a r t ic l e s

Flexible information routing by transient synchrony

Agostina Palmigiano1–4, Theo Geisel1–3, Fred Wolf1–4,6 & Demian Battaglia2,5,6   
Perception, cognition and behavior rely on flexible communication between microcircuits in distinct cortical regions.
The mechanisms underlying rapid information rerouting between such microcircuits are still unknown. It has been proposed
that changing patterns of coherence between local gamma rhythms support flexible information rerouting. The stochastic and
transient nature of gamma oscillations in vivo, however, is hard to reconcile with such a function. Here we show that models
© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.

of cortical circuits near the onset of oscillatory synchrony selectively route input signals despite the short duration of gamma
bursts and the irregularity of neuronal firing. In canonical multiarea circuits, we find that gamma bursts spontaneously arise
with matched timing and frequency and that they organize information flow by large-scale routing states. Specific self-organized
routing states can be induced by minor modulations of background activity.

Perception and cognition rely on context-dependent selection of rel-
evant inputs and flexible interareal brain communication. However,
despite its fundamental role, the basis of dynamic information routing
remains an unsolved problem. In particular, the circuit mechanisms
that underlie the rapid reconfiguration of selective information trans-
fer on behavioral time-scales are poorly understood. Proposals range
from hypothetical circuitry dedicated to routing1–3 to conditional sig-
nal propagation4–6 to the hypothesis that neuronal oscillations direct
interareal communication, either by enabling conditional readout
through frequency filtering7 or by exploiting coherence in networks of
regular oscillations8. Windows for efficient communication along dif-
ferent pathways might potentially also arise from fluctuations of local
network excitability. In particular, this latter ‘communication through
coherence’ hypothesis has been invoked in the interpretation of many
experimental findings in which boosted interareal coherence seems
associated with enhanced interareal communication9. For instance,
increased interareal gamma-band coherence and phase synchroniza-
tion might be markers of selective attention, in which different input
information streams have to be transmitted or attenuated according
to task relevance10,11.
Conceptually, however, there are fundamental obstacles to a func-
tion of neuronal oscillations in routing. First, oscillatory synchroni-
zation is confined to short episodes lasting on the order of ~100 ms,
and the statistical properties of these episodes bear strong signatures
of stochasticity12,13. Second, neuronal spike firing is irregular and
only weakly synchronized, with spikes emitted at every possible phase
within the ongoing oscillation cycle14. Third, interareal synaptic
transmission delays are long and diverse and may counteract reli-
able phase synchronization15. Fourth, the frequency of the transient
oscillatory bursts fluctuates over time, varies between recording sites
and depends on characteristics of the presented stimuli16,17. A pri-
ori, the likelihood of gamma bursts to spontaneously match should
thus be very low.
Here we investigate a class of circuit models that naturally exhibits
extensive power, frequency and timing variability and that contains key
types of heterogeneity, including heterogeneous transmission delays.
In these models, below the onset of developing oscillatory synchrony,
collective oscillations are short-lived, with durations on the order
of a few cycles, are weakly synchronized and exhibit stochastically
drifting frequencies. When multiple circuits with these characteris-
tics are coupled by long-range excitatory connections, the resulting
large-scale dynamics spontaneously generates temporally co-occur-
ring bursts of synchrony. We find that the drifting frequencies of each
region track each other, giving rise to transient phase-locking within
the gamma bursts. Through state-resolved information-theoretical
analyses, we assessed whether these transient patterns of coherence
can gate information flow. We further examined the propagation of
external input signals relayed from different source regions, as well as
under which conditions these signals can be decoded from the activity
of a downstream target region. We find that the transient coherence
between the activities of the local circuits dynamically shapes the flow
of information between them: information transfer is either selec-
tively boosted or suppressed along different routes according to the
transient phase pattern. These distinctive routing states have a direc-
tion set by the transient phase-relations and are modulated by the
fluctuating level of synchrony. We find that the propagation of exter-
nally supplied information can be selectively gated on or off along
different pathways depending on the routing state. Unexpectedly, the
stochastic and fleeting nature of ongoing oscillations more effectively
modulates information flow than stronger and more coherent forms
of synchronized circuit activity.
RESULTS
The transient synchrony regime
Oscillatory neuronal activity in vivo is comprised of epochs of syn-
chronous activity arising from an overall poorly synchronized state.

1Max Planck Institute for Dynamics and Self-organization, Göttingen, Germany. 2Bernstein Center for Computational Neuroscience, Göttingen, Germany. 3Institute for
Nonlinear Dynamics, Georg-August University School of Science, Göttingen, Germany. 4SFB-889 Cellular Mechanisms of Sensory Processing, Göttingen, Germany.
5Aix-Marseille Université, Inserm, Institut de Neurosciences des Systèmes, Marseille, France. 6These authors contributed equally to this paper. Correspondence

should be addressed to A.P. (agos@nld.ds.mpg.de) or D.B. (demian.battaglia@univ-amu.fr).

Received 22 November 2016; accepted 25 April 2017; published online 22 May 2017; doi:10.1038/nn.4569

nature NEUROSCIENCE  advance online publication 

 a r t ic l e s

a b
No. 300

...
No. 1

16%

10 mV

100 ms

c Synchronization d Mean population e LFP spectrogram f Distribution of

index frequency (Hz) (a.u.) gamma bursts (a.u.)
10 3 10–3
40 40
2.5 2
0.8 80
2 2
50
© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.

2 50

Frequency (Hz)
Frequency (Hz)
0.6 60
4 4
�in (kHz)
�in (kHz)

1.5
60 60 1
6 0.4 6 40
1
70 70
8 0.2 8 20 0.5

10 10 80 80 0
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1 1 2 0 0.1 0.2 0.3 0.4
PI PI Time (s) Burst length (s)

Figure 1  The transient synchrony regime. (a) Scheme of the local connectivity of a brain region or area (inhibitory and excitatory neurons in blue and
red, respectively). (b) Simulated activity in the transient synchrony regime. Transient oscillatory bursts in LFP time-series (bottom) and multiunit activity
(middle, measured as percentage of active neurons). Individual neurons fire irregularly and not at every oscillation cycle, as indicated by the raster
plot of 300 excitatory neurons (top). (c) The degree of synchronization of local neuronal activity, as quantified by the synchronization index (Online
Methods), grows with the probability PI of recurrent inhibitory connections and the rate νin of the Poisson background drive to the region.
(d) Dependence of the frequency of collective oscillation on PI and νin (see Supplementary Fig. 1c for single-neuron firing rates). Dashed circles in c
and d highlight two working points, with similar collective oscillation frequencies in the gamma range but very different synchronization levels. Upper
(black) dashed circle (transient synchrony): PI = ~0.25 and νin = ~2.5 kHz. Central (gray) dashed circle (high synchrony): PI = ~0.55 and νin = ~5.5 kHz.
(e) Transient gamma bursts are evident in the spectrogram of LFP activity at the transient synchrony working point. (f) Joint distribution of the
frequencies and durations of these gamma bursts. Note the predominance of short-lived variable-frequency bursts.

We identified and characterized a regime of local circuit dynamics,
below the onset of oscillatory synchrony, that mimics this funda-
mental feature and reproduces the substantial fluctuations in both
frequency and power seen experimentally in local field potential
(LFP) recordings12,13.
The dynamics of a local circuit (or region) with transient synchrony
is characterized in Figure 1. A circuit is modeled as a large network of
excitatory and inhibitory spiking neurons with heterogeneous local
connectivity (Fig. 1a), synaptic conductances and latencies (Online
Methods; parameters summarized in Table 1). Figure 1b depicts the
evolution over time of the average membrane potential in the local
circuit, together with the spiking patterns of individual neurons and
a histogram of the multiunit activity rate (Online Methods). In the
following we call the population-averaged membrane potential, a
convenient descriptor of the collective neuronal activity, an LFP-like
signal18. As indicated in Figure 1b, the amplitude of the LFP trace
exhibits transient epochs of increased oscillatory amplitude. These
oscillatory events are more difficult to detect when inspecting the
spiking patterns of a few single units whose firing is always stochas-
tic19,20. Individual neurons can fire at every phase of the ongoing
oscillation, as revealed by an average circular variance of ~0.6 of the
overall distribution of the phases of firing.
In our local circuit, as in experimental observations21, the genera-
tion of oscillatory activity relies on delayed, recurrent interactions,
either occurring through direct delayed connections or mediated
by inhibitory–excitatory–inhibitory loops20. The level of recurrent
inhibition controls the overall level of synchrony of the network.
Enhancing recurrent inhibition induces a graded transition toward
higher synchrony. In Figure 1c this is shown via an increase of
the probability PI of establishing local inhibitory connections (see
Supplementary Fig. 1a,b for alternative mechanisms), as captured
by the network synchronization index (Online Methods). Increasing
local inhibition also leads to an overall reduction of the mean fre-
quency of the collective population oscillation (Fig. 1d) and of the
firing rates of individual neurons (Supplementary Fig. 1c).
Proportionally increasing both the strength of recurrent inhibi-
tion and the background drive (Supplementary Figs. 2 and 3) essen-
tially maintains the population frequency (Fig. 1d) and firing rates
(Supplementary Fig. 1c) but increases the synchronization in a way that
becomes smoother as the adopted parameter’s heterogeneity increases
(Supplementary Fig. 2). We define two working points (Fig. 1c,d
and Supplementary Figs. 1 and 2): one below the transition to syn-
chrony and a second one at a stronger synchrony level, both giving
rise to a collective oscillation frequency close to 60 Hz. Bursts of
transient oscillatory power are visible in the LFP spectrogram at the
working point below the transition to strong synchrony (Fig. 1e),
while higher synchrony levels lack this transient nature and generate
a persistently strong power peak in the spectrogram (Supplementary
Fig. 4a,b). Statistical analysis of the duration and frequency of these
bursts (their joint distribution is shown in Fig. 1f) shows that the
working point below the transition gives rise to gamma bursts with
frequencies broadly distributed over a range between 45 and 75 Hz

 advance online publication  nature NEUROSCIENCE

 a r t ic l e s

Table 1  List of default network parameters in the transient- a b LFP spectrograms (a.u.) 103
2
synchrony regime 40 Area X

Model parameter Symbol Value 1.5

60

Frequency (Hz)
Excitatory synaptic decay time τ1E 3 ms
80
Inhibitory synaptic decay time τ1I 4 ms 1
Area Y
40
Excitatory synaptic time τ2E 1 ms
60 0.5
Inhibitory synaptic time τ2I 1 ms
Mean synaptic delay d 1.5 ms 80
Area Y Area X 1 2 3
Mean synaptic excitatory conductance g sE 5 nS Time (s)
Mean synaptic inhibitory conductance g sI 200 nS c Same-frequency d Cross-frequency
burst-timing overlap burst-timing overlap
Mean input synaptic conductance g sν 3 nS
Chance level 40
Standard deviation delay σd 0.1 ms 0.5 0.45
95% c.i.

Burst-timing correlation

Frequency Area Y (Hz)

Standard deviation synaptic excitatory conductance s 1 nS 0.4 50
g sE 0.35

s 0.3
Standard deviation synaptic inhibitory conductance 10 nS 60 0.25
g Is
0.2
© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.

Standard deviation input synaptic conductance s 1 nS 0.15

g vs
Probability of local inhibitory connection PI 0.2
Probability of local excitatory connection PE 0.3
Probability of interareal excitatory connection PELR 0.08
  (two-regions model)
Probability of feedforward excitatory connection   PEff 0.08
  (three-regions model)
Probability of feedback excitatory connection PEfb 0.01
  (three-regions model)
Poisson input rate ν 10·PI kHz
and a short mean duration close to 100 ms (i.e., four to six oscillation
cycles). We call this the ‘transient synchrony regime’. It represents a
multidimensional range of parameters (cf. Supplementary Figs. 1–3)
leading to a qualitatively similar mixed dynamic in which asynchro-
nous and synchronous epochs both stochastically occur. Its properties
are in good agreement with in vivo LFP recordings12,13.
In the following we consider model networks with transient syn-
chronies, representing multiple brain regions and linked into simple
structural-connectivity motifs, and analyze their emergent dynamics.
In particular, we study their coordinated oscillatory bursting dynamics
and their phase locking properties and, finally, test whether these fea-
tures can subserve the selective and flexible routing of information.
Simultaneous emergence of phase-locked gamma bursts
We first treated the simplest case of two local circuits (corresponding
to two generic areas or regions X and Y), coupled by long-range exci-
tatory projections (Fig. 2a) established with equal probability (PELR )
and strength in both directions (i.e., X-to-Y and Y-to-X). If the dura-
tion and the frequency of oscillatory epochs fluctuate independ-
ently in each brain region, it would be extremely unlikely to observe
simultaneous oscillatory bursts in a source and a target region aligned
in a well-defined phase relation. To analyze the inter-regional
coordination (or lack of thereof) we examined the relative timing,
frequencies and phase relationships of gamma bursts in the two
connected regions.
We found that our model spontaneously generates correlated
episodes of matched timing and with tracking frequencies between
stochastically emitted gamma bursts in the two connected regions.
This is depicted in the spectrograms in Figure 2b. The burst-tracking
phenomenon can be quantified by the normalized overlap between
the occurrence times of oscillatory bursts in distinct regions as a
function of their instantaneous frequency (Online Methods). Figures
2c,d reveal that, for frequencies in the gamma band, the overlap
between the timing of bursts with matching frequencies can reach
70
0.1
0.05
0 80
40 50 60 70 80 40 50 60 70 80
Frequency (Hz) Frequency Area X (Hz)
Figure 2  Co-emergence and frequency tracking of gamma bursts.
(a) Scheme of a canonical circuit of two brain regions reciprocally
connected through long-range excitation (dashed lines). (b) Spectrograms
of LFP activity for coupled circuits in the transient synchrony regime. The
times and frequencies of emergent coordinated bursts match in the two
regions. (c) Burst-timing normalized overlap between same-frequency
gamma bursts compared with chance level (gray). Shaded band in blue
denotes 95% c.i. (d) Burst-timing overlaps between oscillatory bursts with
arbitrary frequencies. Strict frequency-matching is not required for burst-
timing coordination.
relatively high values, peaking at ~0.5 for nearby collective oscilla-
tion frequencies of 60 Hz. A similar frequency-tracking behavior has
been observed experimentally22. We find that coordination between
oscillatory episodes occurs not only between bursts with an exactly
matching main frequency (Fig. 2c) but also in a cross-frequency
manner (Fig. 2d).
Beyond time and frequency coordination, we find that oscillatory
bursts also exhibit transient phase synchronization (Fig. 3). We per-
formed a time-resolved cross-covariance (XC) analysis of the LFP sig-
nals from the two coupled regions of the structural motif in Figure 2a
(Online Methods). Transiently rising values of XC denote the onset
of epochs of increased inter-region oscillatory coherence (Fig. 3a).
The time dependent maximum of XC (denoted XC*(t)) occurs at a
fluctuating time-lag τ*(t), generally different from zero (in-phase)
and from the average half-period (antiphase), indicating out-of-phase
locking, as observed empirically23. The relative phase of the rhythms
of the regions, ∆Φ (Online Methods), is a strongly fluctuating quan-
tity whose distribution, ρ(∆Φ), is broad (Fig. 3b).
The precision of phase-locking markedly increases with the instan-
taneous level of cross-covariance (Fig. 3c). As shown by the joint
distribution of ∆Φ and XC* (Fig. 3c), events of higher-value XC* are
less frequent than those with lower values, but they tend to occur at
specific relative phases. Conditioning the joint distribution on the
values of XC* better reveals this phenomenon (Fig. 3c). For low val-
ues, corresponding to a poorly synchronized baseline activity, the
distribution of the relative phase ∆Φ remains broad (Fig. 3d). During
epochs of strong cross-covariance, however, it becomes prominently
bimodal (Fig. 3d).
A similar preference for out-of-phase interareal phase-locking
is observed at the higher-synchrony working point of our system

nature NEUROSCIENCE  advance online publication 

 a r t ic l e s
a Intermittent gamma
LFP cross-covariance (XC)
c log(P(∆Φ, XC*))
1 0
−10
Lag (ms)
XC*
0 0 0.5
10
−1 1 −20
0 1 2 3
Time (s)
b d
Low XC*
0.5
0.25 ∆Φ *
↑ ∆Φ↑* ∆Φ↓*
�(∆Φ) = 0.01
0.5 0 ∆Φ↑*
∆Φ↓* 0 0.5 1
0.75 ∆Φ High XC*0.5
© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.
Figure 3  Transient phase-locking. (a) Time-resolved cross-covariance (XC)
between simulated LFPs of the two coupled brain regions of Figure 2a.
Circuits are set at the same working point. Horizontal dotted lines mark
τ* = 0 ms (white) and t * = ± T / 2 (black) with T equal to the length of the
average period. The positions of the highest XC maxima, displaced with
respect to these guiding lines, reveal the occurrence of transient episodes
of interareal phase-locking with out-of-phase relations. (b) Distribution
of instantaneous interareal phase-differences between simulated LFPs,
in polar (left) and linearly unwrapped (right) histogram forms. Note the
two peaks at relative phases ∆Φ↓* > 0.5 and ∆Φ↑* < 0.5. (c) Left: joint
probability distribution of the relative phase ∆Φ and the time-dependent
maximum XC* of the cross-covariance XC. Right: the joint distribution
is conditioned on the value of XC*. Note the enhanced peaks associated
with the high-synchrony events. (d) Polar (left) and linearly unwrapped
(right) histograms of relative phase, conditioned on XC* being lower (top)
or higher (bottom) than threshold levels (Online Methods). Relative phases
close to ∆Φ↓* or ∆Φ↑* dominate when XC* is high.
(Supplementary Fig. 4). In this case, the simulated LFPs
persistently oscillate and the two phase-locking configurations are
stable attractors (Supplementary Fig. 4c), giving rise to a promi-
nently bimodal ∆Φ distribution (Supplementary Fig. 4d). The
gamma-bursting dynamics can be viewed as transiently replicating
oscillatory modes that would be stable in the high-synchrony regime
(Supplementary Fig. 5).
Spiking activity is only weakly modulated by oscillations on aver-
age. Nevertheless, we found that the phase-concentration of neuronal
firing transiently rises to higher values during high XC* transients
(Supplementary Fig. 6). The instantaneous circular variances of the dis-
tribution of firing phases drops to ~0.3 during the strongest coherence
oscillatory bursts and rises as high as ~0.8 during interburst periods.
Emergent routing states
The canonic two-regions motif of Figure 2a can be used as a model of
two interconnected brain regions at two different levels in the cortical
structural hierarchy, such as, for example, a sensory region and a pre-
frontal cortical (cf. the scheme in Fig. 4a) interacting in an attention-
modulated manner10 or a frontal and parietal pair of regions during
different stages of motor planning24. Others have observed that the
direction of influence between such region pairs can change depend-
ing on task and behavioral stage10,25. We thus examined how the tran-
sient dynamics of gamma bursts emerging in our models (Fig. 3)
modulate the direction and strength of information flow. We quantify
 advance online publication  nature NEUROSCIENCE
log(P(∆Φ|XC*)) the efficiency and direction of inter-region information exchange of
−4 −1 spontaneous activity depending on the system’s states by means of
−3
information theoretical analyses.
A useful measure is provided by transfer entropy (TE) 26,27, here
−12
−5
computed based on LFPs from different regions (see Supplementary
−7 Fig. 7 for an equivalent analysis based on multiunit activity). TE, in a
0 0.5 1 0 0.5 1
model-free fashion, quantifies how much the knowledge of the past
∆Φ ∆Φ
activity of a putative source region improves the prediction of the
0.25 ∆Φ *
↑
∆Φ↑* ∆Φ↓* future activity of a target system; in this sense it constitutes a generali-
�(∆Φ) = 0.01
zation of the Granger causality approach (exactly as mutual informa-
0 tion (MI) generalizes linear correlation). The information-theoretical
0 0.5 1
setting naturally allows conditioning the estimation of TE or MI on
∆Φ↓*
an arbitrary set of constraints. We designed state-selecting filters that
0.75 ∆Φ
∆Φ↓*
0.25 ∆Φ * condition information measures on transient dynamical features (for
↑
example, on specific relative phase conditions to be fulfilled) and so
�(∆Φ) = 0.01
0
define a state-specific directed functional connectivity (or, in short,
a routing state). As summarized in Figure 4b, transients of a sys-
0.75 ∆Φ↓*
tem’s activity fulfilling specific state-filtering conditions are pooled,
0 0.5 1
∆Φ through time and trials, into separate statistical samples. Different
distributions of activity were obtained from these separate samples
and quantified information-theoretically (Online Methods). Using
this approach we probed information transfer specifically within
state-filter-selected epochs.
Using a TE analysis in which the relative phase is constrained to a
narrow band, we study state-specific directed functional connectivity
between the collective neuronal activity of the regions X and Y (Fig. 4c).
For both directions of interaction, TE exhibits a broad peak cen-
tered on a specific combination of the state delay τ and ∆Φ (Fig. 4d
and Online Methods). In both panels, the state delay, τ, that maxi-
mizes information transfer corresponds to a lag of ~4 ms, at which
XC(t) peaks during coordinated bursting events (Fig. 3a). This opti-
mal state delay τ depends on the metastable phase locking (∆Φ↑* , ↓)
and therefore differs from the mean interareal transmission delay
d of 1.5 ms (Supplementary Figs. 5 and 18 and “Discussion” section).
For longer state delays, TE does not rise significantly above chance
level (P < 0.05, bootstrap confidence interval (c.i.) comparison over the
whole white-colored range in Supplementary Fig. 8a). Optimal func-
tional coupling occurs at different relative phases for different direc-
tions of influence. This phenomenon is quantified by the functional
anisotropy index, ∆TE, proportional to (TEY→X − TEX→Y) (Fig. 4e
and Online Methods). Two patches with equal absolute values, |∆TE|,
but different signs correspond to effectively unidirectional configu-
rations with opposite directions (Fig. 4f), supported by the same
fixed bidirectional structural connectivity. We refer to the states cor-
responding to these two configurations, filtered by the ∆Φ↑ (Y leads
X, 0 < ∆Φ < 0.5) and ∆Φ↓ (X leads Y and 0.5 < ∆Φ < 1) conditions, as
the Top and Bottom routing states, respectively. Information transfer
proceeds from the region leading in phase to the region lagging in
phase, as if long-range synaptic connections from the laggard to the
leader region were effectively nonfunctional.
Based on analysis of the relative phase-dependency of TE in Figure 4d,e
we also adopted coarser state-filtering criteria, using solely the ∆Φ↑
and ∆Φ↓ conditions. Figure 4g,h shows the net directed information
transfer in the Top (Fig. 4g) or Bottom routing states (Fig. 4h) com-
pared to chance-level expectations (Online Methods). Information
transfer in the opposite laggard-to-leader direction falls below chance
level, thus actively quenching information transfer.
Information transfer in the leader-to-laggard direction was highly
significant despite the coarseness of the ∆Φ↓ and ∆Φ↑ conditions
(P < 0.0001, bootstrap c.i. comparison). This indicates that the emer-
gence of information transfer anisotropy requires just a weak degree
 a r t ic l e s

a b
X Y X leads Y Y leads X X leads Y Y leads X Top

Population
activity
X ...
... ...
SY TE/MI
Y
analysis Bottom
SX
1

State-filtering
condition
∆Φ↑ 0
1 State-dependent State-dependent
∆Φ↓ 0 distributions functional states

c d TEY→X (bits) e ∆TE g ∆Φ↑

i j ∆Φ↑
l 0.06 ∆Φ↑
0.25 1
1 1 0.25 0.2 TEY→X MISY↔X
Y Y TEX→Y MISX↔Y

TE (bits)

MI (bits)
3

TE (bits)
3 0.15
τ (ms)

τ (ms)
0 0.15 0.04
0.1
5 5 SY
0.05 0.05
7 7
−1 0 0.02
0 0.5 1 0 0.5 1 X Y Y X 0 5 10 0 5 10
∆Φ ∆Φ τ (ms) τ (ms)
© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.

TEX→Y (bits) f h SX
k m
0.25 ∆Φ↓ ∆Φ↓ ∆Φ↓
1 0.25 0.2 0.06
TEY→X MISY↔X
3 TEX→Y MISX↔Y

TE (bits)
TE (bits)

MI (bits)
0.15
τ (ms)

0.15 X 0.04
X Bottom 0.1
5
0.05 0.05
7 Top
0 0.02
0 0.5 1 X Y Y X 95% c.i. 0 5 10 0 5 10
∆Φ τ (ms) τ (ms)
Chance level
∆Φ↑ ∆Φ↓

Figure 4  Information transfer during transient burst events. (a) Cartoon illustrating potentially interacting areas in a manner similar to X and Y; for
example, visual area V4 and frontal eye field (FEF). Inputs to these areas are modeled as temporally fluctuating signals SX and SY. (b) Scheme of the
state-filtering pipeline. LFP epochs fulfilling the state-selecting constraints (in this case probing the relative phase ∆Φ) are pooled, sampling separate
neural-activity distributions. State-dependent information-transfer direction and efficiency are evaluated from these distributions. (c) Scheme of the
model examined in panels d–h (see also Fig. 2a). (d) Maps of TE between LFPs of the X and Y regions, resolved as a function of ∆Φ and the state-delay τ.
White, TE values below chance. Information transfer is always significantly higher from the phase leading to the phase lagging region than vice versa.
(P < 0.05 at least over all the white-colored range, bootstrap c.i. comparison). (e) Functional anisotropy index, ∆TE. Colored lobes with different
signs indicate the existence of two distinct routing states, Top and Bottom, associated with effectively unidirectional information transfer in opposite
directions. (f) Graphical representation of the routing states as functional motifs (colors match the regions represented in c and e). (g,h) Information
transfer in different directions, in (g) a ∆Φ↑ configuration (Top state) or (h) a ∆Φ↓ configuration (Bottom state), for the optimal state delay. Boxes and
shaded bands show the 95% c.i.; shaded gray bands indicate chance-level values. (i) Scheme illustrating the input-driven routing model analyzed
in panels j–l. Input streams SX and SY are provided as inputs to every neuron in each area (see also Supplementary Fig. 13). (j,k) TE as a function of
the state delay τ for the model in i, in a Top state (j) and a Bottom state (k). The routing states, as measured by TE, are unidirectional and switch as a
function of the phase relation. (l,m) MI as a function of the state-delay, τ, between the input stream to one source region and the neuronal activity in
the target region, in a Top state (l) and in a Bottom state (m). The input to a source region only reaches the target region when the rhythm of the source
region leads in phase that of the target area.

of phase-locking. In fact, we find unidirectional transfer of informa- In particular, it is not obvious that streams of external information
tion during Top and Bottom transients to be robust, persisting within entering the circuits are routed in a fashion similar to the spontane-
extended ranges of local synaptic strength (Supplementary Fig. 9), ous activity analyzed so far. To examine this aspect in our models,
interareal coupling (Supplementary Figs. 9 and 11a) and synchroni- we introduce external streams of input currents, SX and SY, modeling
zation levels (Supplementary Figs. 10 and 11b), and we find that such incoming activity from surrounding areas or direct sensory stimula-
information transfer emerges for distinct mechanisms for the gen- tion (for a characterization of this state, see Supplementary Fig. 13).
eration of oscillations (Supplementary Fig. 12) or indicators of the We then studied how these input streams propagate through the inter-
population’s activity (Supplementary Fig. 7). We further restricted connected regions conditional on the current routing state (Fig. 4i
the analyses in Figure 4g,h to transients with peak XC* rising above and Online Methods). The addition of external input did not modify
(Supplementary Fig. 8c,d) or dropping below (Supplementary Fig. 8e) the system’s states as defined by TE, which remain markedly unidi-
different threshold levels (Online Methods). We found that informa- rectional (Fig. 4j,k). State-resolved information theoretical analysis
tion is transmitted primarily inside the bursts. The interareal syn- between the input streams SY (or SX) to the source region Y (or X) and
chrony thus works as a gain control that boosts information transfer the neuronal activity of the target region X (or Y) revealed that the
in one of the two directions when synchronization is high and limits information supplied to a source region propagated toward an inter-
it during asynchronous epochs. connected distant target, following the path imposed by the routing
states, defined by the state-specific TE analysis. The phase-relations
Flexible routing of information streams between the neuronal rhythms define the direction of information
The existence of a phase-gated dynamic switching between Top and flow of input signals, as revealed by the MI analysis of Figure 4l,m
Bottom routing states only starts to address the question of routing. and the TE analysis of Supplementary Figure 14.

nature NEUROSCIENCE  advance online publication 

 a r t ic l e s

a Z leads X leads Y Z leads Y leads X

X leads Y leads Z

... ... ...

Population
X TE/MI
activity Y ... ... analysis

... ... ...

∆Φ2XYZ
State-filtering
condition

∆Φ3ZXY

∆Φ4ZYX State-dependent State-dependent

distributions functional states

c Y
d
∆Φ1XZY ∆Φ2XYZ ∆Φ3ZXY ∆Φ4ZYX ∆Φ5YXZ ∆Φ6YZX
b X Z
X Y
SX

Y MIS
e f MIS Z Y
© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.

Z Y
MIS MIS
0.04 X Y ∆Φ2XYZ X Y
∆Φ4ZYX
X Z

MI (bits)
SZ
Z 0.02

SX SZ 0 95% c.i.
0 5 10 0 5 10
τ (ms) τ (ms) Chance level

Figure 5  Flexible routing of input signals through a hierarchy of areas. (a) Scheme showing the state-filtering pipeline for three interconnected regions,
analogous to the two-regions case described in Figure 4b. Six state filters are defined for all permutations of X, Y and Z. (b) Cartoon of a macaque brain
with three recoding sites: brown and blue in primary visual cortex V1 and red in extrastriate cortex, for example, V4. Also indicated: screen with two
stimuli (slashed circles), fixation spot (×) and the receptive fields of the recording sites in matched colors. This experimental configuration provides
a possible exemplification of the three-regions model studied here. (c) Scheme of three populations, modeling the two V1 and the V4 sites as circuits
in the transient synchrony regime. The connectivity is asymmetric, with stronger feedforward (i.e., V1 to V4) excitatory projections. (d) Functional
motifs for the associated state selecting filters. Configurations in which Y is lagging in phase behind the other areas (∆Φ1XZY and ∆Φ3ZXY) give rise to a
convergent functional motif in which both regions transmit information to Y, although with different time-courses (Supplementary Fig. 14). Functional
motifs in which Y leads both X and Z in phase are functionally disconnected, reflecting the weakness of the back-projections. Functional motifs
associated with ∆Φ2XYZ and ∆Φ4ZYX are functionally unidirectional despite the symmetrical connectivity to Y from the lower order regions X and Z.
(e) As in c except that additionally, each region in the model can receive independently fluctuating input signals (Online Methods). For clarity, only SX
and SZ are shown. (f) Left: MI between the input signal SX to region X, one of the two possible source regions, and the LFP activity of a target region Y
(as a function of the state delay; blue curve) is strongly enhanced in the associated phase configuration ∆Φ2XYZ. This phase configuration, in which X
leads Y in phase, implements a routing state mimicking attention directed toward the stimulus in the blue receptive field. No significant information
can be retrieved from Y about the signal SZ, representing the nonattended stimulus, as evidenced by the second MI curve (brown) at baseline levels.
Right: phase configuration ∆Φ4ZYX, implementing a routing state that mimics attention directed to the stimulus in the brown receptive field. MI between
the LFP activity of the target region Y and the input signal SZ is now highly significant. The routing states associated with each phase configuration are
represented by the functional motifs drawn next to the MI plots.

Next, we further extended the model to include three intercon-
nected regions, X, Y and Z. As shown in Figure 5a, the state-selection
procedure can be extended straightforwardly to this case. In this sys-
tem, six different conditions can arise, denoted by the state-selecting
YXZ
filters ∆Φ1XZY to ∆Φ6 (where ∆Φ1XZY indicates a first-phase con-
figuration in which X leads Z leads Y within a cycle; Supplementary
Fig. 15 and Online Methods). Depending on parameters and inputs,
a multitude of possible functional motifs can arise.
This three-region architecture enabled us to study a setting typical
of visual selective-attention experiments11,28,29. For instance, one can
consider a situation in which two distinct visual stimuli are processed
by distinct populations X and Z in lower-order areas such as parts of
V1, whose distinct receptive fields both fall within the larger field
of a third population, Y, in a higher-order area such as V4 (Fig. 5b).
In such an experiment, the focus of attention may be alternatively
directed to one stimulus or the other, defining an information-rout-
ing problem in which the transmission toward Y of the input stream
entering X (or Z) and associated with the attended stimulus must
be enhanced, while the other, unattended, stream entering Z (or X)
should be gated off.
A routing-state analysis of the three-region system (Fig. 5c) is
shown in Figure 5d and Supplementary Figure 15. In this setting,
X and Z send convergent projections toward a third higher-order
region, Y, which, in its turn, sends weaker feedback projections
to X and Z (for a system with equally strong back-projections see
Supplementary Fig. 15). TE analysis of the LFP signals of the three
coupled regions shows that multiple routing states can arise, depending
on the phase pattern. In particular, in the routing states defined by the
phase configurations ∆Φ2XYZ and ∆Φ4ZYX, information transfer along
one of the possible feedforward pathways is completely blocked.
This allows selection of only one of the two possible information
sources when external input streams are supplied to the system (Fig. 5e).
We focus here on routing of the signals SX and SZ, coding for
different stimuli (for more routing examples see Supplementary
Fig. 16). In a phase configuration in which X leads Y leads Z, only
information about the input stream SX is present in the down-
stream region Y (Fig. 5f), although no other mechanism besides an
unsuitable phase relation prevents information about SZ from arriv-
ing at Y. Switching between the routing states detected by the ∆Φ2XYZ
and ∆Φ4ZYX conditions thus produces rerouting effects analogous

 advance online publication  nature NEUROSCIENCE

 a r t ic l e s

Y Y Y

Y X Y X Y X X Z X Z X Z

∆�Y = 4 kHz ∆�Y = 2 kHz ∆�X,Y = 0 kHz ∆�X = 2 kHz ∆�X = 4 kHz ∆�X = 4 kHz ∆�X = 2 kHz ∆�X,Z = 0 kHz ∆�Z = 2 kHz ∆�Z = 4 kHz

a d

∆ΦXY ∆ΦXY ∆ΦXY ∆ΦXY ∆ΦXY

b �(∆Φ) = 0.01
∆TE
1 1

3
τ (ms)

0
5

7
−1
0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1 ∆ΦZY ∆ΦZY ∆ΦZY ∆ΦZY ∆ΦZY
∆ΦXY ∆ΦXY ∆ΦXY ∆ΦXY ∆ΦXY

c
© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.

Fraction of time in which X leads Y Fraction of time in which Z leads Y

Fraction of time Fraction of time Fraction of time in which Y leads X Fraction of time in which Y leads Z
∆Φ in which Y leads X ∆Φ in which X leads Y

Figure 6  Steering information transfer. (a) The application of a constant bias is modeled as an increase ∆ν of the background drive νin of only one of the
two interconnected regions. As a result, the phase distribution of phase relations is modified. Middle position: no input bias. Extreme positions: strong
background bias ∆ν = +4 kHz (in addition to a baseline drive of ν = 3 kHz) applied to population X (extreme right) or to Y (extreme left). Intermediate
columns: weaker cases of ∆ν = +2 kHz. (b) The predominant direction of information transfer is biased toward the region with lower background drive,
as indicated by the maps of the functional anisotropy index, ∆TE. (c) Pie charts of the time the system spends in a Top or a Bottom state. The functional
motifs for the corresponding routing states are indicated below. (d) As in a but for a three-regions model. Center: without background bias, the dynamics
of a three-region system spends the same fraction of time in a state in which both X leads Y or Z leads Y (note that there is already a weak asymmetry
in the baseline distributions of ∆ΦXY and ∆ΦZY, reflecting the different strengths of feedforward and feedback connectivity; cf. Fig. 5c). When a mild
bias is applied to population X (or Z; left and right positions respectively), the fraction of time in which X leads Y and Y leads Z (or, correspondingly,
in which Z leads Y and Y leads X) increases, eventually suppressing the fraction of time spent in the mirror configuration. Information is preferentially
transmitted to the target downstream region Y from the region receiving the background bias. See Supplementary Figure 17 for details.

to those that would be associated with a reorientation of the applied bias (extreme right or left columns) eventually mutes informa-
attentional spotlight. tion transfer in the subdominant direction.

Steering information routes DISCUSSION

In the examples of Figs. 4 and 5, transients of the different routing
states (Top and Bottom for the two-regions model and ∆Φ1XZY to
∆Φ6YXZ for the three-regions model) arise spontaneously, with simi-
lar probabilities of occurrence. We found that applying a weak bias,
making the bursts of a given type more likely to occur, can enforce a
dominant direction of information transfer (Fig. 6).
Figure 6a–c refers to routing-state control in the two-regions model
(Fig. 4), and Figure 6d refers to one in the three-regions model (Fig. 5).
Raising the baseline input to either one of the two possible source
regions by ∆ν increases the probability that the biased region will
transiently become a phase leader, favoring the occurrence of spe-
cific routing states. As indicated by the pie charts in Figure 6c,
a bias to circuit X (or Y) increases the probability of occurrence of
a Bottom (or Top) state, as seen in the rightmost (or leftmost) col-
umns. Equivalently for the three-regions model, a bias to the region
X increases the probability of X leading Y and, simultaneously, of
Y leading Z (Fig. 6d). Similarly, a bias to the region Z increases the
probability of Z leading Y and, simultaneously, of Y leading X (right
columns). As the relative phase-distributions (Fig. 6b and
Supplementary Fig. 17) take on more asymmetric shapes, maps of
the functional anisotropy index reveal a strong asymmetry in the
direction of information transmission, showing that information
flows preferentially from the biased region toward the target one.
This is detailed in Figure 6b for the two-regions model in and in
Supplementary Figure 17 for the three-regions model. The highest
Our results demonstrate that, in coupled circuits with transient syn-
chrony, frequency tracking and out-of-phase locking of oscillatory
bursts are emergent features of the large-scale circuit’s collective
dynamics. Bursts are generated in coordinated sets of near-simulta-
neous onset and mediate transient phase-locking between the regions’
activities. The transient patterns of inter-regional coherence define
flexible routing states that determine the direction of information
flow of both endogenously generated rate fluctuations and externally
supplied signals.
While the mechanisms of transient interareal frequency tracking
and phase-locking in biologically plausible models were largely unex-
plored, evidence for such patterns of coherence has been reported
experimentally22 (but questioned elsewhere15). Distinct from the
typical behavior of chains of weakly coupled oscillators 30, burst
coordination in our model is fast enough to conciliate transient
phase-locking with a rapid dissolution of oscillatory synchrony after
a few cycles. This intrinsic and spontaneous coordination is achieved
despite the sparseness of interareal connections and the heterogene-
ity of conduction delays. Our work demonstrates that fast dynamic
frequency-matching is a robust emergent feature of interacting neu-
ronal populations with transient synchrony and does not require an
external controller.
Our study of the dynamics of a general regime of neuronal net-
work activity with transient synchrony sheds a new light on long-
standing arguments against coherence-mediated communication.

nature NEUROSCIENCE  advance online publication 

 a r t ic l e s
A first widespread concern has been the low and inconsistent power of
gamma oscillations in vivo12,13. In our models, LFP oscillations show
inconsistent power and fluctuate stochastically between 40 and 70 Hz,
and their time-averaged LFP spectrum presents broadband gamma
modulation (Supplementary Fig. 6f), consistent with other models of
transiently synchronous neuronal activity31. Despite the weak average
gamma power, TE analyses revealed that the transient power rise dur-
ing oscillatory bursts is sufficient to impact on information transfer
between areas and on the propagation of externally supplied informa-
tion. A second concern relates to the low level of spiking correlations.
While several studies identify marked and context-dependent oscil-
latory synchrony32, others find only weaker spiking correlations33.
In our simulations, spiking activity always remained highly irregular.
In particular, neuronal firing was poorly phase-concentrated, with
the exception of burst events (Supplementary Fig. 6). These short
epochs nevertheless convey the largest net contribution to directed
information transfer (Supplementary Fig. 8c,d). A third concern is
© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.
raised by the large variability of interareal conduction delays, imped-
ing the phase-matched arrival of presynaptic spikes with respect to
the oscillation in the target region. The regulation of information
transfer, however, is the outcome of collective emergent states and
not just of direct monosynaptic interactions. In fact, observed inter-
areal phase relations are largely independent of the distance between
the interacting sites34. The average interareal cross-correlation lag
is not determined solely by the neuronal conduction delay but also
depends on the phase response properties of the coupled popula-
tions35,36, which in turn reflect local circuit parameters37. As shown in
Supplementary Figure 5, different interareal phase-locking relations
can be implemented with a fixed interareal delay just by changing the
level of local inhibition. It is thus conceivable that the fine control of
local inhibition exerted by specialized interneuron types38 might play
an active role in setting interareal phase relationships.
Our model predicts that transitions between alternative routing
states, multiplexed within a fixed structural circuit39, are reflected by
switching patterns of interareal phase differences40. Such a routing
mechanism would be compatible with both rate coding and more-
complex representations relying on finer organization of the spiking
patterns41,42. Given the irregular spiking in our models, coordinated
excitability fluctuations may gate the transmission of complex spike
patterns. Future studies could, for instance, explore whether the
simultaneous emergence of synchronized bursts can modulate infor-
mation propagation through embedded systems of interacting synfire
chains5,6 or whether stimulus-dependent cell assemblies43,44 could be
selectively gated by interareal phase coordination.
The phase relations we find may be potentially useful in establish-
ing dynamic functional hierarchies of interareal communication23,
with information flowing preferentially from phase-leading toward
phase-lagging regions. Since the structures of these phase patterns are
not solely prescribed by connectivity, different regions can compete
flexibly, assuming the functional roles of sender or receiver of infor-
mation, depending on the specific context and modulatory signals
(Fig. 6 and Supplementary Fig. 17).
Interareal interactions may be mediated not only by coherence in
the gamma band but also by that in the beta band45 or by coordinated
faster and slower oscillations, as in the hippocampus46,47. For simplic-
ity, our model was designed to generate transient oscillations within
a single gamma band. This specification, however, is not inherent to
our approach and can be relaxed to reproduce multifrequency oscilla-
tions48. At a brain-wide scale, functional networks associated with mul-
tifrequency spectral fingerprints may be actively generated49,50. Future
studies could thus also examine more complex models, including
 advance online publication  nature NEUROSCIENCE
multifrequency interactions or multiareal networks of greater
complexity. Our study lays out a methodology that can be used to
assess the dynamic routing properties of such models stringently and
critically and can guide the design of future experimental studies.
Features that at first sight appear to be noncompliant with informa-
tion routing may actually provide the brain with a particularly flexible
routing mechanism.
Methods
Methods, including statements of data availability and any associated
accession codes and references, are available in the online version of
the paper.
Note: Any Supplementary Information and Source Data files are available in the
online version of the paper.
Acknowledgments
We thank U. Ernst, C.M. Gray, A. Kreiter, K. Pawelzik and R. Shapley for discussions.
This work was partially supported by the Federal Ministry for Education and
Research (BMBF) under grant no. 01GQ1005B (to A.P., T.G., F.W. and D.B.), by a
GGNB Excellence Stipend of the University of Göttingen (to A.P.), through CRC
889 by the Deutsche Forschungsgemeinschaft and by the VolkswagenStiftung
under grant no. ZN2632 (to F.W.), and by the FP7 Marie Curie career development
fellowship IEF 330792 (DynViB) (to D.B.).
AUTHOR CONTRIBUTIONS
A.P. performed the simulations of the models and analyzed the results;
A.P., F.W. and D.B. conceived the study, designed models and developed analysis
pipelines; A.P., T.G., F.W. and D.B. wrote the paper. All authors discussed the results
and implications.
COMPETING FINANCIAL INTERESTS
The authors declare no competing financial interests.
Reprints and permissions information is available online at http://www.nature.com/
reprints/index.html. Publisher’s note: Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
1. Olshausen, B.A., Anderson, C.H. & Van Essen, D.C. A neurobiological model of
visual attention and invariant pattern recognition based on dynamic routing of
information. J. Neurosci. 13, 4700–4719 (1993).
2. Vogels, T.P. & Abbott, L.F. Gating multiple signals through detailed balance of
excitation and inhibition in spiking networks. Nat. Neurosci. 12, 483–491
(2009).
3. Zylberberg, A., Fernández Slezak, D., Roelfsema, P.R., Dehaene, S. & Sigman, M.
The brain’s router: a cortical network model of serial processing in the primate
brain. PLoS Comput. Biol. 6, e1000765 (2010).
4. Abeles, M., Hayon, G. & Lehmann, D. Modeling compositionality by dynamic binding
of synfire chains. J. Comput. Neurosci. 17, 179–201 (2004).
5. Kumar, A., Rotter, S. & Aertsen, A. Conditions for propagating synchronous spiking
and asynchronous firing rates in a cortical network model. J. Neurosci. 28,
5268–5280 (2008).
6. Hahn, G., Bujan, A.F., Frégnac, Y., Aertsen, A. & Kumar, A. Communication through
resonance in spiking neuronal networks. PLoS Comput. Biol. 10, e1003811–
e1003816 (2014).
7. Akam, T. & Kullmann, D.M. Oscillations and filtering networks support flexible
routing of information. Neuron 67, 308–320 (2010).
8. Harnack, D., Ernst, U.A. & Pawelzik, K.R. A model for attentional information
routing through coherence predicts biased competition and multistable perception.
J. Neurophysiol. 114, 1593–1605 (2015).
9. Fries, P. A mechanism for cognitive dynamics: neuronal communication through
neuronal coherence. Trends Cogn. Sci. 9, 474–480 (2005).
10. Gregoriou, G.G., Gotts, S.J., Zhou, H. & Desimone, R. High-frequency, long-range
coupling between prefrontal and visual cortex during attention. Science 324,
1207–1210 (2009).
11. Grothe, I., Neitzel, S.D., Mandon, S. & Kreiter, A.K. Switching neuronal inputs by
differential modulations of gamma-band phase-coherence. J. Neurosci. 32,
16172–16180 (2012).
12. Burns, S.P., Xing, D. & Shapley, R.M. Is gamma-band activity in the local field potential
of V1 cortex a “clock” or filtered noise? J. Neurosci. 31, 9658–9664 (2011).
13. Xing, D. et al. Stochastic generation of gamma-band activity in primary visual cortex
of awake and anesthetized monkeys. J. Neurosci. 32, 13873–13880a (2012).
14. Jia, X., Tanabe, S. & Kohn, A. γ and the coordination of spiking activity in early
visual cortex. Neuron 77, 762–774 (2013).
15. Ray, S. & Maunsell, J.H.R. Do gamma oscillations play a role in cerebral cortex?
Trends Cogn. Sci. 19, 78–85 (2015).

16. Ray, S. & Maunsell, J.H.R. Differences in gamma frequencies across visual cortex
restrict their possible use in computation. Neuron 67, 885–896 (2010).
17. Jia, X., Xing, D. & Kohn, A. No consistent relationship between gamma power and
peak frequency in macaque primary visual cortex. J. Neurosci. 33, 17–25
(2013).
18. Okun, M. et al. Diverse coupling of neurons to populations in sensory cortex. Nature
521, 511–515 (2015).
19. Brunel, N. & Hakim, V. Fast global oscillations in networks of integrate-and-fire
neurons with low firing rates. Neural Comput. 11, 1621–1671 (1999).
20. Brunel, N. & Wang, X.-J. What determines the frequency of fast network oscillations
with irregular neural discharges? I. Synaptic dynamics and excitation-inhibition
balance. J. Neurophysiol. 90, 415–430 (2003).
21. Bartos, M., Vida, I. & Jonas, P. Synaptic mechanisms of synchronized gamma
oscillations in inhibitory interneuron networks. Nat. Rev. Neurosci. 8, 45–56
(2007).
22. Roberts, M.J. et al. Robust gamma coherence between macaque V1 and V2 by
dynamic frequency matching. Neuron 78, 523–536 (2013).
23. Bastos, A.M., Vezoli, J. & Fries, P. Communication through coherence with inter-
areal delays. Curr. Opin. Neurobiol. 31, 173–180 (2015).
24. Chakrabarti, S., Martinez-Vazquez, P. & Gail, A. Synchronization patterns suggest
different functional organization in parietal reach region and dorsal premotor cortex.
J. Neurophysiol. 112, 3138–3153 (2014).
© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.
25. Buschman, T.J. & Miller, E.K. Top-down versus bottom-up control of attention in
the prefrontal and posterior parietal cortices. Science 315, 1860–1862 (2007).
26. Schreiber, T. Measuring information transfer. Phys. Rev. Lett. 85, 461–464
(2000).
27. Wibral, M., Vicente, R. & Lizier, J.T. Directed Information Measures in Neuroscience
(Springer, 2014).
28. Bosman, C.A. et al. Attentional stimulus selection through selective synchronization
between monkey visual areas. Neuron 75, 875–888 (2012).
29. Grothe, I. et al. Attention selectively gates afferent signal transmission to area V4.
Preprint at https://doi.org/10.1101/019547 (2015).
30. Somers, D. & Kopell, N. Rapid synchronization through fast threshold modulation.
Biol. Cybern. 68, 393–407 (1993).
31. Mazzoni, A., Panzeri, S., Logothetis, N.K. & Brunel, N. Encoding of naturalistic
stimuli by local field potential spectra in networks of excitatory and inhibitory
neurons. PLoS Comput. Biol. 4, e1000239 (2008).
32. Kreiter, A.K. & Singer, W. Stimulus-dependent synchronization of neuronal responses
in the visual cortex of the awake macaque monkey. J. Neurosci. 16, 2381–2396
(1996).
33. Ecker, A.S. et al. Decorrelated neuronal firing in cortical microcircuits. Science
327, 584–587 (2010).
nature NEUROSCIENCE  advance online publication 
a r t ic l e s
34. Canolty, R.T. et al. Oscillatory phase coupling coordinates anatomically dispersed
functional cell assemblies. Proc. Natl. Acad. Sci. USA 107, 17356–17361 (2010).
35. Witt, A. et al. Controlling the oscillation phase through precisely timed closed-loop
optogenetic stimulation: a computational study. Front. Neural Circuits 7, 49
(2013).
36. Tiesinga, P.H. & Sejnowski, T.J. Mechanisms for phase shifting in cortical networks
and their role in communication through coherence. Front. Hum. Neurosci. 4, 196
(2010).
37. Battaglia, D., Brunel, N. & Hansel, D. Temporal decorrelation of collective oscillations
in neural networks with local inhibition and long-range excitation. Phys. Rev. Lett.
99, 238106 (2007).
38. Burkhalter, A. Many specialists for suppressing cortical excitation. Front. Neurosci.
2, 155–167 (2008).
39. Battaglia, D., Witt, A., Wolf, F. & Geisel, T. Dynamic effective connectivity of inter-
areal brain circuits. PLoS Comput. Biol. 8, e1002438 (2012).
40. Dotson, N.M., Salazar, R.F. & Gray, C.M. Frontoparietal correlation dynamics reveal
interplay between integration and segregation during visual working memory.
J. Neurosci. 34, 13600–13613 (2014).
41. Osborne, L.C., Palmer, S.E., Lisberger, S.G. & Bialek, W. The neural basis for
combinatorial coding in a cortical population response. J. Neurosci. 28,
13522–13531 (2008).
42. Rigotti, M. et al. The importance of mixed selectivity in complex cognitive tasks.
Nature 497, 585–590 (2013).
43. Harris, K.D., Csicsvari, J., Hirase, H., Dragoi, G. & Buzsáki, G. Organization of cell
assemblies in the hippocampus. Nature 424, 552–556 (2003).
44. Buzsáki, G. Neural syntax: cell assemblies, synapsembles, and readers. Neuron 68,
362–385 (2010).
45. Salazar, R.F., Dotson, N.M., Bressler, S.L. & Gray, C.M. Content-specific fronto-
parietal synchronization during visual working memory. Science 338, 1097–1100
(2012).
46. Lisman, J. The theta/gamma discrete phase code occuring during the hippocampal
phase precession may be a more general brain coding scheme. Hippocampus 15,
913–922 (2005).
47. Colgin, L.L. et al. Frequency of gamma oscillations routes flow of information in
the hippocampus. Nature 462, 353–357 (2009).
48. Cannon, J. et al. Neurosystems: brain rhythms and cognitive processing. Eur. J.
Neurosci. 39, 705–719 (2014).
49. Hipp, J.F., Hawellek, D.J., Corbetta, M., Siegel, M. & Engel, A.K. Large-scale cortical
correlation structure of spontaneous oscillatory activity. Nat. Neurosci. 15, 884–890
(2012).
50. Bastos, A.M. et al. Visual areas exert feedforward and feedback influences through
distinct frequency channels. Neuron 85, 390–401 (2015).
 ONLINE METHODS
Network models. The networks used to model local brain regions consisted of
1,000 excitatory and 250 inhibitory conductance-based Wang-Buzsáki neurons51
with random connectivity. Synaptic conductances had a difference-of-exponen-
tials time-course, proportional to a postsynaptic channel open probability Ps:
 − t − t0 − d −
t − t0 − d 
 t1 t2  (1)
Ps = B e −e  (1)
where t0 is the time of the presynaptic spike, d is a combined axonal and synaptic
transmission delay, τ1 and τ2 define the rise and decay synaptic time-constants
(τrise = τ1τ2/(τ1 − τ2) and τfall = τ1), respectively, and B is a normalization fac-
tor. Excitatory (E) or inhibitory (I) synaptic currents were then given by
Is(E,I) = gs(E,I)Ps(E,I)(Vs(E,I) − V), where gs(E,I) is the peak conductance of the con-
sidered synapse and Vs(E,I) is the synaptic reversal potential. We used VsE = 0 mV
and VsI = −80 mV for excitatory and inhibitory synapses, respectively.
Both peak synaptic conductances and synaptic delays were Gaussian-distrib-
uted random variables with prescribed means gs(E,I) and d and s.d. sg ( E , I ) and
© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.
s
σd. The probability of establishing an excitatory (or inhibitory) connection within
each modeled brain region is denoted by PE (or PI). The background drive to each
neuron in the network was provided by external excitatory inputs, modeled as
Poisson processes with rate νin and statistically independent for different neurons.
Each spike injected by this external driving source induced in the target neuron a
synaptic current as in equation (1). Peak conductances of the external drive were
also heterogeneous and assumed values sampled from a truncated Gaussian dis-
n
tribution with mean g s and s.d. sg n . Long range synaptic connectivity between
s
two or three brain regions was purely excitatory, targeting both excitatory and
inhibitory neurons in the target region. In the two-regions network architec-
ture, the probability of having a long-range excitatory projection, PELR, was held
equal between regions. In the three-regions network architecture, the region
Y (at the highest hierarchical level) received feedforward connections from X
and Z with a probability PE ff that was stronger than or equal to probability PE fb
of feedback connections from Y to both X and Z. Both feedforward (and
feedback) connections had equal conductance strength. No connection was
present between X and Z. Long-range excitatory synaptic conductance values
were drawn from the same distribution as for recurrent excitation within each
local region.
The default set of default network parameters is summarized in Table 1. These
default values were used, unless specified otherwise (specific parameter variations
are listed in the following). Simulations were performed with PI = 0.3 in Figures 1,
2 and 6 and Supplementary Figure 17; PI = 0.6 in Supplementary Figure 4; and
PI = 0.1 in Supplementary Figures 12 and 14f–h. The probability of establish-
ing a local excitatory connection deviated from default values in Figures 4 and 5
and Supplementary Figures 15–17 with PE = 0.35; in Supplementary Figures 7
and 14 with PE = 0.25; and in Supplementary Figure 12 with PE = 0. In Figure 6,
we also adopted a smaller degree of input conductance heterogeneity, sg n = 0.1
s
nS; in Supplementary Figures 9, 10, 13 and 16 we adopted sg n = 0.5 nS; and
s
in Supplementary Figures 15 and 17 we adopted sg n = 0 nS. Large values of
s
σg ν might lead to symmetry-breaking network instances, in which one specific
s
phase-locking configuration is favored over of the other even when there is no bias
applied. For clarity of illustration, these symmetry-breaking instances were excluded
from all analyses. Typical simulation runs were performed with a fourth-order
Runge-Kutta method when the input was a smooth random field (fixed time-step
∆t = 0.05 ms or simple Poisson spike-trains (fixed time-step ∆t = 0.1 ms), and lasted
10 min unless otherwise indicated (details found in the corresponding section).
When the input was modeled as an Ornstein-Uhlenbeck process, an Euler-
Maruyama integration scheme with ∆t = 0.01 ms was used instead. To sample
neuronal activity distributions, we combined multiple simulations with quenched
heterogeneity and connectivity and different realizations of the stochastic drive to
the system. Generally, no statistical methods were used to predetermine sample sizes,
but our sample sizes are similar to those reported in previous publications27,39.
Characterization of network activity. The synchronization index was chosen
to be the previously introduced χ2, where
NsV2 (t )
c2 = N W(t) centered at t, emitted by neurons in subset X of K neurons belonging to
∑ sV2 (t )
i =1 i the region X, we define θkX(t) as the phases of the LFP of area X at spike-times
nature NEUROSCIENCE
Here
N
∑ i = 1Vi (t )
V (t ) =
N
is the LFP-like signal52,53. The variables s V (t ) and s Vi (t ) denote the s.d. of LFP
amplitudes over time or, respectively, of the membrane potential traces Vi(t) of
each individual neuron i. The χ2 coefficient is bounded to the unit interval, with
vanishing values indicating asynchronous dynamics. The frequency of the col-
lective oscillation of a region is calculated as the inverse of the first peak position
of the autocorrelation function of a multiunit rate signal, obtained by convolving
the raster plot of each local network with a Gaussian kernel of s.d. of 2 ms. The
rate signal in Figure 1b was calculated as a histogram with a 1-ms bin width.
Single neuron rates were calculated as the inverse of their mean interspike inter-
val. Spectrograms of LFP activity (Figs. 1e and 2c and Supplementary Figs. 4a
and 13b) were calculated using a standard time–frequency Fourier analysis with
overlapping Hamming windows (300-ms size, 250-ms overlap) and the time-
averaged spectra through the multitaper method (16 tapers).
Analyses of gamma-burst properties and coordination. To extract the joint
distribution of frequencies and durations of oscillatory burst (Fig. 1f and
Supplementary Fig. 4b), we thresholded LFP spectrograms (300-ms size, 280-ms
overlap) at the 95th percentile of instantaneous power values. The thresholded
spectrograms were then scanned at different frequencies (i.e., line by line) to
extract the durations of time intervals over which power at the given frequency
was continuously sustained above threshold. We sampled these burst durations
by combining 20 trials (overall, 100 min of signal). To study time and frequency
coordination between oscillatory bursts occurring in different coupled regions
(Fig. 2c–d), LFP spectrograms of each region were thresholded above their 95th
percentile and scanned line by line, creating a binary string for each popula-
tion and for each frequency z X ,Y ( f ) . These binary strings were then multiplied
against the analogous strings obtained from the other population and normalized
by the s.d. of the binary strings to obtain the overlap
z ( f ) ⋅z X ( f )
J( f ) = Y
s z X s zY
The calculation was made using 50 trials for Figure 2c and 10 trials for
Supplementary Figures 4 and 13. Significance of the estimated overlap (Fig. 2c)
was assessed by comparing 95% c.i. for each overlap value with the associated
permutation-based chance levels (2,500 replicas).
Analysis of transient phase-locking. We calculated time-resolved cross-cov-
ariance (XC) adopting an overlapping sliding window 50 ms long (roughly two
average oscillation periods) and with 0.1-ms steps. Within each time-window,
standard cross-covariance was performed between chunks of LFPs of each of
the coupled populations:
XC(t , d ) = 〈(VX (t ′) − VX ) ⋅ (VY (t ′ − d ) − VY )〉t ′ ∈W (t ) (2) (2)
where 〈⋅〉t ′∈W (t ) denotes averaging over a 50-ms time-window W(t) centered on
time t, and VX and VY are LFP averages over this same window.
The instantaneous phase of each LFP signal ΦX and ΦY was calculated by inter-
polating a straight line over the unit phase interval 0 < Φ < 1 between consecutive
maxima of the population activity of a single repetition. A peak in population
activity was considered such if it was the maximum in a neighborhood of a half-
period radius of the central oscillation frequency of the population activity. Such
a fast approach avoids bandpass-filtering and leads to similar results as a Hilbert
transform analysis, as previously shown35. The relative phase is then defined by
∆Φ = (ΦY − ΦX, mod 1). The joint probability distribution of the time-dependent
phase difference ∆Φ and the time dependent peak of XC, XC*(t) = maxδ(t, δ), was
obtained from 100 simulated trials (overall, 1,000 min of signal). The conditional
probability distributions shown in Figure 3C were obtained by independently
normalizing each line of the probability matrix in Figure 3c. The XC* threshold
for Figure 3d was the 99th percentile (XC* = 0.37).
The circular variance of Supplementary Figure 6 was calculated as follows.
Given {sk}X with k = 1, …, K, the set of spike-times within a window of time
doi:10.1038/nn.4569
 θkX(t) = ΦX(sk). The local circular variance of the firing of the subset X, CVX(t),
is then the circular variance of the set of phases θkX(t). We then average circular
variances over the two coupled regions, i.e.,
1
CV (t ) =
2
(
CVX (t ) + CVY (t ) ) When considering the two-regions model in Figures 4d,e and 6 and
We evaluated these circular variances considering two different conditions.
In Supplementary Figure 5c–e, the subsets X and Y included all the neurons
belonging, respectively, to the regions X and Y. In the right panels of
Supplementary Figure 5d,e, these sets were then restricted to just ten ran-
domly chosen neurons within the regions. In both cases, the length of the sliding
windows W(t) was equal to that of seven average oscillation periods. We used
200 noise realizations with the same network architecture, lasting 1 min each,
to obtain the joint distribution of CV(t) and XC*(t). As before, the conditional
probability distributions shown in Supplementary Figure 6d,e were obtained by
normalizing each line of the probability matrix in Supplementary Figure 6d,e.
Analyses of information transfer. We evaluated TE26,27 between discrete time-
© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.
series of the simulated LFP signals VX(t) and VY(t). These signals are constrained
to take values from the set Ξ, and have distributions p(xt) and p(yt). The set Ξ
corresponds to the set of signal values observed within specific temporal windows
in which certain conditions, defining a consistent functional state, are properly
fulfilled (state-selecting filter; see below). TE from region X to region Y is defined
as a deviation from the Markov condition54,55: p( yt | yt(k ) , xt(l ) ) = p( yt | yt(k ) ).
This reads: the probability of VY taking a value yt at time t, given that the k past
values of VY were yt(k ) = ( yt − t1 , …, yt − t k ) , is independent from the l past
values of VX given by xt(l ). This condition is only fulfilled if there is no influence
from the past values of VX on the current values of VY ; in other words, only
if the distributions are conditionally independent. These probabilities are not
time dependent; they depend only on the time lags (τ1, …, τk). TE can then be
expressed as a measure of the divergence from the above Markov condition:
 p( yt | y(k ) , x(l ) ) 
TE X →Y (t1 , …, t k ) = ∑ p( yt , yt(k ) , xt(l ) )log 2  t t
yt ,…, yt − t k  p( y | y(k ) ) 
t t
∈Ξ
xt ,…, xt − t l
Here we relaxed the Markov condition and adopt a commonly used simpli­
fication39,56–62. We replaced the past activity history yt(k ) and xt(l ) in the previous
expression with unique observations of past values at a single latency at
time yt(k ) = yt − t and xt(k ) = xt − t , and then treated the interaction latency τ
as a variable parameter. With this approximation, the TE expression in
equation (3) can be simplified as
TE X →Y (t ) = H  yt | yt − t  − H  yt | xt − t , yt − t  (4)
   
TE represents the reduction in uncertainty (entropy, H) about the future state
of a given region Y when the history of region X is also taken into account. An
analogous (but not identical) expression for TE in the opposite Y-to-X direction
is obtained by exchanging the indices X and Y. In Supplementary Figure 18,
we compared the results obtained with the simplified expression of equation
(4) against estimations of TE considering different state-delays for sender and
receiver or considering more than a single state-delay.
When more than two regions constitute the interacting system, the above
expression for TE must be modified to account for the possibility of indirect
functional interactions (for example, transfer from X to Z via Y, etc.). We used
partial TE63 to uncover the direct interactions between pairs of populations when
embedded in a larger multiareal system. Given a set of three populations X, Y and
Z, partial TE (pTE) with a single state-delay is defined as
pTE X →Y (t ) = H[ yt | yt − t , zt − t ] − H[ yt | xt − t , yt − t , zt − t ] (5) (5)
This expression reduces to equation (4) when the activity of Z is statistically
independent from that of X and Y.
All the probability densities required for the evaluation of equations (4) and
(5) can be straightforwardly derived from the three-term or four-term joint prob-
ability distribution p( yt , xt − t , yt − t ) in the TE case or p( yt , xt − t , yt − t , zt − t )
in the partial TE case. These probabilities must be sampled uniquely over the
doi:10.1038/nn.4569
time instants in which specific conditions hold. TE and pTE are then evaluated
in a consistent regime, specified by the choice of the filtering conditions, Ξ.
Information-theoretical analysis with a specific filtering condition based on the
instantaneous interareal phase-difference ∆Φ returns the routing state.
Supplementary Figures 8a–c and 17, the ∆Φ range is split into five relative-
phase bins of equal size (the TE vs. τ and the ∆Φ surfaces were then smoothed
by linear interpolation). Figure 6 and Supplementary Figure 17 show the
anisotropy index from single trials. The filtering criteria in just two Top and
Bottom routing states (Fig. 4g–m and Supplementary Figures 7, 8d,e, 9–12,
14 and 17) were selected time instants in which the relative phase fell, respectively, in
the coarser ranges ∆Φ↑, i.e., 0 < ∆Φ < 0.5, for the Top state or ∆Φ↓, i.e., 0.5 < ∆Φ < 1,
for the Bottom state.
In Supplementary Figure 8, we adopted a state conditional to the strength
of interareal cross-covariance, in addition to the instantaneous phase relation.
Conditions on the instantaneous value of XC* were imposed arbitrarily, requiring
it to exceed a threshold XC*high = median(XC*) + 0.5 (Supplementary Fig. 8b,d)
or to be less than XC*low = median(XC*) − 0.5 (Supplementary Fig. 8b,e).
In the three-regions model in which a pairwise analysis was devised (Fig. 6
and Supplementary Fig. 17), the procedure was identical to that described above
for the two-regions model. More complex state-filtering conditions were devised
for complete analysis of the dynamics of the three regions. The six possible hierar-
chical configurations that the LFP signals can take define six filtering conditions,
∆Φ1XZY to ∆Φ6YZX , giving rise to different (or sometimes the same) routing
states, as illustrated in Figure 4d. The conditions can be written in terms of the
pairwise phase relations:
def
∆Φ1XZY = ∆Φ XY > 0.5, ∆Φ XZ > 0.5, ∆Φ XZ > ∆Φ XY
def
∆Φ2XYZ = ∆Φ XY > 0.5, ∆Φ XZ > 0.5, ∆Φ XZ < ∆Φ XY
def
∆Φ3ZXY = ∆ΦYZ < 0.5, ∆Φ XZ < 0.5, ∆Φ XZ < ∆ΦYZ (6)
 (3) def
∆Φ 4ZYX = ∆ΦYZ < 0.5, ∆Φ XZ < 0.5, ∆Φ XZ > ∆ΦYZ
def
(3) ∆ΦYXZ
5 = ∆Φ XY < 0.5, ∆Φ ZY < 0.5, ∆Φ XY < ∆Φ ZY
def
∆ΦYZX
6 = ∆Φ XY > 0.5, ∆Φ ZY > 0.5, ∆Φ XY > ∆Φ ZY (6)
For the two-regions model, we estimated histograms (within the two possible
Top and Bottom routing states) in Figure 4g,h and Supplementary Figure 8d–e
based on 500 min of simulated signal from 50 trials of the same network instance;
in Figures 4d,e and 6b and Supplementary Figures 8c and 17b based on
10 min of signal from a single network realization (the mean in Supplementary
(4) Fig. 8a is based on a total of 150 min from 15 different network instances); for
Supplementary Figures 9 and 10 based on single trial simulations, each lasting
10 min; in Supplementary Figure 11 based on 40 different network instances
(10 min of LFP signal for each considered parameter configuration); in
Supplementary Figure 12 based on 10-min simulations from 50 different network
instances (in Supplementary Fig. 7, only 10 network instances were used, with the
same simulation time); and in Figure 4j–m and Supplementary Figure 14 based
on simulations lasting 50 min each from 40 different network instances (2,000 min
of simulation time). In the three-regions model, we estimated histograms (within
each of the six state-filtering conditions) in Figure 5d and Supplementary Figure 15
based on simulations lasting 3,600 min each from 12 different network instances;
in Figure 5f and Supplementary Figure 16 based on 20 network instances last-
ing 4,000 min in total; and in Figure 6 and Supplementary Figure 17 based on
simulations lasting 10 min each from 25 different network realizations.
TE values obtained from the sampled histograms using the definition in equa-
tion (4) were then further corrected against finite-size bias by quadratic extrapola-
tion toward infinite sample size64. Given a set of time series of length L, we took
three subsets of the same length Lk, with Lk = L/k with k = 1, … , 10. For each
trimmed length Lk, TE is calculated from each of the three subsets and then aver-
aged over them to obtain a value TEk. Finally, we evaluated the limit to infinite
sample size for TE, extrapolating it as the vertical-axis intercept of a second-order
polynomial fit of TEk to k. The 95% c.i. for TE values were approximated as ±2× s.d.
above and below the mean, over the considered network instances or trials (95%
nature NEUROSCIENCE
 rule). These c.i. were then compared to the null hypothesis (lack of functional
coupling), providing a baseline TE range for significance testing. To obtain the
null-hypothesis distribution of TE, we relied on resampling techniques and used
bootstrapping with replacement to build a distribution of 1,500 surrogates for
network realization, and then collected them. To maintain the oscillatory nature
of the analyzed time-series, we built bootstrap surrogates using LFP blocks of a
mean length . This value should be long enough to guarantee that the signal’s
autocorrelation structure would be preserved when building bootstrap surro-
gates. To choose , we analyzed the behavior of the bootstrapped TE versus the
chunk length, . For oscillatory signals in the transient synchrony regime, the
value of the bootstrapped TE initially increases and then saturates at around  =
100 ms. We thus chose  = 300 ms for the TE baseline calculations, and we kept it
the same for the MI calculations, although the random inputs are not oscillatory
and therefore the values of the bootstrapped MI were independent of . After
building a null-hypothesis histogram of TE values, we determined the range
between its 2.5th and 97.5th percentiles and used it as the baseline TE range.
TE values whose 95% c.i. was completely disjoint and fell above (or below) this
baseline range were indicative of significantly enhanced (or damped) information
© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.
transfer with respect to chance level.
Finally, we defined the functional anisotropy index:
TEY → X (t ) − TE X → Y (t )
∆TE(t ) =
max[| TE X →Y (t ) |,| TEY → X (t ) |]
bounded in the −1 ≤ ∆TE ≤ 1 range, as a simpler indicator of the different effi-
ciencies of information transfer in the two possible reciprocal communication
directions between a pair of regions.
Modeling of the external input signals. Exogenous sources of information
were modeled with two different input models. In the first approach (Fig. 5 and
Supplementary Figs. 14 and 16), external inputs were modeled as scalar Gaussian
random field (GRF) input currents. GRFs were generated as follows. First, we
built a scalar random field in Fourier representation, given by
p ˆ
yˆ (k) = C(k)( Ak + i Bk ) (7) (7)
∆k
where Ak and Bk are independent Gaussian random variables with 0 mean and
unit variance. The correlation function Cˆ (k) in Fourier space was obtained by the
discrete Fourier transformation of a set of T points sampled from a continuously
smooth time-domain representation, C(τ) = 2/cosh(τ/τs). The T samples of C(τ)
were taken at regularly spaced times tn = n∆τ, n = 1, … , T, resulting in samples of
Cˆ (k) at k = k1, … , kT, evenly spaced with ∆k = 2π/∆τT. Finally, the real part of the
inverse Fourier transform of equation (7) provided the desired GRF variable.
The currents conveying the externally supplied inputs S(t) = gΨ(t) were gener-
ated independently for each area. Instances of S(t) were first generated in Matlab
and then fed as input currents to each neuron within the corresponding region.
We used the same fourth-order Runge-Kutta integrator (∆t = 0.05 ms) as described
above. The parameters used for the simulations of Figure 5 and Supplementary
Figures 14 and 16 were τs = 1 ms, with a gain factor of g = 2. Qualitatively similar
results were obtained over the entire range 0.5 ms ≤ τs ≤ 5 ms.
In Supplementary Figure 14e–h, we used a different noise model, modeling
inputs as Ornstein-Uhlenbeck processes. The input signal was given by dS(t)/dt =
−γS(t) + gη(t), where η is a Gaussian white noise with 0 mean and unit variance.
In this figure, we used g = 0.1 and γ = 1. An Euler-Mayurama method was used to
simulate the input and the network dynamics, with a time step ∆t = 0.01 ms.
Information sharing between the two endpoints of a communication
line was assessed via the classical measure MI65. MI between an injected
nature NEUROSCIENCE
signal SX to the source region X and LFP from a target region Y can be state-
filtered as TE by pooling only input- and response-value pairs within the epochs
fulfilling the filtering criteria specified by the chosen filter, Ξ, between the
LFP signals VX and VY:
 p( yt , st − t ) 
MISX ↔Y (t ) = ∑ p( yt , st − t )log 2  
st , yt ∈ Ξ  p( yt ) p(st − t ) 
When analyzing three-region structural motifs in Figure 5 and Supplementary
Figure 16, for TE, we computed partial (also called conditional) MI65,66, intro-
ducing an additional condition on the third region signal.
Statistics. Generally, we used nonparametric statistical testing procedures based
on comparisons between confidence intervals obtained by bootstrapping with
replacement (for both the association and the null hypotheses). See details in
each specific Online Methods subsection. A Supplementary Methods Checklist
is available.
Code availability. Custom-written code for model simulations and their analyses
are available on reasonable request to the corresponding authors.
Data availability. All simulations analyzed during the study can be regenerated
using the above code.
51. Wang, X.J. & Buzsáki, G. Gamma oscillation by synaptic inhibition in a hippocampal
interneuronal network model. J. Neurosci. 16, 6402–6413 (1996).
52. Golomb, D. & Rinzel, J. Clustering in globally coupled inhibitory neurons. Physica
D 72, 259–282 (1994).
53. Golomb, D. & Hansel, D. The number of synaptic inputs and the synchrony of large,
sparse neuronal networks. Neural Comput. 12, 1095–1139 (2000).
54. Vicente, R., Wibral, M., Lindner, M. & Pipa, G. Transfer entropy--a model-free
measure of effective connectivity for the neurosciences. J. Comput. Neurosci. 30,
45–67 (2011).
55. Wibral, M. et al. Measuring information-transfer delays. PLoS One 8, e55809
(2013).
56. Gourévitch, B. & Eggermont, J.J. Evaluating information transfer between auditory
cortical neurons. J. Neurophysiol. 97, 2533–2543 (2007).
57. Honey, C.J., Kötter, R., Breakspear, M. & Sporns, O. Network structure of cerebral
cortex shapes functional connectivity on multiple time scales. Proc. Natl. Acad.
Sci. USA 104, 10240–10245 (2007).
58. Lungarella, M., Pitti, A. & Kuniyoshi, Y. Information transfer at multiple scales.
Phys. Rev. E 76, 056117 (2007).
59. Garofalo, M., Nieus, T., Massobrio, P. & Martinoia, S. Evaluation of the performance
of information theory-based methods and cross-correlation to estimate the functional
connectivity in cortical networks. PLoS One 4, e6482 (2009).
60. Ito, S. et al. Extending transfer entropy improves identification of effective
connectivity in a spiking cortical network model. PLoS One 6, e27431
(2011).
61. Stetter, O., Battaglia, D., Soriano, J. & Geisel, T. Model-free reconstruction of
excitatory neuronal connectivity from calcium imaging signals. PLoS Comput. Biol.
8, e1002653 (2012).
62. Orlandi, J.G., Stetter, O., Soriano, J., Geisel, T. & Battaglia, D. Transfer entropy
reconstruction and labeling of neuronal connections from simulated calcium
imaging. PLoS One 9, e98842 (2014).
63. Vakorin, V.A., Krakovska, O.A. & McIntosh, A.R. Confounding effects of indirect
connections on causality estimation. J. Neurosci. Methods 184, 152–160
(2009).
64. Strong, S., Koberle, R., de Ruyter van Steveninck, R. & Bialek, W. Entropy and
information in neural spike trains. Phys. Rev. Lett. 80, 197 (1998).
65. Cover, T.M. & Thomas, J.A. Elements of Information Theory 2nd edn. (Wiley-
Interscience, 2006).
66. Frenzel, S. & Pompe, B. Partial mutual information for coupling analysis of
multivariate time series. Phys. Rev. Lett. 99, 204101 (2007).
doi:10.1038/nn.4569