The Inuence of Intrinsic Motivation on Decisions Between Actions

Extensive research explains how pre-frontal cortical areas process explicit rewards, and how pre-motor and motor cortices are recipients of that processing to energize motor behaviour. However, the specics of motor behaviour, decisions between actions and brain dynamics when driven by no explicit reward, remain poorly understood. Are patterns of decision and motor control altered wen performing under social pressure? Are the same brain regions that typically process explicit rewards also involved in this expression of motivation? To answer these questions, we designed a novel task of decision-making between precision reaches and manipulated motivation by means of social pressure, dened by the presence or absence of virtual partner of a higher/lower aiming skill than our participants. We assessed the overall inuence of this manipulation by analysing movements, decisions, pupil dilation and electro-encephalography. We show that the presence of a partner consistently increased aiming accuracy along with pupil diameter, furthermore the more skilled the partner. Remarkably, increased accuracy is attained by faster movements, consistently with a vigour effect that breaches speed-accuracy trade-offs typical of motor adaptation. This implicated an ensemble of cortical sources including pre-frontal areas, concerned with the processing of reward, but also pre-motor and occipital sources, consistent with the nature of the task. Overall, these results strongly suggest the role of social pressure as a motivational drive, enabling an increase of both vigour and accuracy in a non-trivial fashion.


Introduction
Motivation has been often studied within paradigms in which rewarding stimuli elicit subsequent behaviour, assessing motivation through proxies of physiological variables [1][2][3] . For example, the effect of explicit reward on action invigoration 4,5 , preparatory attention, working memory and decision-making 6-9 , identi ed multiple pre-frontal cortical regions that encode speci c aspects of reward processing 3,10-13 .
However, anecdotal evidence suggests that, beyond explicit rewards, the brain's motivational systems evolved to deal with situations in which reward could be either non-explicit or just absent, e.g., animals tends to explore their environment during minutes or hours driven by motivation alone, in the prospect of nding food. Consistent with this, earlier theories de ned motivation as a substance, held back in a container, capable of energizing behaviour, and subsequently released in action 14,15 . In other words, motivation is viewed as a mid eld physiological actor, sensitive to reward and subliminal priming 16 , which releases through behaviour. Fifty years after Bindra's de nition, our understanding of the processes underlying motivation in the absence of external reward remain elusive. Despite motivation's covert nature, physiological proxies have shown that motivation decreases when reward is obtained, contingent on performance 17,18 , and that it increases with behavioural activation 19 , encompassed by connectivity modulations between the right medial frontal gyrus and the right posterior cingulate cortex. Likewise, a differential manipulation of extrinsic reward and intrinsic motivation in an imaginary task identi ed the angular gyrus 20 and the anterior insular cortex 21 , as their respective regions of interest 22 . In summary, although we are progressively gathering a remarkable amount of information about motivational in (requiring the reaching movement to stop at the target), and cross-over (hitting and crossing over the target) 28 ---FIG 1C. In both cases, our hypothesis is that social pressure motivates our participants to increase precision, and to consequently adjust their decisions and principles of selection of motor parameters (see conceptual framework described in FIG 1E). We recorded movement kinematics and pattern of decision from movement data, ocular and electro-encephalographic data.

Kinematic Correlates of Motivation
The purpose of the two control regimes is having two conditions in which the requirement of control as a function of motor noise 29,30,31 . In both cases, we predicted that our modulation of motivation could act in one of two ways: either participants were sensitive to social pressure per se and reduced their aiming error regardless of relative aiming skill, or they were concerned by the social ranking and reduced their accuracy only when the partner's accuracy was better than the participant's. In either case, if the adaptive strategy were consistent with a speed-accuracy trade-off, the error reduction should extend deceleration times during stop-in trials and peak velocities during cross-over ones. In other words, increased precision during stop-in trials should correlate with longer movements and weaker late movement kinematics, from Peak Velocity (PV) to the end of the movement at target arrival. By contrast, cross-over movements enter the target around PV, which make any strategy to control precision contingent on modulating early kinematic markers, from the onset of movement to PV. Detailed analyses for each control regime are shown next (FIG 2 & Suppl. FIG 1).
We for the effect size in addition to the p-value 32 , by comparison with a ten-thousand sample surrogate distribution for each regression coe cient (FIG 2A-B). This surrogate distribution was obtained by tting the GLM to shu ed kinematic marker values across trials per participant, which were then pooled across all participants.
Consistent with prediction, the arrival error diminished with increasing motivation (p=0.00071; z=-3.38), encompassed by a reduction of PV (p=0.017; z=-2.37) and PD (p=0.019; z=2. 34), and of TTPV (p=0.0012; z=-3.21) and DT (p=0.028; z=-2. 19) intervals (FIG 2B). Also, accuracy co-varied with velocity and movement duration when performing Solo. However, the joint PV & TTPV reduction for motivated states (M=1-Easy, 2-Hard) did not exhibit any signi cant correlation with error (FIG 2E), thus violating the speedaccuracy trade-off when performing with a partner. In other words, although both the error and PV diminished within motivated states 1 and 2, they were uncorrelated. Along with Motivation, the error also diminished as the experimental session unfolded, with #Block (p=0.00067; z=-3.40; FIG 2F) 1H). Previous evidence suggested two relevant behavioral differences with respect to stop-in: rst, target arrival when crossing-over is necessarily more di cult, as this occurs around peak velocity, the time of maximum vigor and motor noise of the movement 29 ; second, the piece of the movement beyond target crossing is inconsequential to arrive to a precise target location. Our hypothesis was that a lower social categorization (being ranked second best in motivated states 1 or 2) positively motivates the participant to increase precision by decreasing the intensity of the earlier two metrics of movement: Peak  1E). In other words, if the control requirements were low (stop-in), an error reduction could be attained regardless of kinematic variables. However, as the requirement of control increased (cross-over), a multivariate adaptive strategy became necessary, thus leading to a speed-accuracy trade-off adjustment.

Decisional Correlates of Motivation
In line with the analyses of movement, our prediction was that social pressure would equally in uence decisions between actions implying opposite biomechanical cost 33 1C).

Oculometric Correlates of Motivation
We monitored pupil diameter to assess whether our manipulation of motivation could be traced independently at a physiological level [34][35][36][37][38] . To test this, we pulled together the pupil diameter data across both regimes (stop-in & cross-over), z-scored each participant's recording, and averaged across two 1s intervals of interest, preceding two task events: the time the rst origin cue was shown (tShowOrigin; tSO) and the GO signal (tGO) ---see Methods, traces of two subjects in FIG 4A. We then quanti ed their dependency with respect to motivated state, #Block, #Trial, motor cost and their interactions with our GLM analysis (see METHODS). FIGs 4B-C report the resulting GLM group average and standard error regression coe cients for both intervals of interest. Two main group effects were reported: a signi cant pupil diameter increase with Motivation (F-test, F(11,1)=3.25, p=0.027 around tSO / F(11,1)=4.51, p=0.018 around tGO), and a reduction, possibly associated to fatigue, as the #Block increased (M×#B interaction, signi cant only around tGO, F(11,1)=2.95, p=0.032). Furthermore, we also assessed statistical signi cance by reiterating our GLM t at intervals of 30ms, during both intervals of interest, and recalculating a sliding t-test across group GLM regression coe cients. Remarkably, these yielded p-values < 0.05 over both intervals of interest (FIG 4D-E), demonstrating a consistent, baseline increase of pupil diameter caused by social pressure, which extended over most of the trial duration.

Electro-Encephalographic (EEG) Correlates of Motivation
Furthermore to assessing the in uence of social pressure on motor decisions and pupilometry, our analyses of electro-encephalographic data aimed to identify the cortical network responsible for the changes of behaviour we reported. In other words, to asses the network responsible for Motivation as de ned in this study, we adapted techniques previously used in fMRI studies 22 to our EEG recordings. We focused on the EEG activity recorded at an interval of 1.2s preceding movement onset for each trial, as to prevent movement artefacts and to focus on baseline effects during single block performance of brain activity ---keep in mind that Motivation and the control regime were the only constant experimental factors within block. The data was pre-processed by means of an ad-hoc pipeline for de-noising, participant-wise normalization across both sessions, and band-passing into the typical EEG a, b and g frequency bands ---see METHODS. Our analyses combined two complementary levels of description: local electrode power activity and functional connectivity as a proxy for pairwise interactions in the electrode network and quanti ed by correlations (equivalent to spectral coherence within each band).
Our rst set of analyses assessed whether and how three levels of Motivation (0-Solo, 1-Easy, 2-Hard) correspond to distinct brain activity states, as quanti ed by the electrode signals (FIG 5A-B). This we tested by a decoding scheme based on two usual machine-learning classi ers: multi-layer logistic regression (MLR) and k-Nearest Neighbour (1NN with k=1). For classi cation, we used the two sets of local/network features mentioned above (electrode power and pairwise electrode correlations) within each frequency band. Supplemental FIG 2A-B shows the classi cation accuracies obtained for a typical participant (Subject #2) and for the group, color-coded as a function of the frequency band. Remarkably, its accuracy increases from 65% in the a-band to a 95% in the g-band, when using the matrix of electrode power to classify. This held across participants, with remarkable consistency also for correlations, with a success rate from 50% in the a-band, to 85% in the g-band when using electrode pair-wise correlation as feature. On all accounts, the MLR outperformed the 1NN classi er by 5-10%, indicating that the activity states were better characterized by changes in speci c features rather than by global changes of activity.
A more detailed account of the state-by-state classi cation is also provided by the confusion matrices (Suppl . FIG 2A-B), which con rms the identi ability of three distinct motivated states for the higher frequency β-and g-bands, which is consistent across participants and local/network metrics.
We also identi ed the informative electrodes that support this robust classi cation, corresponding to electrodes whose power or electrode pairs whose correlation varied most signi cantly across the three motivated states. In this fashion, we gained access to the electrode network responsible for modulating motivation in our task. We relied on the recursive feature elimination (RFE) algorithm, which provides a The classi cation procedure in the source space was analogous to the one just described, transposed to the source signals obtained from using independent component analysis on the pre-processed data 39,40 . We used a generic pattern to estimate the electrode anatomical locations provided for the 60-electrode Acticap con guration to localize each source. Sources were obtained for each participant by pulling together both sessions and control regimes. Note that the source space had a reduced dimension of 40-50 after eliminating noisy and artefactual sources.
Reinforcing our results at the electrode space, the level of Motivation could also be decoded in source space for each participant and at the group level (FIG 5C-D). Again, accuracy increased with frequency, reaching 95% for the g-band, and ranked similarly regardless of metric (source power or correlation) per participant. Overall, the decoding performance and identi ability of the motivated states (see confusion matrices) were similar both in electrode and source spaces.
To identify the top relevant sources contributing to the classi cation, we performed RFE on the power sources. In brief, eight sources su ced to attain a classi cation plateau of about 95% for all participants in the g-band (see for example Suppl. FIG 3). From those eight sources per participant, we identi ed the common sources that generalize across participants, pooling all sessions and control regimes across participants (two sessions per subject, two control regimes).
We used the Girvan-Newman community detection procedure based on spatial similarity on the cortical surface (see METHODS). This resulted in six common sources (communities of best sources), whose average spatial pro les are displayed in FIG 6A for each frequency band. In essence, these head maps correspond to the centroid of sources in each session belonging to the same community. We obtained sources in the prefrontal / premotor/ occipital cortices for a, b and g bands (FIG 6A). This broad distribution of sources related to motivation contrasts with the results of RFE applied to electrode power in Suppl. Figure 2D, which mainly identi es prefrontal electrodes. This also suggests a more complex network of processing involving not only prefrontal related areas, related to reward processing, but also premotor and visual cortical areas, consistent with the speci cs of our experimental design.
Lastly, we veri ed whether the interactions across these common sources were modi ed by Motivation.
To that end, we recalculated the classi cation performance using as features the correlations between the common sources per session alone. This is equivalent to selecting within the correlation matrices, the elements corresponding to the common sources obtained using community detection, akin to a massive subsampling. It is worth noting that a good performance was not guaranteed here because the best sources were identi ed on the grounds of their power alone, regardless of their cross-correlation (see the chance level performance of interactions within the alpha band in FIG 6B). Nevertheless, the accuracy turned out to be signi cantly above chance, for both b and g bands (FIG 6B). These results thus strongly suggest that these common sources also experienced network communication across distant brain areas induced by Motivation, highlighting the roles of b and g rhythms for this cognitive function.

Discussion
Here we studied how social pressure in uences movements and decisions between them 41,42 , and analyzed the network of cortical sources supporting this process. We introduce a novel experimental protocol of decision-making between precision reaches using social pressure to manipulate intrinsic motivation in a controlled fashion, while bypassing the use of explicit external reward. Our participants performed either alone, or in the presence of virtual partners, capable of higher/lower accuracies than theirs. Participants were instructed to focus on their own performance and to disregard their partner's. Our results reported increased aiming accuracy (reduced error) whenever a partner was present, regardless of their skill. However, the participant's accuracy increased further when the partner was more skilled than her/him. At a physiological level, accuracy improvements were encompassed by an increase of pupil diameter. We quanti ed the in uence of social pressure on movement under two regimes of control: stopping at the target (stop-in) or hitting and crossing over it (cross-over). Remarkably, our results show that, under these conditions, improved accuracy often favoured movement intensity, thus breaching the speed-accuracy trade-off typical of motor adaptation. We also used the participants' electroencephalograms (EEGs) to characterize changes on the cortical network for each level of social pressure.
In brief, we examined patterns of power activity and functional connectivity within the a, b and g bands for each participant, both at the electrode and source levels. Our results show that the MLR classi er could robustly categorize trials from each participant according to their Motivation state with a 95% accuracy in the g-band, both in the electrode and source spaces. We identi ed six common cortical sources that generalized across participants, mainly localised in the frontal, motor and visual cortices.
Notably, the interaction across these common sources, achieved over 90% categorization accuracy in the g-band and 70% in the b-band, indicating motivation speci c communication across these cortices. In summary, these behavioural, physiological and neuronal results point towards a transversal physiological effect, that is, an internal physiological drive that in uences decisions and motor executions in the absence of explicit external reward.
From social pressure to intrinsic motivation and motor control Social pressure is well-known to in uence behaviour, by seeking outcomes that avoid being penalized 43 and achieve social success [24,44]. In our experimental design, each participant subjectively experiences success or failure at each trial across two full sessions, reporting two fundamentally different behavioural responses: when playing Solo, and when playing alongside a partner. Even under a condition of explicitly discouraged competition, and even if the partner is virtual, our experimental results (error reduction, peak velocity modulation, pupil diameter magni cation, identi ability of separate brain states) are consistent with our participants experiencing some concern for performing worse than their partner, which yields better accuracies whenever a partner is present. However, the behavioural component of this adaptation can be along one of two strategies, always aimed at attaining that bit of accuracy that would equal or outperform the partner: a speed-accuracy trade-off and a cognitively controlled strategy.
In other words, if the goal of each movement is success, de ned as attaining above average accuracy or a lesser error than the partner, the dynamics of a speed-accuracy trade-off adaptation dictate that going slower should increase your chances of reducing your aiming error 30 . However, our results show that this occurs only when playing Solo in the stop-in regime and essentially in all conditions for the cross-over regime. Our interpretation is that the participants improve accuracy by slowing down when there is no external agent (Solo; Stop-in), and when the di culty to improve precision is extreme due to signal dependent noise (Cross-Over). However, when there is social pressure and the di culty to increase precision are moderate (Easy, Hard; Stop-in), the general trend is that of increasing precision and vigour simultaneously. Although further research will have to be conducted to ascertain this, this range of behavioural responses to social pressure, as a function of motor di culty, suggests that adaptation within contexts of moderate di culty ---those in which motor noise is low, hinges on a cost-bene t rather than speed-accuracy trade-off 4 , which may be cognitively controlled precisely because motor noise is within manageable boundaries. Under these conditions, the potential advantage is that of adapting by varying a single motor dimension, either amplitude or time, in a fashion independent of each other 45 . By contrast, as noise increases beyond some threshold, the only possible option to increase accuracy is to trade-off speed to increase accuracy.
One of the major rigors to claim that social pressure is a motivating factor is the requirement of attaining either a direct metric of its underlying physiological dynamics, or at least of a proxy thereof 2 . We argue that this proxy may be pupil diameter, which has been previously reported to co-vary, as a function of the experimental setup, with physical and mental effort, with arousal [35][36][37][38] , and also with coping or with success 46 . The reason is that, in our experimental design, the motivating effect should transpire from the subliminal intention of 'making it' whenever accompanied by a partner, an effect that spills into the motivation related vigour effect and error reduction, and is consistent with the increase of pupil diameter we report. Our analyses of pupil diameter extend throughout the entire trial duration. Remarkably, our results show that pupil dilation positively co-varied with our manipulation of social pressure during entire trials, and for as long that speci c block of trials (and exposure to the same type of partner) extended. In brief, this con rms the modulation of internal motivation regardless of movement metrics, and that an increase of social pressure is also consistent with a systemic rise of arousal. Beyond this motivational effect, we also observed an effect of fatigue on pupil diameter, which exerted an opposing depressing in uence over time.

Motivated Motor control and Decision-Making
In our experiment, increased social pressure yielded a consistent improvement of precision. While this be ts the participant's intent to outperform the partner, his/her adaptive strategy varied as a function of the context. As for the case of subliminal reward processing, we observed movement invigoration when the partner was less skilled than our participant during the stop-in regime alone. By contrast, if the partner was more skilled than the participant or during cross-over trials in all cases ---arriving to target around peak velocity, subjects switched to the more cautionary strategy of weakening movement intensity and extending the movement, possibly to control the in uence of signal-dependent noise on variability 29 and to ensure higher accuracies.
This introduces two main implications: rst, although a low-skilled partner during stop-in trials exerts social pressure, it should not be that concerning, as the range of accuracies is such that participants may simultaneously increase velocity and accuracy. Importantly, these simultaneous changes are at odds with the well-known speed-accuracy trade-off typical of the goal directed movements 30,[47][48][49][50] , suggesting that the dynamics of motivation may be, to some extent, cognitively reduced if properly motivated, and if the participant's skill allows 51 . Second, this effect is absent during the cross-over regime, possibly because of the intrinsic di culty of controlling precision during ballistic, which overbears the capacity of cognitively controlling variability. The consequence is a more cautionary behaviour, consisting of longer time-to-peakvelocities (until target contact), and reduced velocity intensities when presented with a more skilled partner.
Moreover, consistently with the notion that decisions between actions abide by the same principles of motor control 33,52,53 , our analyses of the pattern of decisions between movements yielded two main effects related to effort and motivation. First, decisions between movements were consistent with the participant's sensitivity to motor cost, typically preferring low cost movements in the Solo condition 26,27 . Second, motivation decreased the participant's sensitivity to motor cost and increased the likelihood of selecting high cost movements as the demand for end-point precision increased, consistently with the need of nding a path where to be more accurate 28 . In light of cost-bene t assessments 54 , it could be interpreted that increasing accuracy is an abstract form of reward, that may be traded-off with the larger energy expenditure required of more stable movement paths. Future research needs to be conducted to clarify this point.

The cortical network of intrinsic motivation
In addition to the modulation of pupil diameter, we also analysed the participant's EEGs to characterize the activity and the communication across the related brain sub-networks typical of each level of social pressure. Our EEG analyses focused on a time interval preceding the presentation of both movement options and the decision-making phase. Consistent with this, we did not nd the usual a-band occipital modulation typical of sensory tasks increasing attentional load 55  As to analyze the structure of the network differences from baseline onto motivated states, we performed an analysis of source contribution to each accuracy calculation (RFE analysis across subjects), revealing that several distant cortical regions (FIG 6), contributed to this network, exhibiting similar spatial patterns across frequency bands with various levels of intensity. Consistently with previous studies, we found prefrontal cortical sources responsible for the modulation of motivation, suggesting the engagement of assessment processes and some of the neural circuitry responsible for reward processing 13,62,63 . Likewise, we identi ed sources in motor cortical sources (located in central regions), consistent with the nature of our task. Finally, the involvement of sources within the parietal and visual cortices, points at multi-modal integration, an introspective state 64 , and the baseline modulation on visual attention. Together, our results hint at a distributed cortical network effect engaged by intrinsic motivation, which exceeds pre-frontal localization often associated with the encoding of reward and also involves cortical areas naturally associated in our visuo-motor task.
Limitations and Future Work Our brain network interpretation is necessarily constrained by the limitations of EEG recordings, as more detailed head models with higher EEG density would be required for a ner spatial source localization 65 .
Furthermore, the fact that we found physiological correlates of motivation on pupil dilation and that the g-band is the one the most affected by motivation (and not in the a-band), would suggest that a process other than attentional load is responsible. Further experimental work should be conducted to con rm this. Finally, our work is constrained by a model-free analysis of connectivity, which could be extended by a more realistic bio-physical modelling of brain dynamics and communication, as recently achieved for fMRI signal analyses 66 . This would also allow to interpret cross-band communication in a uni ed framework. Ultimately, our goal was to entirely disentangle motivation speci c from reward speci c cortical communication. In addition to the proper analytical tools, this would require a novel experimental context where neural activity may be compared between two experimental conditions, one where explicit reward and motivation were present, and another where motivation alone drives behaviour. We leave this for a future study.

Conclusion
We showed that motivation may be manipulated by social pressure in the absence of an explicit external reward. Furthermore, this modulation is re ected in neuro-physiological signals that leads the way to the design of sophisticated future experimental work aimed at the careful disentanglement of the neural dynamics of motivation, reward and attentional processes in a quantitative manner. away from the participant's sitting position, we placed a vertically-oriented, 24" Acer G245HQ computer screen (1920x1080). This monitor was connected to an Intel i5 (3.20GHz, 64-bit OS, 4 GB RAM) computer that ran custom-made scripts which controlled task ow, programmed using OpenFrameworks v.0.9.8 software. The screen was used to show the geometrical arrangements and related stimuli on each trial. A small cross (1x1cm), whose position was synchronized with the planar coordinates of the end-point as it slid along the horizontal plane (table), was used to show the participant's corresponding movement in the vertical plane on the screen.

Materials And Methods
As part of the experiment, subjects had to respond by performing overt movements with their arm along the table plane. Their movements were recorded with an Optitrak motion tracking system (Optitrak, Inc; Corvallis, OR, USA) sampling at 100Hz, which tracked the position of a spherical marker placed on the nail of the right-hand index nger, as it slid on the table plane. We performed a spatial calibration to synchronize the axes and units for movements on the The subject was required to maintain posture at a xed distance from the table and to place his/her chin on the chinrest. Eye gaze and pupil diameter from both eyes were tracked and recorded with an EyeTribe oculometer (Oculus, Menlo Park, CA, USA), sampling at 60Hz. We used a chinrest to stabilize posture and x the head position at a predetermined distance from the screen and from the oculometer.

Experimental Task
The participants performed a decision-making task between two reaching movements, seeking precision at target arrival. Maximum precision is attained when arriving at the centre of the wide side of one of both rectangle targets. To provide a positive metric of precision other than the error, we de ned a metric of reward that decreased linearly with the error (FIG 1B). During the experiment, we varied three factors: motivation, biomechanics, and the requirement of stopping at the target (FIG 1B-C

Manipulation of Motivation
Our manipulation of motivation was performed by means of a social hierarchy de ned within the experiment, as a function of the participant's aiming skill with regard to that of simulated partners. We informed the participant that he/she would perform within a community of participants, and that at each block, he/she would have a different partner from this community to perform alongside. To reinforce the belief of a real partner, at the end of each trial we showed a horizontal green bar displaying the accuracy attained, normalized between 0 and 100%, alongside a red bar displaying the accuracy of their partner.
Each nine trials we also showed a photograph of the participant and of their partner, ranked in a top-down fashion, at a height proportionally to their average end-point accuracy. Since the goal of the simulated partner was to introduce a subliminal bias that would modulate the participant's intrinsic motivation, we informed each participant this was not a task of competition, that the partner was for companionship purposes, that the ranking was only to keep track of your performance, and that the participant should focus only on performing the task and not on outperforming their partner. To parameterize the bias introduced by the presence of the virtual partner, we classi ed partners of two types, as a function of their worse of better aiming accuracy with respect to the participant (FIG 1D). We also matched the participant and the partner gender to avoid cross-gender effects.
Despite the instruction, we predicted that participants will be concerned by their accuracy with respect to their partner, and that they will adjust their process of selection of motor parameters and/or their policy to select between movements. However, this could be attained by simply trading-off speed by accuracy, or by varying one of these two metrics in a cost-bene t scenario. The in uence of the motor cost on the choices was varied in a random fashion on a single trial basis, as we used one of two geometrical arrangements (FIG 1A). By contrast, the social position and the control regime were maintained constant during each block and varied across blocks (10 block types), presented in a counter-balanced fashion.
Each participant performed two sessions of six blocks of one-hundred and eight trials each. Each session included six blocks, three per control regime; one block solo, and two blocks alongside both kinds of partners. Subjects were given real-time visual feedback on their trajectories. interactions between them, as de ned by equation 1. Note that the BM×C, was de ne to group low-cost and high-cost movements within arrangement (0-Low Cost/1-High Cost), and to ultimately quantify the in uence of motor cost on our metrics.
For each explaining factor, we identi ed the regression coe cients b that signi cantly differ from 0 (as reported in FIG 2A-B and Suppl. FIG 1A-B) by running a Mann-Whitney-Wilcoxon (rank sum) test between the distribution of b pooled over all subjects and a null distribution. The null distribution was built by shu ing the regressed values f k across trials within each subject, then pooling the surrogate coe cients of each type across all subjects together. Note that the z-scoring applied to the kinematic markers allows for the alignment of the distributions across the different subjects. In this way we identi ed the explaining factors (and interactions thereof) that affected the kinematic markers (as quanti ed by the regression) to compare the effect of motivation with other expected factors like habituation or fatigue (#Block and #Trial) and biomechanical constraints (BM). We set the threshold for group signi cance at p<0.05 with the Mann-Whitney-Wilcoxon (rank sum) test to compare the distribution of regression coe cients estimated from the data and the corresponding null/surrogate distribution.

Oculometry Analyses
Like the movement markers, we also z-scored our recordings of pupil diameter and regressed this metric with the same GLM used to test behaviour (Eq. 1). We calculated the GLM b-coe cients per subject at each 30ms during two intervals, centred around the tSO and at the tGO events (FIG 4C-D). We determined group signi cance for each variable by running a t-test across the b-regression coe cients obtained across subjects (FIG 4A). Again, the threshold for statistical signi cance was set at p<0.05 (without correction for multiple comparison, considering each time step to be independent).

EEG Pre-processing
We selected an interval of interest of 1200ms for further analyses, starting 800ms before the rst stimulus onset ---the initial cue, and ending 400ms later (see Fig. 1A). We selected this interval for two main reasons, rst to seek for baseline changes in the network of motivation, which should not depend on each speci c trial but on the block of trials ---the partner was the same during each block of trials, and so should the modulation of motivation. Second, this interval preceded any movement, and was therefore likely to contain fewer artifacts than those following the stimulus presentation.
We applied a 4 th order notch lter around 50, 100 and 150Hz to prevent electrical interference from the . The procedure to identify eye-movement related sources was semi-automatized, rst correlating each source obtained with the signal from the electrodes recording eye movements to obtain a rst metric of relatedness. Second, we visually inspected all sources to corroborate that their shape and spatial location matched those of ocular artefacts. Eye-related sources were removed and the cleaned signal obtained by inverting the ICA process.
Extraction of Neuronal Signals (Source Space) The pre-processed EEG datasets were transformed from electrode into source space via custom-made MATLAB scripts based on the default EEGLAB ICA algorithm in combination with the electrode spatial location map. The Brain Products Unicap 64 con guration was used to establish a spatial reference between the electrode placement and the sources location, and we assumed a spherical head model.

EEG Analyses in both Electrode and Source spaces
We used cross-validated classi cation to assess how much information about motivation is present in the EEG signals. Following the task description, we de ned three motivated states as a function of the types of partner the participant performed the task with: Solo (M=0), when alone; Easy (M=1), when performing with a partner of lesser skill than the participant; Hard (M=2), when performing with a partner more skilled than the participant (see behavioural task, FIG 1D). The signals, in electrode or source space, were rst ltered in three frequency bands: a [8-12Hz], b , g . We cut out our interval of interest from the EEG temporal series, from 800ms before the rst stimulus onset, until 400ms after (yielding 1200 time points at a resolution of 500Hz), i.e. at the tShowTarget event (FIG 1A). We applied our classi cation pipeline to each recording session, type of signal and frequency band. This pipeline relied on two speci c metrics to capture complementary aspects of the brain network of motivation: signal power and pairwise correlation between signals (as a proxy for interactions, which is equivalent to spectral coherence averaged over each frequency band used here weighted sum of the entries, which is then recti ed by a sigmoid function, to calculate an output ranging from 0 to 1 for each category. This indicates the con dence of the classi cation in the predicted category; the decision is made by selecting the category whose con dence is the highest. The training procedure tunes the weights (or regressors) for all features to reduce the prediction error on the train set. By contrast, the 1NN calculates a similarity measure between the sample in the test set and all samples in the train set (using the Pearson correlation between the two vectors of features) and attribute to the test sample the category of the most similar train sample. Here the 1NN does not involve training.
To identify the most informative (or "best") features for the classi cation, we used Recursive Feature Elimination (RFE) on the MLR; because the MLR gives a better performance than the 1NN. In brief, its strong weights (in absolute value) correspond to important features (also because the MLR gives a better performance than the 1NN). RFE provides a ranking of features by their importance by iteratively excluding features that weakly contribute to the classi cation and optimising the remaining weights. Ultimately, the most relevant features survive the pruning process. In practice, we performed RFE for each train set (in each train/test split) and then evaluate the stability of the feature rankings by calculating the Pearson correlation coe cient between all pairs of rankings. In the case of stable ranking (typically Pearson correlation coe cients above 0.6), the mean ranking over all train/test splits is calculated. This gave two sets of best features: electrodes for power and interactions between pairs of electrodes for correlation (FIGS 5-6 and Suppl. FIG 2).

Common sources
The common sources were calculated for each participant using the power of source signals. Following the heuristic that the eight best sources obtained using RFE were su cient to attain an over 90% accurate classi cation (FIG 6B), we pooled the eight best sources over all sessions of all subjects, over both motion types. We used the Girvan-Newman algorithm [70] to calculate communities of similar sources (spatially distributed over the cortical surface) among the original 25-45 sources obtained per subject.
Similarity between pairs of sources was established using the Pearson correlation over the vectorized head maps, after retaining only the more localized part (similar to a spatial recti cation to favour narrow bumps). Once the sources were grouped in communities, the centroid head map of each community was calculated by averaging all corresponding sources. We performed this operation independently for each frequency band, yielding six communities of common sources for the a, six for the b-band and ve for the g-band.
Declarations Figure 1 A. Timeline of a typical trial. It starts with the presentation of a white screen during 500ms. After this time, a 1cm diameter pale blue origin circle cue is presented on the centre bottom part of the screen. 1ms after the subject's ngertip crossed its border, the origin cue changed to green, and two rectangular blue targets (10 x 1cm) and two pale blue via-points (1cm diameter) we presented to the right/left of the origin cue. Path distances from the origin to either target extended 15cm. The GO signal was reported by removing the origin cue, at which time the participant could report his/her choice by moving the cursor from the origin to the selected target. The rectangle target turned from blue to green when the cursor entered the target. After 500ms, everything disappeared. 500ms after this, a green horizontal bar was shown during 500ms to report the precision of the movement just performed. If the participant was accompanied, a red bar was also shown below, reporting the simulated partner's precision. An empty screen ITI followed, the     All Sources Space EEG Analyses. A. EEG pre-processing pipeline. It starts with a band-pass lter between 0.1 and 100Hz, a notch lter to remove power supply noise at 50Hz and its harmonics (100, 150Hz). This is followed by a de-composition into independent components to identify and remove eye and muscular components, to ultimately re-compose the original signal with the original components. B. Pre-processed EEG analysis pipeline. It rst includes an ICA decomposition for the case the analyses were performed into source space (FIG 6). This is followed by a decomposition into the α, β and γ-bands, with are used for further analysis. C. Motivated State classi cation accuracy and confusion matrices, as by the MLR and 1-NN classi ers, as a function of #Motivated state, for each frequency band and metric (electrode power, correlation), performed on the EEGs of Subject #1. D. Same as C, averaged across all subjects.

Figure 6
Selected Common Sources Space EEG Analyses. A. Common sources across subjects related to the task for the α, β and γ-bands . B. MLR Classi cation accuracy for each frequency and metric (source power, source covariance, source correlation).

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download. CosDecoGilsonFigs2020v28Suppl.pdf