedation following the session in Figure 1. Arrowheads indicate K complexes,
sometimes associated with spindles (*). LFP traces are ordered and color coded as in Figure 1.
(B) Equivalent recordings during natural sleep at the end of the session in Figure 1.
(C) Equivalent recordings during retrieval of food from small wells in a Klüver board. In addition to the wrist muscles named in Figure 2, EMG was recorded from
flexor digitorum profundus (FDP), flexor digitorum superficialis (FDS), and first dorsal interosseous (1DI), which act on the fingers.
(D) Coherence spectra between M1 LFP-PCs and wrist and finger muscles during free reaching exhibit low-frequency coherence peaks.

Neuron 83, 1185–1199, September 3, 2014 ª2014 The Authors 1189

 Neuron
Low-Frequency Motor Cortical Dynamics

A Submovement K Complex measure kinematics during these complex whole-limb move-

150 600 ments, but characterized behavior instead using EMGs recorded
from multiple hand and wrist muscles. When projected onto
Cursor speed (%/s)

Raw LFP (µV)

the PC plane determined from isometric task recordings, the first
two LFP-PCs were coherent with rectified EMG in the delta band
(Figure 3D), suggesting a consistent relationship between LFP
0 0 and muscle activity even during unrestrained reach-to-grasp.

−1 0 1 −1 0 1 Common Low-Frequency LFP Dynamics across

Time relative to submovement (s) Time relative to K complex (s) Behavioral States
Figure 5A shows power spectra for a representative M1 LFP un-
B 2 3 M1 neuron der the four behavioral states: isometric task performance, free
Normalised firing rate

Normalised firing rate

PMv neuron reaching, natural sleep, and ketamine sedation. It is clear that
2
the spectra vary substantially across conditions. Awake isomet-
1
1 ric and naturalistic reaching behaviors are characterized by a
peak in the beta band around 20 Hz, whereas sleep and sedation
0 0 recordings show increased power at low frequencies. A clear
peak in the delta band is seen during sedation (but not sleep),
−1 0 1 −1 0 1 whereas awake behaviors are associated with a broad distribu-
Time relative to submovement (s) Time relative to K complex (s) tion of power at low frequencies.
The event-triggered analysis used in Figure 4 was not appli-
C 50 500 cable to these naturalistic recordings because the sleep data
lacked K complexes, and the timing of submovements could
not be accurately determined for free movements. Therefore,
M1-LFP (µV)

M1-LFP (µV)

we examined whether a conserved dynamical structure could

0 0
be found in the correlation structure between multichannel
LFPs, and between LFPs and spiking activity, under the four
different behavioral states. In all analyses, we used the same
-50 -500 2D projection of the LFP data, which was determined by PCA
-1 0 1 -1 0 1 of the low-pass-filtered isometric torque data (LFP-PCs).
Time relative to submovement (s) Time relative to K complex (s)
Despite differences in LFP power spectra, the low-frequency
D correlation structure between LFP-PCs was preserved under all
50 500 four behavioral states (Figure 5B). Cross-correlation of both unfil-
tered and low-pass-filtered LFP-PCs revealed strong, consistent
LFP-PC2 (µV)

peaks and troughs separated by about 150 ms, corresponding to

LFP-PC2 (µV)

an oscillatory cycle of around 3 Hz. This is not a trivial conse-

0 0
quence of PCA decomposition, because although PCs must be
uncorrelated at zero lag, there is no reason why they should be
strongly correlated at any other lag. Moreover, it is not trivial
−50 −500 that LFPs recorded under other behavioral states, when
−50 0 50 −500 0 500
LFP-PC1 (µV) LFP-PC1 (µV)

E F Monkey D Monkey R Monkey S 197 K complexes. Data are from monkey D, same sessions as in Figures 1
720 720 and 3.
(B) Average normalized firing rate of seven neurons in M1 (blue) and six neu-
rons in PMv (red) relative to submovement (left) and K complex (right).
KC phase (deg)

KC phase (deg)

(C) Average low-pass-filtered, mean-subtracted LFP from ten M1 electrodes

360
relative to submovement (left) and K complex (right). Traces in both plots are
360
color coded according to phase relative to submovements.
(D) Average submovement-triggered (left) and K complex-triggered (right)
LFP-PC trajectories, plotted over 200 ms on either side of the trigger event
(indicated by circles). All data are projected onto the PC axes determined from
0 0 LFPs recorded during isometric task performance.
0 360 720 0 360 720
SM phase (deg) SM phase (deg) (E) LFP phase relative to submovement (SM phase) plotted against LFP phase
relative to K complex (KC phase) for each M1 electrode (unwrapped over two
full cycles). Dashed lines indicate equality. Points are color coded according to
Figure 4. Submovement- and K Complex-Related Activity Share a LFP phase relative to submovements.
Common Low-Frequency Phase Structure (F) SM phase plotted against KC phase for all LFP recordings over 13 sessions
(A) Left: average cursor speed aligned to a peak speed of 2,063 submove- in three monkeys.
ments. Right: average surface-referenced unfiltered LFP aligned to the peak of Data are presented in Table S2. See also Figure S1.

1190 Neuron 83, 1185–1199, September 3, 2014 ª2014 The Authors

Neuron
Low-Frequency Motor Cortical Dynamics

A Power spectrum of raw LFP D Isometric task vs. ketamine sedation

40 N.S.

N.S.
Isometric task
30 Free reaching 0.5
*** ***
LFP power (dB/Hz)

Coefficient of determination
Sleep
Sedation *** ***
0.25
20

0
10
Task fit Sedation fit
Task-sedation
0-5 Hz
-0.25 generalization
10-30 Hz
0
0 10 20 30 40 50
E Isometric task vs. natural sleep
Frequency (Hz)
N.S.
B Cross-correlation between LFP-PCs N.S.
0.6 1
Unfiltered
Isometric task
Low-pass filtered (<5 Hz) 0.5
* ***

Coefficient of determination
Free reaching
Sleep *** ***
Cross-correlation (R)

Cross-correlation (R)

0.3 Sedation 0.5

0.25

0 0
0
Task fit Sleep fit
-0.3 -0.5 Task-sleep
0-5 Hz
-0.25 generalization
10-30 Hz
-0.6 -1
-1 0 1 -1 0 1 F Isometric task vs. free reaching
Time-lag (s) Time-lag (s) N.S.

C Cross-correlation frequency N.S.

4
Isometric task N.S.
Free reaching
0.5
***
Coefficient of determination

3
Sleep ** ***
Frequency (Hz)

Sedation
0.25
2

1 0
Task fit Task-free Free fit
generalization
0-5 Hz
0
-0.25 10-30 Hz
Monkey D Monkey S

Figure 5. Consistent Low-Frequency LFP Dynamics across Behavioral Conditions

(A) Power spectrum of unfiltered M1 LFPs during isometric task performance, free reaching, natural sleep, and ketamine sedation.
(B) Cross-correlation (normalized R values for each time lag) between unfiltered (left) and low-pass-filtered (right) LFP-PCs under four behavioral states. All data
are projected onto the PC axes determined from low-frequency LFPs recorded during isometric task performance.
(C) Average frequency of LFP-PC correlation (determined from the time interval between cross-correlation peak and trough) for monkeys D and S. In both animals
a small but consistent reduction in frequency is observed during free reaching and ketamine sedation.
(D) Coefficient of determination for a linear dynamical model fit to delta band (black) and beta band (gray). The plot compares the quality of the model fitted on
isometric task data and tested on the same task data (task fit), the model fitted on task data and tested on sedation data (task-sedation generalization), and the
model fitted to sedation data and tested on sedation data (sedation fit). Thirteen session pairs in three animals; data are presented in Table S3.
(E) Equivalent plot showing the model fitted on isometric task data generalizes to natural sleep. Nine session pairs in two animals; data are presented in Table S4.
(F) Equivalent plot showing the model fitted on isometric task data generalizes to free reaching. Ten session pairs in two animals; data are presented in Table S5.
Error bars indicate SD. N.S., not significant, p > 0.05; *p < 0.05; **p < 0.01; ***p < 0.001; paired t test.

projected onto the PC axes determined from the isometric data, lation peak and trough). In both animals for which we recorded
should exhibit the same correlation structure. Indeed, this anal- in all four conditions, the frequency of correlation was highest
ysis revealed a subtle but systematic difference in the frequency (3 Hz) during isometric movements and natural sleep, and
of oscillation (determined from the interval between cross-corre- slightly lower (2.8 Hz) during free reaching and ketamine

Neuron 83, 1185–1199, September 3, 2014 ª2014 The Authors 1191

 Neuron
Low-Frequency Motor Cortical Dynamics

sedation (Figure 5C). Although small, this difference was individ- parameters best fit to the sedated state explained the delta-
ually significant in both animals (one-factor ANOVA; monkey D: band data only marginally better, with a CoD of 0.26 ± 0.13 (Fig-
F3,16 = 11.4, p = 3 3 104; monkey S: F3,10 = 8.0, p = 0.005). ure 5D; Table S3), whereas the best fit to the beta-band activity
Another way to visualize the similarity in correlation structure is remained poor (CoD = 0.02 ± 0.03).
to plot the LFP trajectory over time in the PC plane. Movie S4 Similar results were obtained for the generalization of the iso-
shows real-time LFP data recorded in the different conditions, metric task model to data recorded during natural sleep (Fig-
alongside its PC projection (note that the awake data are ure 5E; Table S4) and free reaching (Figure 5F; Table S5). In
expanded 2-fold to compensate for the increase in slow-wave both cases, the model parameters that best fit the task data
amplitude during sleep and sedation). In all cases the LFP trajec- captured a significantly higher proportion of LFP dynamics in
tory rotated in the same direction, with a frequency of around the delta band compared with the beta band, and the quality
3 Hz. We quantified the extent to which these LFP trajectories of the fit was only marginally improved by fitting model parame-
could be captured by a single first-order linear dynamical equa- ters to the corresponding behavioral state.
tion of the form These analyses confirm the consistent correlation structure
in multichannel M1 LFP activity under all four behavioral states,
_ = A:xðtÞ;
xðtÞ (1)
albeit with a minor (10%) reduction in frequency during
where the time evolution of the first two LFP-PCs, xðtÞ, is deter- free reaching and sedation. Next, we examined whether similar
mined only by a 2 3 2 matrix, A, with a trace equal to zero. Fig- slow LFP oscillations were also observed in ventral premotor
ure S2A shows this procedure applied to LFPs recorded during cortex (PMv), and how they related to neural activity in each
the isometric task. Solutions of Equation 1 are closed elliptical area.
trajectories with constant frequency and direction of rotation
(Figure S2C), similar to the real data (Figure S2D). The three Distinct Low-Frequency LFP Oscillations in M1 and PMv
free parameters of A (the fourth is fixed by the trace constraint) during Task Performance
effectively determine the frequency, orientation, and eccentricity Figure 6 compares M1 and PMv activity during a single trial of
of trajectories. Therefore, the extent to which a single matrix A isometric task performance, and Figures 7A–7C show average
can describe the time evolution of LFP-PC trajectories provides data for an entire session aligned to the end of each successful
a measure of the consistency of the underlying dynamics. trial. Firing rates in M1 (Figures 6B and 7B) were highest during
We quantified the fit over sessions of isometric task perfor- the rising torque phase, as the animal made multiple submove-
mance using the coefficient of determination (Equation 10 in ments to acquire peripheral targets. By contrast, PMv firing rates
Experimental Procedures) and obtained an average (±SD) CoD were highest after the end of the trial, as the animal took a food
of 0.20 ± 0.08 (n = 13 sessions in three animals; Table S3). The reward with the ipsilateral limb. This is consistent with greater
quality of this fit is not a trivial consequence of the orthogonality bilateral tuning of premotor neurons (Hoshi and Tanji, 2006), as
of PCs, because most orthogonal signals cannot be described well as with a strong preference for object-grasping movements
by Equation 1. When the same analysis steps (low-pass filtering, within the bank of the arcuate sulcus (Umilta et al., 2007).
mean referencing, PCA, and model fitting) were applied to These distinct periods of high neuronal activity were each
equivalent lengths of white noise, the 95% percentile of the dis- associated with low-frequency LFP activity within the same
tribution of the resultant CoD was only 0.0013. Moreover, not all cortical area (Figure 6C). In both M1 and PMv, the low-pass-
oscillatory activity can be described by Equation 1. Equivalent filtered LFP could be decomposed into two orthogonal compo-
analysis of beta-band LFP data (filtered between 10 and 30 Hz; nents (Figure 6D). Submovements during the trial were phase
Figure S2B) yielded an average CoD of only 0.06 ± 0.05, signifi- locked to the M1 cycle (Figure 6E), but had no consistent rela-
cantly worse than the low-frequency fit (n = 13, t = 6.8, p = 2 3 tionship to the PMv LFP. The areal velocity of the LFP-PC trajec-
105, paired t test; Figure 5D). This was not due to an absence tory in M1 and PMv was maximal during periods of high neural
of signal at this frequency, because beta-frequency oscillation activity in the same cortical area (Figures 6F and 7C). LFP-PCs
was evident in the raw signal (Figure S2A) and power spectrum within each area exhibited a similar low-frequency correlation
(Figure 5A). Rather, the oscillation at this frequency comprised structure (Figure 6G), indicating a consistent phase lag
predominantly a single phase leading to a high proportion of vari- throughout the recording. However, the oscillations in each
ance in the first PC (Figure S2E), whereas the second PC had no area were largely independent of each other during the isometric
consistent phase relationship. Therefore, trajectories in the PC task, occurring at different phases of the task. As a result, corre-
plane lacked rotational structure (Figure S2D), and hence could lations between LFP-PCs across areas were weaker than within
not be described by first-order linear dynamics. areas (Figure 6H).
Next, we tested how well the model that best described the By contrast, neurons in M1 and PMv were coactive during free
isometric task data could explain LFPs recorded under ketamine reaching-to-grasp with the contralateral limb (Figure S3). In this
sedation. We applied the best-fit parameters obtained from the case, slow oscillations in both areas were phase locked, leading
task recordings to predict the time derivative of the sedation to robust correlation between LFP-PCs across areas. Finally,
data using Equation 1 and achieved a comparable CoD of neural activity in both M1 and PMv under sedation was synchro-
0.20 ± 0.10. This was significantly better than the generalization nized to K complexes (Figures 7D and 7E; Figure S4). Each K
of the beta-band model, which failed to fit these frequencies in complex was also associated with synchronous bursts of low-
the sedation data (mean CoD = 0.05 ± 0.05, n = 13, t = 8.5, frequency oscillation, resulting in peaks of LFP-PC areal velocity
p = 2 3 106, paired t test). For comparison, the model with in both areas (Figure 7F).

1192 Neuron 83, 1185–1199, September 3, 2014 ª2014 The Authors

Neuron
Low-Frequency Motor Cortical Dynamics

A Radial position End hold Reward given

Figure 6. M1 and PMv Are Active during
100
Distinct Phases of Isometric Task Perfor-
%

50
mance
0
(A) Radial cursor position for a representative trial
ECU-EMG of the isometric task. After the peripheral target is
500
acquired (End hold), the monkey takes a food
µV

0
reward with the hand ipsilateral to the recording
−500
1 sec sites. Also shown is raw (black) and rectified,
B M1 neurons
smoothed (red; not to scale) EMG from a wrist
extensor muscle, in which the submovement
structure is more clearly evident.
(B) Spike rasters for eight neurons in M1 (blue) and
six neurons in PMv (red). Note that M1 neurons fire
PMv neurons with contralateral isometric wrist submovements,
whereas PMv neurons are active as the monkey
takes food with the ipsilateral limb.
(C) Low-pass-filtered, mean-subtracted LFPs
C 100 M1 LFPs recorded from ten electrodes in M1 and eight
electrodes in PMv.
µV

0
−100
(D) LFP-PCs calculated from M1 and PMv re-
100 PMv LFPs cordings.
(E) Radial cursor speed and M1 LFP-PC2 overlaid.
µV

0
−100 Arrowheads indicate identified submovements
PC1
D 200 PC2
with a peak speed exceeding 30%/s. Note that
M1 LFP-PCs
submovements are phase locked to the M1 oscil-
µV

0
lation, although in this trial they do not occur on
-200 PC1
200 PMv LFP-PCs PC2 every cycle.
µV

0 (F) Areal velocity in the PC plane for M1 and PMv

-200 LFPs. Increased areal velocity in M1 coincides with
M1 neural activity, whereas increased areal ve-
E Radial speed
Speed locity in PMv coincides with PMv neural activity.
500 M1 PC2
(G) Cross-correlation between LFP-PCs within the
%/s

0
same cortical area.
-500
(H) Cross-correlation between LFP-PCs across
F M1 cortical areas.
0.2 Areal velocity PMv Data are from monkey D. See also Figures S3
mV2/s

0.1
and S4.
0

1 sec
G 1 M1 PC1 x M1 PC2 H 1 M1 PC1 x PMv PC2
PMv PC1 x PMv PC2 PMv PC1 x M1 PC2
tories within and across areas, consistent
0.5 0.5 with the synchronization of low-frequency
Cross-correlation (R)

Cross-correlation (R)

rhythms under these conditions.

0 0 To assess phase locking to M1 LFP cy-
cles across different data sets, we rotated
-0.5 -0.5 the M1 LFP-PC plane such that the peak
speed during isometric submovements
-1 -1 occurred at a phase of zero. Across the
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
population, M1 neurons were significantly
Time-lag (s) Time-lag (s)
phase locked to M1 LFP oscillations in all
three animals (Figure 8B), with an average
Neuronal Firing in M1 Is Phase Locked to Slow preferred phase that preceded peak speed, consistent with
Oscillations during Movement, Sleep, and Sedation these neurons having a causal role in movement. During task
Last, we examined how spiking activity was related to the phase performance, PMv neurons in two out of the three animals (mon-
of the low-frequency oscillation in each area. Figure 8A and keys D and R) did not show consistent locking to M1-LFP cycles,
Movie S5 show sample spike-triggered average trajectories of whereas in the third animal (monkey S) the distribution of
M1 and PMv LFP-PCs for the same set of neurons recorded preferred phases was significantly nonuniform but nevertheless
during the four behavioral states. During isometric task perfor- broad relative to M1 neurons. However, during free reaching,
mance, neurons showed greater locking to LFPs within the natural sleep and ketamine sedation, neurons in both M1 and
same cortical area, as expected from the dissociation of activity PMv became synchronized and fired at a similar preferred phase
patterns during different task phases (Figures 7B and 7C). How- of the M1 cycle in all three animals.
ever, during free reaching, sleep, and sedation, spike activity in In summary, the relationship between spiking activity and
both M1 and PMv was associated with cyclical LFP-PC trajec- LFPs suggests that each cortical area is governed by its own

Neuron 83, 1185–1199, September 3, 2014 ª2014 The Authors 1193

 Neuron
Low-Frequency Motor Cortical Dynamics

A D Figure 7. M1 and PMv Activity during Iso-

End hold K Complex metric Task Performance and Ketamine
60 500
Sedation

Raw LFP (µV)

Radial cursor
position (%)

40 (A) Radial cursor position aligned to the end of suc-

0 cessful hold periods for peripheral targets, averaged
20 across 40 trials from the session shown in Figure 6.
(B) Average normalized (to zero mean and unity
0 -500
B E standard deviation) firing rate for eight M1 neurons
and six PMv neurons, aligned to the end of the hold
Normalised firing rate

Normalised firing rate

2 M1 neurons 2 M1 neurons
PMv neurons PMv neurons period. M1 activity is highest as the monkey gener-
1 1 ates torque with the contralateral wrist to reach tar-
gets. PMv activity is highest after the successful trial,
0 0
corresponding to taking food reward with the ipsi-
−1 −1 lateral limb.
(C) Average areal velocity in the PC plane of M1 and
C F PMv LFPs, aligned to the end of the hold period. The
M1 LFP M1 LFP
Areal velocity (mV2/s)

Areal velocity (mV2/s)

0.05 1 profile of areal velocity during task performance
PMv LFP (x2) PMv LFP (x2)
mirrors the dissociation seen in neural activity across
0.025 0.5
areas. Note that the vertical scale for PMv areal
0 0 velocity is expanded 32 for ease of comparison.
(D) Average unfiltered, surface-referenced M1 LFP
-0.025 1s -0.5 1s aligned to the peak of 48 K complexes during keta-
mine sedation.
(E) Average normalized firing rate for the same M1 and PMv neurons, aligned to K complexes. Neural activity in both M1 and PMv is maximal during the rising phase and
peak of the K complexes.
(F) Average areal velocity in the PC plane of M1 and PMv LFPs, aligned to K complexes.
Shading indicates SEM across trials or K complexes.

intrinsic dynamics, allowing distinct slow oscillations to emerge became coupled across different areas, which may explain the
in M1 and PMv when those areas are individually active during increased slow-wave amplitude seen in these brain states.
different phases of the isometric task. However, during free
behavior involving coactivation of M1 and PMv, as well as during Functional Role of Slow Oscillations
sleep and sedation, the slow oscillations become coupled It has previously been thought that the frequency of submove-
across cortical areas. ments during visuomotor tracking was determined by feedback
loop delays, because their rhythmicity is disrupted under condi-
DISCUSSION tions of absent or delayed visual feedback (Miall et al., 1986,
1993; McAuley et al., 1999; Miall and Jackson, 2006). However,
A Common Structure Underlies Low-Frequency Motor our finding of a common oscillatory structure in the cortical LFP
Cortex Activity during Movement and Sedation that is (1) coherent with movement speed and (2) present during
We have described a common 3 Hz correlation structure in LFP sleep and sedation reveals an intrinsic periodicity in motor cir-
recordings during an isometric movement task, free reaching, cuitry at the submovement frequency. Submovement durations
natural sleep, and ketamine sedation. Individual LFPs exhibited are relatively unaffected by movement speed (Miall et al., 1986;
oscillatory activity, albeit of lower amplitude in the awake Roitman et al., 2004; Pasalar et al., 2005; Selen et al., 2006),
state, with a consistent distribution of phase across electrodes target size (Selen et al., 2006), arm stiffness (Selen et al.,
relative to submovements and K complexes. Because this 2006), or learning novel visuomotor mappings (Sailer et al.,
phase distribution was preserved across behavioral states, the 2005). One possibility is that the intrinsic dynamics are tuned
multielectrode LFP could be decomposed into two orthogonal appropriately for visuomotor control such that different phases
components that evolved according to the same underlying dy- of the cycle are associated with the various computations
namics during all behavioral conditions. During isometric task involved in planning and generating the next submovement
performance, M1 neurons fired at a consistent phase of the based on feedback from the previous one. In fact, adaptation
cortical cycle, and this modulation of descending drive led to a to delayed visual feedback is extremely limited (Miall and Jack-
3 Hz submovement structure in muscle activity and movement son, 2006), suggesting that the motor system may in fact be
kinematics. A similar cycle was evident during free reaching tuned only to a narrow range of naturally occurring loop delays.
movements, which also comprise submovements (Milner and Indeed, Kunesch et al. (1989) concluded that the temporal char-
Ijaz, 1990; Roitman et al., 2004), although peripheral interactions acteristics of manipulative hand movements requiring tactile
with limb biomechanics and afferent feedback may lower the feedback were determined not by (shorter) sensorimotor loop
frequency and obscure the clear rhythmicity seen in isometric delays but instead by central neural mechanisms responsible
tasks. Interestingly, the isometric task also revealed a dissocia- for interpreting sensory inputs. This would be consistent with a
tion of neural activity within M1 and PMv during different task common intrinsic oscillator shaping the structure of feedback-
phases, each associated with a distinct low-frequency oscilla- controlled movements, irrespective of the feedback modality.
tion. However, during sleep and under sedation, these rhythms Finally, it is interesting to note that during verbal articulation there

1194 Neuron 83, 1185–1199, September 3, 2014 ª2014 The Authors

Neuron
Low-Frequency Motor Cortical Dynamics

is coherence between cortical signals and mouth EMG at a fre- source of GABAB-mediated inhibition (Carracedo et al., 2013).
quency of 2–3 Hz, which reflects the spontaneous rhythmicity The slow kinetics of this G protein-coupled receptor lead to
of speech (Ruspantini et al., 2012). sustained hyperpolarizing currents that can be delayed by
several hundred milliseconds relative to inhibitory cell activity.
The Origin of Low-Frequency Cortical Dynamics These slow currents are observed in the LFP (Dine et al.,
Care must be taken when inferring neural substrates of LFP 2014), and might be expected to contribute a low-frequency
activity, because synaptic and intrinsic currents from multiple component with a substantial phase lag relative to ionotropic
neuronal populations contribute to the extracellular field (Buzsáki currents. Nevertheless, occasional bursting has been reported
et al., 2012). Moreover, rotation in the PC plane does not require in the thalamus in the awake state (Guido and Weyand, 1995),
underlying oscillatory sources that are orthogonal, because any and the relative contributions of cortical and corticothalamic
consistent phase difference, or a traveling wave appearing with a mechanisms in generating delta activity in vivo during behavior
different phase on each electrode, could equally be decom- and sleep remain an important area for further investigation.
posed into orthogonal components (Rubino et al., 2006; Murphy
et al., 2009; Nauhaus et al., 2009; Ray and Maunsell, 2011). Kinematic Information in LFP Trajectories
Importantly, the distribution of preferred phase for neural firing Low-frequency LFPs have several practical advantages for BMIs
was narrow compared to the LFP (Figure S5A). Moreover, during (Rickert et al., 2005; Bansal et al., 2011; Hwang and Andersen,
sleep and sedation, this phase was common to neurons in both 2013), but our understanding of how these signals arise and
M1 and PMv (Figure 8B). This appears incompatible with a trav- how best to extract information from them is limited. We found
eling wave, which would cause neurons at different locations to that the areal velocity swept out by LFP trajectories was propor-
fire at different preferred phases of the global cycle. Churchland tional to movement speed, and suggest that this may prove a
et al. (2012) reported complex, multiphasic patterns of cortical useful feature to examine for BMI applications, as it is robust
activity that could be projected onto a plane using the jPCA to sources of synchronous noise (because correlated signals
method to reveal consistent cycles with notable similarity to lead to radial trajectories). In 3D PC space, there was a slight
the LFP trajectories we describe here. However, it is not clear variation in the axis of rotation for different directions of move-
from that study whether all phases of the cycle were represented ment. In effect, the first two PCs captured the LFP trajectory
equally, because the jPCA method is again based on orthogonal that was common across all submovements, whereas the third
projections of the neural activity. Consistent with our observa- component reflected more subtle variation in the neuronal sour-
tions, Riehle et al. (2013) found that movement-related potentials ces associated with different directions (Waldert et al., 2009).
were composed of multiple components with amplitude and la- These observations suggest that understanding the lawful
tency that varied systematically across the cortical surface, even dynamics that generate low-frequency behaviors may inform
though recorded neurons tended to be maximally active around and constrain the search for more sophisticated approaches to
movement onset. decoding kinematics from LFPs.
A parsimonious explanation of the consistent correlation
structure we describe is that the multichannel LFP comprises a Conclusions
mixture of at least two underlying sources with a fixed time/ By examining the dynamics of motor cortex activity, we can unite
phase delay (Figures S5B and S5C). If one source reflects (rela- two previously distinct phenomena: the rhythmicity of submove-
tively) synchronous neural activity occurring around submove- ments during isometric tracking and delta oscillations during
ments and K complexes, what then is the source of the second sleep and under sedation. In both cases, cortical neurons fire
component? One possibility is that neural activity at other at distinct phases of the same underlying 3 Hz LFP cycle, and
phases is undersampled in our recordings, either because the thereby impose this frequency on behavior via modulation of
neurons are located in a different area of cortex or a subcortical the descending drive to muscles. We suggest that this intrinsic
structure or have smaller soma size (for example, inhibitory inter- rhythmicity reflects an underlying organization of motor cortical
neurons). An alternative explanation is that the field potential circuits engaged in feedback control of movement.
associated with synchronous neural activity may be composed
of multiple sources with different time courses. These sources EXPERIMENTAL PROCEDURES
are cortical, because submovement- and K complex-related
LFP oscillations underwent polarity reversal within the gray Isometric Movement Task
Experiments were approved by the local ethics committee and performed un-
matter, and we speculate that they may reflect excitatory and
der appropriate UK Home Office licenses in accordance with the Animals (Sci-
inhibitory synaptic potentials contributing to the generation of entific Procedures) Act 1986. Three purpose-bred female rhesus macaques
low-frequency rhythms. Delta oscillations can arise in the thal- (monkey R: 5 years old, 5 kg; monkey D: 6 years old, 6.5 kg; monkey S: 5 years
amus due to low-threshold calcium currents active in the hyper- old, 5.4 kg) were trained to control a cursor by generating isometric flexion-
polarized state (Amzica et al., 1992; Destexhe and Sejnowski, extension (vertical) and radial-ulnar (horizontal) torque with the left wrist
2003). However, in the awake state, thalamic neurons are depo- restrained in pronated posture to move to eight peripheral targets presented
in a pseudorandomized center-out sequence on a computer monitor. Wrist
larized and generally fire in a tonic mode (Steriade and Llinás,
torque was measured using a six-axis force/torque transducer (Nano25; ATI
1988), suggesting that the low-frequency dynamics we observe
Industrial Automation). Cursor position was expressed as percentage of the
during behavior may relate to a cortical delta rhythm that has distance to screen edge, with 100% corresponding to a torque of 0.67 Nm.
recently been characterized in slice preparations. This rhythm Targets were centered at 70% of the distance to screen edge and had a diam-
originates from intrinsic bursting cells in layer V that activate a eter of 25%.

Neuron 83, 1185–1199, September 3, 2014 ª2014 The Authors 1195

 Neuron
Low-Frequency Motor Cortical Dynamics

A Isometric Free Sleep Sedation

task reaching
M1 neurons
PMv neurons
M1 LFP

3 µV 6 µV 6 µV 12 µV
PMv LFP

2 µV 4 µV 4 µV 8 µV

B Phase of M1 LFP (relative to peak speed)

Isometric Free Sleep Sedation

task reaching
Peak speed
M1 neurons

P=0.027 P<0.001
P<0.001 P<0.001 P<0.001 P<0.001
P<0.001 P=0.024 P<0.001 P<0.001
PMv neurons

P=0.45 P<0.001
P=0.64 P<0.001 P<0.001 P<0.001
P<0.001 P<0.001 P<0.001 P<0.001

Figure 8. Phase Locking of Neural Activity to the Cortical Cycle during Isometric Movement, Free Reaching, Natural Sleep, and Ketamine
Sedation
(A) Spike-triggered average LFP-PC trajectories for eight M1 neurons (blue) and six PMv neurons (red) over 200 ms before and after spike time
(indicated by circles). Top row: averages of M1 LFPs; bottom row: averages of PMv LFPs. In all cases, the data are projected onto PC axes

(legend continued on next page)

1196 Neuron 83, 1185–1199, September 3, 2014 ª2014 The Authors

Neuron
Low-Frequency Motor Cortical Dynamics

Surgical Procedures Areal Velocity of LFP-PC Trajectories

After training, we implanted EMG electrodes onto forearm and hand muscles, Submovements were binned into nine groups of equal size according to
tunneled subcutaneously to a connector on the head. In a separate surgery, increasing peak cursor speed, or alternatively into six groups according
two custom arrays of 12 moveable 50 mm diameter tungsten microwires to submovement direction. 2D LFP-PC trajectories from 200 ms before to
(impedance 200 kU at 1 kHz) were implanted into the right M1 and PMv 200 ms after the midpoint of each submovement were averaged within
(Jackson and Fetz, 2007). All surgeries were performed under sevoflurane each group and quantified using areal velocity ðnareal Þ and frequency of rota-
anesthesia with postoperative analgesics and antibiotics. tion ðfÞ:
1
_
vareal ðtÞ = xðtÞ 3 xðtÞ (2)
Electrophysiological Recording 2
Head-free recordings were made using unity-gain headstages followed by
wide-band amplification and sampling at 24.4 ksp/s (sp, sample) (System 3; 1 2
jv areal ðtÞj = jxðtÞj :2p:fðtÞ: (3)
Tucker-Davis Technologies). LFPs were digitally low-pass filtered at 300 Hz 2
and recorded at 488 sp/s. EMGs were amplified (31,000) and band-pass _ is its derivative with
Here, xðtÞ is a 2D or 3D vector of LFP-PCs at time t, xðtÞ
filtered between 10 and 5,000 Hz (model 1700; AM Systems) before sampling respect to time, and 3 denotes the vector cross-product. For each submove-
at 12.2 ksp/s. ment group, rotation frequency and areal velocity were measured at the time of
peak cursor speed.
Data Set
Kinematic decoding was performed on 13 sessions (monkey R: 4; monkey Areal Velocity Decoding of Single-Submovement Kinematics
D: 6; monkey S: 3). Sedation data (at least 5 min per session) were recorded We used the average 3D areal velocity vector (Equation 2) from 200 ms
on separate days after induction with ketamine (10 mg/kg, intramuscularly; before to 200 ms after the time of peak speed to decode the kinematics of
i.m.) and medetomidine (0.02 mg/kg, i.m.). We report 13 pairs of movement individual submovements with leave-one-out cross-validation, as follows.
and sedation sessions (separated by no more than 3 days) for all animals
(monkey D: 5; monkey R: 4; monkey S: 4). In monkeys D and S, we (1) We parameterized the relationship between areal velocity v i (for sub-
collected natural sleep data at the end of behavioral sessions, and report movement i) and the speed and direction of that submovement
four and five pairs of sessions, respectively. In monkeys D and S, we also ðsi ; qi Þ assuming (1) for a given direction of submovement, the areal ve-
collected data while animals retrieved food from a Klüver board, and report locity magnitude increased linearly with cursor speed, and (2) for a
six and four such sessions. On average, 15 neurons were recorded per given speed of submovement, the orientation of areal velocity vector
session. varied with submovement direction. Specifically,

vi = si :bðqi Þ; (4)
Data Preprocessing
Offline analyses were performed in MATLAB (MathWorks). LFPs were visu- with the direction-dependent component composed of a Fourier
ally inspected and electrodes with excessive mains noise or artifacts were series,
discarded. Remaining LFPs recorded during isometric task performance bðqi Þ = b0 + b1 cosðqi Þ + b2 sinðqi Þ + b3 cosð2qi Þ + b4 sinð2qi Þ: (5)
were separated by area (M1 and PMv) and processed by low-pass filtering
(5 Hz, four-pole, zero-phase digital Butterworth filter), mean referencing (i.e., We included only terms up to 2qi to prevent overfitting. The 15 free pa-
subtraction of the mean LFP across electrodes within the same cortical rameters were obtained by linear regression over the entire data set,
area), and dimensionality reduction using standard PCA. The PC plane excluding one submovement that was used for cross-validation.
was oriented such that the predominant rotational structure during task (2) We assessed how well the model predicted the speed (or direction) of
performance was in the anticlockwise direction. In all analyses, LFPs re- the excluded submovement from the magnitude (or orientation) of the
corded during other behavioral states (free reaching, sleep, and sedation) areal velocity vector associated with that submovement. The decoded
were always projected into the same PC space obtained from the corre- submovement speed was proportional to the magnitude of the areal
sponding isometric task data set. We refer to these projections throughout velocity vector,
as LFP-PCs.  
v j 
Cursor speed was calculated as the derivative of the magnitude of the 2D Decoded speedð sbj Þ = ; (6)
jb0 j
torque vector, that is, the radial component of velocity with a positive sign
for movements away from the center of the screen. Submovements were whereas the decoded submovement direction was that which mini-
defined by a peak speed exceeding 30%/s. K complexes were identified mized the angular deviation between actual and predicted areal veloc-
from a single surface-referenced LFP channel as the peak of a positive deflec- ity vectors, calculated by the vector dot product,
tion that exceeded 250 mV.    
v j :bðqÞ
Coherence spectra were calculated between unfiltered cursor position and Decoded direction b q j = arg maxq : (7)
jbðqÞj
speed; rectified EMG and LFP-PCs used a 2,048-point rectangular window
with no overlap. Although the PC axes were determined from low-pass-filtered Steps 1 and 2 were repeated with a different submovement excluded
data, we used unfiltered LFP projections for coherence spectra so as to until the speed and direction of all N submovements had been
include frequencies above the filter cutoff. estimated.
Online, semisupervised spike classification used principal component (3) We quantified decoding performance using coefficients of determina-
feature extraction and K means clustering (SpikePac; Tucker-Davis Tech- tion,
nologies). Firing-rate profiles for each neuron were calculated offline by
binning spike events (into 488 bins/s), low-pass filtering at 5 Hz, and P 2
j ðsj  sbj Þ
normalizing to zero mean and unity standard deviation across the entire CoDðspeedÞ = 1  P 2
(8)
recording. j sj

determined from the isometric task recordings. Nevertheless, a consistent rotational structure is observed across all behavioral states. See also
Movie S5.
(B) Summary of the preferred phase of neural firing within the M1 LFP cycle, relative to the LFP phase at peak movement speed. Data are for 98/125 (isometric
task), 71/83 (free reaching), 61/78 (sleep), and 89/122 (sedation) neurons from M1/PMv, respectively. p values indicate the significance of the Rayleigh test of
circular nonuniformity. Monkey R, open circles; monkey D, filled circles; monkey S, crosses. See also Figure S5.

Neuron 83, 1185–1199, September 3, 2014 ª2014 The Authors 1197

 Neuron
Low-Frequency Motor Cortical Dynamics

1X   Amzica, F., Nuñez, A., and Steriade, M. (1992). Delta frequency (1–4 Hz) oscil-
CoDðdirectionÞ = cos qj  b
qj : (9) lations of perigeniculate thalamic neurons and their modulation by light.
N j
Neuroscience 51, 285–294.
Antelis, J.M., Montesano, L., Ramos-Murguialday, A., Birbaumer, N., and
(4) Finally, we determined significance thresholds (p < 0.05) for CoD values Minguez, J. (2013). On the usage of linear regression models to reconstruct
by repeating the entire procedure for 1,000 surrogate data sets in which limb kinematics from low frequency EEG signals. PLoS ONE 8, e61976.
either the speed or direction was shuffled across submovements.
Bansal, A.K., Vargas-Irwin, C.E., Truccolo, W., and Donoghue, J.P. (2011).
Phase of LFP Relative to Submovements and K Complexes Relationships among low-frequency local field potentials, spiking activity,
The phase of each LFP relative to submovements and K complexes was deter- and three-dimensional reach and grasp kinematics in primary motor and
mined at the time of the event from a Hilbert transform of the event-triggered ventral premotor cortices. J. Neurophysiol. 105, 1603–1619.
average. The correlation between LFP phase relative to each event was tested Berens, P. (2009). CircStat: a MATLAB toolbox for circular statistics. J. Stat.
using the circular-circular correlation coefficient available in the CircStat Softw. 31, 1–20.
toolbox (Berens, 2009).
Bradberry, T.J., Gentili, R.J., and Contreras-Vidal, J.L. (2010). Reconstructing
three-dimensional hand movements from noninvasive electroencephalo-
2D LFP-PC Trajectory Model graphic signals. J. Neurosci. 30, 3432–3437.
The simplest linear system with oscillatory dynamics is a 2D state vector xðtÞ
that evolves according to Equation 1. Under the conditions trace ðAÞ = 0 and Buzsáki, G., Anastassiou, C.A., and Koch, C. (2012). The origin of extracellular
det ðAÞ > 0, this system exhibits stable periodic solutions. We regressed the fields and currents—EEG, ECoG, LFP and spikes. Nat. Rev. Neurosci. 13,
_ against 2D LFP-PCs, x, to find the three free parameters of
time derivative, x, 407–420.
A (the fourth is fixed by the trace constraint). Because the LFP-PCs during Carracedo, L.M., Kjeldsen, H., Cunnington, L., Jenkins, A., Schofield, I.,
the isometric task are orthogonal, we expect the elements of matrix A to be Cunningham, M.O., Davies, C.H., Traub, R.D., and Whittington, M.A. (2013).
zero on the diagonal. However, we did not impose this constraint during model A neocortical delta rhythm facilitates reciprocal interlaminar interactions via
fitting, because it will not necessarily hold when data from other conditions are nested theta rhythms. J. Neurosci. 33, 10750–10761.
projected onto the same PC axes. We measured the quality of fit to the LFP-PC Cash, S.S., Halgren, E., Dehghani, N., Rossetti, A.O., Thesen, T., Wang, C.,
d
_ in each case with a vector coefficient of determination defined by
derivative xðtÞ Devinsky, O., Kuzniecky, R., Doyle, W., Madsen, J.R., et al. (2009). The human
R  d 2 K-complex represents an isolated cortical down-state. Science 324,
_  xðtÞ
 xðtÞ _  dt 1084–1087.
CoD = 1  R ; (10)
_ 2 dt
jxðtÞj Churchland, M.M., Cunningham, J.P., Kaufman, M.T., Foster, J.D.,
Nuyujukian, P., Ryu, S.I., and Shenoy, K.V. (2012). Neural population dynamics
where integration was performed over the entire recording. The same analysis
during reaching. Nature 487, 51–56.
was applied separately to LFP data that had been filtered into delta- (0–5 Hz)
and beta- (10–30 Hz) frequency bands using four-pole, zero-phase digital But- Colrain, I.M. (2005). The K-complex: a 7-decade history. Sleep 28, 255–273.
terworth filters before mean referencing and PCA. Craik, K.J. (1947). Theory of the human operator in control systems; the oper-
ator as an engineering system. Br. J. Psychol. Gen. Sect. 38, 56–61.
Spike-Triggered Average LFP-PC Trajectories
Destexhe, A., and Sejnowski, T.J. (2003). Interactions between membrane
M1 and PMv LFPs were averaged from 200 ms before to 200 ms after each spike.
conductances underlying thalamocortical slow-wave oscillations. Physiol.
The LFP averages were then projected onto the PC axes determined from the iso-
Rev. 83, 1401–1453.
metric task data and normalized by the standard deviation of the firing rate. To
allow comparison across different data sets, the M1 LFP-PC plane was rotated Dine, J., Kühne, C., Deussing, J.M., and Eder, M. (2014). Optogenetic evoca-
such that the average submovement-triggered trajectory had zero phase at the tion of field inhibitory postsynaptic potentials in hippocampal slices: a simple
moment of peak cursor speed, and the phase of the spike firing was measured and reliable approach for studying pharmacological effects on GABAA and
relative to this. Significant phase locking across the population of neurons was GABAB receptor-mediated neurotransmission. Front. Cell. Neurosci. 8, 2.
assessed using the Rayleigh test for circular uniformity in the CircStat toolbox. Dipietro, L., Poizner, H., and Krebs, H.I. (2011). EEG correlates of submove-
ments. Conf. Proc. IEEE Eng. Med. Biol. Soc. 2011, 7429–7432.
SUPPLEMENTAL INFORMATION Doeringer, J.A., and Hogan, N. (1998). Intermittency in preplanned elbow
movements persists in the absence of visual feedback. J. Neurophysiol. 80,
Supplemental Information includes five figures, five tables, and five movies 1787–1799.
and can be found with this article online at http://dx.doi.org/10.1016/j.
Grillner, S. (2006). Biological pattern generation: the cellular and computational
neuron.2014.07.022.
logic of networks in motion. Neuron 52, 751–766.

ACKNOWLEDGMENTS Guido, W., and Weyand, T. (1995). Burst responses in thalamic relay cells of
the awake behaving cat. J. Neurophysiol. 74, 1782–1786.
We thank J. Tulip and K. Nazarpour for assistance and S.N. Baker for discus- Hoshi, E., and Tanji, J. (2006). Differential involvement of neurons in the dorsal
sion. This work was supported by the Wellcome Trust (086561), Medical and ventral premotor cortex during processing of visual signals for action plan-
Research Council (MR/G0802195/1, MR/K501396/1), and Engineering and ning. J. Neurophysiol. 95, 3596–3616.
Physical Sciences Research Council (EP/H051570/1). Hwang, E.J., and Andersen, R.A. (2013). The utility of multichannel local field
potentials for brain-machine interfaces. J. Neural Eng. 10, 046005.
Accepted: July 14, 2014
Jackson, A., and Fetz, E.E. (2007). Compact movable microwire array for long-
Published: August 14, 2014
term chronic unit recording in cerebral cortex of primates. J. Neurophysiol. 98,
3109–3118.
REFERENCES
Jerbi, K., Lachaux, J.P., N’Diaye, K., Pantazis, D., Leahy, R.M., Garnero, L.,
Amzica, F., and Steriade, M. (1997). The K-complex: its slow (<1-Hz) rhyth- and Baillet, S. (2007). Coherent neural representation of hand speed in humans
micity and relation to delta waves. Neurology 49, 952–959. revealed by MEG imaging. Proc. Natl. Acad. Sci. USA 104, 7676–7681.
Amzica, F., and Steriade, M. (1998). Electrophysiological correlates of sleep Kiehn, O. (2006). Locomotor circuits in the mammalian spinal cord. Annu. Rev.
delta waves. Electroencephalogr. Clin. Neurophysiol. 107, 69–83. Neurosci. 29, 279–306.

1198 Neuron 83, 1185–1199, September 3, 2014 ª2014 The Authors

Neuron
Low-Frequency Motor Cortical Dynamics

Kunesch, E., Binkofski, F., and Freund, H.J. (1989). Invariant temporal charac- Riehle, A., Wirtssohn, S., Grün, S., and Brochier, T. (2013). Mapping the spatio-
teristics of manipulative hand movements. Exp. Brain Res. 78, 539–546. temporal structure of motor cortical LFP and spiking activities during reach-to-
grasp movements. Front. Neural Circuits 7, 48.
Massey, J.T., Lurito, J.T., Pellizzer, G., and Georgopoulos, A.P. (1992). Three-
dimensional drawings in isometric conditions: relation between geometry and Roitman, A.V., Massaquoi, S.G., Takahashi, K., and Ebner, T.J. (2004).
kinematics. Exp. Brain Res. 88, 685–690. Kinematic analysis of manual tracking in monkeys: characterization of move-
ment intermittencies during a circular tracking task. J. Neurophysiol. 91,
McAuley, J.H., Farmer, S.F., Rothwell, J.C., and Marsden, C.D. (1999).
901–911.
Common 3 and 10 Hz oscillations modulate human eye and finger movements
while they simultaneously track a visual target. J. Physiol. 515, 905–917. Rubino, D., Robbins, K.A., and Hatsopoulos, N.G. (2006). Propagating waves
mediate information transfer in the motor cortex. Nat. Neurosci. 9, 1549–1557.
Miall, R.C., and Jackson, J.K. (2006). Adaptation to visual feedback delays in
Ruspantini, I., Saarinen, T., Belardinelli, P., Jalava, A., Parviainen, T., Kujala, J.,
manual tracking: evidence against the Smith Predictor model of human visu-
and Salmelin, R. (2012). Corticomuscular coherence is tuned to the sponta-
ally guided action. Exp. Brain Res. 172, 77–84.
neous rhythmicity of speech at 2–3 Hz. J. Neurosci. 32, 3786–3790.
Miall, R.C., Weir, D.J., and Stein, J.F. (1986). Manual tracking of visual targets Sailer, U., Flanagan, J.R., and Johansson, R.S. (2005). Eye-hand coordination
by trained monkeys. Behav. Brain Res. 20, 185–201. during learning of a novel visuomotor task. J. Neurosci. 25, 8833–8842.
Miall, R.C., Weir, D.J., and Stein, J.F. (1993). Intermittency in human manual Schöner, G., and Kelso, J.A. (1988). Dynamic pattern generation in behavioral
tracking tasks. J. Mot. Behav. 25, 53–63. and neural systems. Science 239, 1513–1520.
Milner, T.E., and Ijaz, M.M. (1990). The effect of accuracy constraints on three- Selen, L.P., van Dieën, J.H., and Beek, P.J. (2006). Impedance modulation and
dimensional movement kinematics. Neuroscience 35, 365–374. feedback corrections in tracking targets of variable size and frequency.
J. Neurophysiol. 96, 2750–2759.
Murphy, M., Riedner, B.A., Huber, R., Massimini, M., Ferrarelli, F., and Tononi,
G. (2009). Source modeling sleep slow waves. Proc. Natl. Acad. Sci. USA 106, Shenoy, K.V., Sahani, M., and Churchland, M.M. (2013). Cortical control of arm
1608–1613. movements: a dynamical systems perspective. Annu. Rev. Neurosci. 36,
337–359.
Nauhaus, I., Busse, L., Carandini, M., and Ringach, D.L. (2009). Stimulus
Steriade, M., and Llinás, R.R. (1988). The functional states of the thalamus and
contrast modulates functional connectivity in visual cortex. Nat. Neurosci.
the associated neuronal interplay. Physiol. Rev. 68, 649–742.
12, 70–76.
Umilta, M.A., Brochier, T., Spinks, R.L., and Lemon, R.N. (2007). Simultaneous
Pasalar, S., Roitman, A.V., and Ebner, T.J. (2005). Effects of speeds and force recording of macaque premotor and primary motor cortex neuronal popula-
fields on submovements during circular manual tracking in humans. Exp. Brain tions reveals different functional contributions to visuomotor grasp.
Res. 163, 214–225. J. Neurophysiol. 98, 488–501.
Ray, S., and Maunsell, J.H. (2011). Network rhythms influence the relationship Waldert, S., Preissl, H., Demandt, E., Braun, C., Birbaumer, N., Aertsen, A.,
between spike-triggered local field potential and functional connectivity. and Mehring, C. (2008). Hand movement direction decoded from MEG and
J. Neurosci. 31, 12674–12682. EEG. J. Neurosci. 28, 1000–1008.
Rickert, J., Oliveira, S.C., Vaadia, E., Aertsen, A., Rotter, S., and Mehring, C. Waldert, S., Pistohl, T., Braun, C., Ball, T., Aertsen, A., and Mehring, C. (2009).
(2005). Encoding of movement direction in different frequency ranges of motor A review on directional information in neural signals for brain-machine inter-
cortical local field potentials. J. Neurosci. 25, 8815–8824. faces. J. Physiol. Paris 103, 244–254.

Neuron 83, 1185–1199, September 3, 2014 ª2014 The Authors 1199