Contribution
A mechanistic account of the readiness potential (RP): a decision threshold applied to autocorrelated noise, here the output of a leaky stochastic accumulator. The account fits the RP’s shape from behavioural parameters alone, and its prediction, verified in a new “Libetus interruptus” experiment, is that speeded responses to unpredictable cues are preceded by a slow negativity when they are fast.
Key insight
Average enough noise traces aligned to the moment each crossed a threshold, and the average rises smoothly toward the threshold, even though no individual trace was “building up” toward anything. Movement-locked averaging is a backward selection: only epochs ending in a movement are analysed. So an RP is what one should see if, as in a decision-making task, the decision to move is a threshold crossing, and in an uncued task the accumulator’s input is mostly spontaneous fluctuation. What the experimenter calls planning may be the trace of noise that happened to cross.
Assumptions
- Same machinery as perceptual decisions (section “The Stochastic-Decision Model”): an accumulator plus threshold, fed in this task with internal noise instead of evidence, because the instructions forbid basing the decision on evidence.
- Model (Eq. 1): δx_i = (I − kx_i)Δt + cξ_i√Δt, with c = 0.1 and Δt = 0.001. I is a constant urgency from the task’s demand characteristics; k is leak; ξ is Gaussian noise. With leak, urgency raises the baseline toward the threshold rather than ramping to it.
- Threshold expressed as a percentile of output amplitude over 1,000 simulated trials of 50,000 steps. Fitted values k = 0.5, I = 0.11, β = 0.298 (80th percentile).
- Independent trials and a reset to zero after each crossing; no memory between trials.
- The fit window ends at −150 ms, the time of maximum slope of the lateralised RP, taken as the end of the pre-commitment phase. Activity after it is attributed to motor execution.
- Interrupted responses are assumed to come from the same accumulator as self-initiated ones (citing Hughes et al.). A click is simulated as a steep linear ramp added at the interruption.
Key results
- Classic Libet task (16 recruited; 2 excluded for no RP at any electrode; 14 for EEG). The mean reported urge was 152 ms before movement (±33 ms SEM), consistent with earlier reports. Normalised waiting-time distributions were broad and right-skewed, fitted “excellent[ly]” (Fig. 1B).
- RP from behaviour (Fig. 1C). With the waiting-time parameters, the sign-reversed average of simulated traces fitted the RP from −3 to −0.15 s (r² = 0.96, P < 10⁻⁹). Nearby parameters fitted worse (Fig. S1, not read). The fit was about as good with the window ending at −200 or −100 ms.
- Mean–SD relation (Fig. 2). Mean and SD of waiting times were linearly related across participants (r = 0.9, P < 0.00001), as for first-passage times of drift-diffusion processes (Wagenmakers & Brown 2007) and for the model.
- Libetus interruptus (13 participants; 150 trials). Adding random clicks left the RP on uninterrupted trials unchanged in the data and the model (Fig. 3). Responses to clicks, split into fastest and slowest thirds: fast responses were preceded by a more negative potential, over −0.8 to −0.3 s before movement and over 0.5 s before the click (both P < 0.005, two-sided signed-rank; Fig. 4). The simulation predicted the same (Fig. 4 E–F). Pre- click times of fast and slow responses were distributed equally across the trial (P = 0.64, n = 13), which excludes a slowly building general readiness (CNV). “Coincidence” trials (4%) were excluded.
- Residual misfits. Waiting times on successive trials were positively correlated, and participants preferred the “bottom” of the clock (Figs. S3–S4). Some participants moved earlier when clicks were possible.
- Reinterpretation (Discussion). The RP has two nonlinear components, a pre-commitment phase dominated by fluctuations and a post-commitment motor phase in the last ~150 ms. The neural decision to move is the threshold crossing, together with the lateralised RP and an abrupt rise in corticospinal excitability. Its time coincides with the reported urge, so the brain’s estimate of its own decision time is “reasonably accurate”, and there is no special gap for motor decisions as compared with sensory ones.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | A leaky accumulator fed with noise and weak urgency, fitted to waiting times alone, reproduces the RP’s shape | moderate to strong: three parameters, behaviour-only fit, r² = 0.96 | Fig. 1 |
| C2 | Fast responses to unpredictable cues are preceded by a slow negativity, as the model predicts and the planning account does not | moderate: n = 13, one experiment, P < 0.005 | Fig. 4 |
| C3 | The neural decision to move occurs about 150 ms before movement, not at RP onset | interpretive: follows within the model; the authors say the data show the RP “could” be spontaneous activity, not that it is | Discussion |
| C4 | Libet’s conclusion that the neural decision precedes awareness by half a second or more is unfounded | moderate as a rebuttal of an inference; it does not establish the reverse | Discussion |
| C5 | The same bounded-integration mechanism serves perceptual and spontaneous decisions | weak to moderate: the mean–SD relation is consistent with it but shared by many processes | Fig. 2 |
| C6 | Conserved 1/f autocorrelation explains why crayfish and humans show similar premovement buildup | speculation | Discussion |
Concepts
- readiness potential (RP, Bereitschaftspotential): the slow negative EEG deflection preceding self-initiated movement in the movement-locked average (Kornhuber & Deecke 1965).
- neural decision to move now: the threshold crossing that commits to a movement, distinct from the conscious decision and from RP onset. Likened to tipping the first domino: ballistic, not deterministic, since a veto can still intervene.
- waiting time (WT): time from trial start to the spontaneous movement.
- urgency: the constant input I, from the task’s implicit demand to move within about 20 s.
- backward selection bias: only epochs ending in a movement are averaged, so the fluctuations that produced the crossing are recovered.
Connections
- Usher & McClelland 2001 (LIT-565). The source of the model (their ref. 27), here a single accumulator with no competitor. The paper calls it “a well-known accumulator model (DDM)… an extension of an earlier model developed by Ratcliff”. With leak it is strictly an Ornstein–Uhlenbeck process, not a pure diffusion.
- Bogacz et al. 2006 (LIT-569). In their terms this is a stable O-U process, with an attractor at I/k below the threshold. My mapping; neither paper draws it.
- Ratcliff & McKoon 2008 (LIT-576). The perceptual-decision framework the paper borrows: commitment as a threshold crossing.
- Dennett & Kinsbourne 1992 (LIT-431, NOTE-367). Both papers deny that Libet showed an unconscious decision preceding a conscious one. Dennett and Kinsbourne argue from the timing of represented content and treat the clock report as an artefact; Schurger et al. argue from the mechanism of the RP and accept the clock report as an approximate estimate of the time of commitment. They are compatible, but they are not the same argument, and they disagree on how far the urge report can be trusted.
Bearing on the record
- Agency and free will. It removes an empirical premise, not a philosophical position. THEORY-029 and THEORY-040 do not rely on Libet, so it does not bear on them. It is the record’s evidence that the RP does not establish a neural decision before awareness, and THEORY-059 states that claim.
- Behavioral integration. It extends the evidence-accumulation account of decision from choosing which to choosing when, with goals acting by moving a baseline rather than by issuing a command. That THEORY is filed as THEORY-061.
- No ML instruction. The brain–computer-interface remark (the RP’s limits as a self-paced BCI trigger) is a prediction about neural data, not an ML practice. Nothing here belongs in the anthology.
Limitations
- Possibility, not proof. “Although our study demonstrates that the readiness potential could reflect non–goal-directed (spontaneous) neural activity, it does not prove that this possibility is in fact the case.”
- Narrow task. One movement per trial, spontaneous timing, no reason to move at a particular time; deliberate choices with stakes are not tested.
- Small samples. 14 and 13 participants for EEG; electrode chosen per participant from the classic task.
- Model limits. No inter-trial memory although waiting times were serially correlated; history cannot extend before trial start, so the early simulated RP is noisy.
- Silent on the urge. The model does not say what the conscious urge is or whether it plays a causal role.
Open questions
- Does the account carry over to deliberate decisions with consequences, or to the Kornhuber–Deecke self-paced task with time estimation?
- What sets the urgency and the baseline, and could slow fluctuations of the threshold explain the trial-to-trial correlation?
- Is the reported urge time a readout of the threshold crossing, as the authors propose, and how would one test that independently of the clock method?
Corrections
- none to a seeded skim (there was no seed)