The present study was designed to test whether the visual system can generate the Sustained Posterior Negativity (SPN)—a reliable neural marker of symmetry perception—even when stimuli are presented too briefly to support prolonged processing or top-down modulation. To achieve this, we asked participants to discriminate symmetrical and asymmetrical octagonal shapes and applied two key constraints to the task. First, stimuli were presented for only 20 ms, an exposure far shorter than the typical onset of the SPN (~250 ms; Bertamini and Makin, 2014; Makin et al., 2022; Norcia et al., 2002). Second, the symmetrical stimuli were constructed in such a way that they were frequently elongated along one axis (see Method section for details). This elongation both distorted and defined an object-centered reference frame whose orientation and salience varied randomly across trials, rendering the stimuli less predictable than completely regular octagons. When the object’s principal elongation axis overlapped with one of the symmetry axes, that axis tended to dominate perceptual organization, rendering it more salient. As illustrated in Figure 1, the resulting stimuli preserved mirror symmetry but varied systematically in contour and in the relative dominance of competing axes. To minimize reliance on the external reference frame of the display, the two symmetry axes were oriented obliquely (45° and 135°), preventing alignment with either the cardinal axes or the screen’s diagonals (Treder, 2010; Wenderoth, 1996). Although feedback processes cannot be entirely ruled out, the short presentation strongly restricts the temporal window available for recurrent activity (Lamme and Roelfsema, 2000).

Importantly, any neural activity occurring after target onset—including visual transients or the initiation of motor preparation—can interfere with or obscure SPN. To eliminate such contamination, participants were not only instructed to delay their response, but were also prevented from preparing any motor action, as the correct response mapping (left vs. right key) was revealed only after a one-second interval. This ensured that the EEG signal during the SPN window remained uncontaminated by lateralized motor potentials and could be attributed exclusively to perceptual integration mechanisms.

Together, these methodological features define a constrained environment in which top-down modulation and prolonged visual input are minimized. This offers an opportunity to isolate the minimal conditions required for SPN generation. If a robust and fully expressed SPN emerges despite the brief and unpredictable stimuli, this would indicate that feedforward input is sufficient for initiating and sustaining symmetry-selective neural activity. In contrast, if no SPN is observed, it would support the view that feedback and temporal integration are necessary for constructing global representations of regularity (Alp et al., 2018; Machilsen et al., 2009).

Critically, if the SPN is present but substantially attenuated, this would suggest that feedforward processing is sufficient to trigger SPN onset, but that feedback mechanisms are necessary for full consolidation and for achieving the typical magnitude of the response. This would imply a hybrid dependency in which initial symmetry detection operates automatically via early visual pathways, but full engagement of symmetry-sensitive networks—especially those associated with sustained activity in ventral visual areas—relies on reentrant and top-down amplification. Such an outcome would also align with the known temporal profile of the SPN: even though the stimulus is visible for only 20 ms, the SPN typically begins around 250 ms and extends well beyond, making it likely that recurrent processing contributes to its full expression.

This is the first study to attempt eliciting the SPN under such brief exposure. Although symmetry-related effects have occasionally been reported in earlier ERP components (e.g., P1 or N1, Dering et al., 2024; Makin et al., 2012), these lack the temporal and functional specificity of the SPN, which reflects a sustained perceptual response to global regularity. Nevertheless, an early feedforward contribution cannot be excluded a priori. Accordingly, we adopted an agnostic approach and tested explicitly for symmetry-related activity at any latency and scalp distribution. To this end, we used a nonparametric, data-driven cluster-based permutation analysis, which enables unbiased detection of spatiotemporal clusters in the EEG signal where responses to symmetrical and asymmetrical stimuli diverge, providing a statistically rigorous and assumption-light means of identifying the presence, onset, and topography of symmetry-related neural activity.

Finally, as a confirmatory analysis, we computed the SPN using a canonical literature-derived posterior ROI (O1, O2, Oz, PO3, PO4, PO7, PO8) and the conventional 250–600 ms window. This spatiotemporal definition is widely adopted in the SPN literature and summarized in recent reviews and catalogues (Bertamini et al., 2018; Dering et al., 2024; Derpsch et al., 2024; Makin et al., 2022; Wright et al., 2015).

24 participants (14 female; mean age = 24.83 years, SD = 4.81) were recruited from the University of Padova. All participants reported normal or corrected-to-normal vision and no history of neurological or psychiatric disorders. Each provided written informed consent. The study protocol was approved by the university of Padova ethics committee (Comitato etico della ricerca psicologica—area 17), protocol number 588-a (17/05/2024). Data collection formed part of a broader research initiative investigating metacontrast masking.

The experimental sessions took place in a quiet, dimly lit room. Participants were seated 57 cm from a calibrated PHILIPS100 LCD monitor (60 cm width, 1920 × 1,080 resolution, 100 Hz refresh rate), with head position stabilized via a chinrest to maintain a fixed viewing geometry. Stimuli were generated in PsychoPy3 (version 2021.2.3; Peirce et al., 2019).

Octagonal stimuli for both the symmetric (S) and asymmetric (A) conditions were generated dynamically with a parametric algorithm. The base template was a regular octagon inscribed in a circle of radius 0.75° of visual angle (1.5° diameter). Each of the eight vertices was independently rotated about the origin by an angular jitter uniformly sampled from −20° to +20°, keeping radial distance fixed but producing irregular, non-equidistant contours. This ensured that no perfectly regular octagon was ever displayed.

The two conditions differed only in how these vertices were constrained. In the asymmetric condition, all eight vertices were independent, yielding stimuli with no mirror- or rotational-symmetry aside from rare accidental configurations. In the symmetric condition, two anchor vertices were random and the remaining six were generated by successive reflections across the vertical and horizontal axes. After construction, the stimulus was rotated by either 45° or 135°, so the symmetry axes (originally vertical and horizontal in object coordinates) appeared at 45° and 135° in display coordinates. Importantly, these orientations did not coincide with the monitor diagonals (e.g., ~29°/151° on a 16:9 display), preventing the screen’s reference frame from serving as a cue. Because the same base geometry and radial extent were used in both conditions, average global size, perimeter, and enclosed area were matched, minimizing luminance and contrast-energy confounds.

Both S and A stimuli were elongated (anisotropy). In A stimuli, elongation simply established a principal (object-centered) axis without introducing mirror structure. In S stimuli, elongation sets the object reference frame to align with one of the two oblique symmetry axes (45° or 135°), and thus to be misaligned with the other. This trial-to-trial variability in which axis captured the reference frame prevented the adoption of fixed strategies or canonical templates. The task therefore required trial-by-trial global shape extraction to detect symmetry under minimal visual input and variable geometry.

Each trial followed a fixed temporal sequence. A central black fixation dot (≈ 0.2° visual angle) appeared on a uniform mid-gray background for a variable duration (850–1,150 ms, mean = 1,000 ms), after which a single filled black octagon was presented centrally for exactly 20 ms (two frames at 100 Hz). Participants were instructed that a geometric shape would briefly appear at fixation on each trial and were asked to judge whether it was symmetric or asymmetric.

A critical design feature of the present paradigm was the trial-by-trial randomization of the mapping between symmetry judgments and motor responses. Specifically, participants were informed that the labels “symmetric” and “asymmetric” would be variably assigned to either the left or right arrow key on each trial. To eliminate anticipatory motor preparation and ensure that the SPN reflects perceptual rather than premotor activity, the response mapping cue (e.g., “LEFT = symmetric; RIGHT = asymmetric” or its reverse) was presented only after stimulus offset—a protocol consistent with best practices in symmetry-related ERP research (Makin et al., 2022). This post-stimulus cueing effectively decouples lateralized readiness potentials (LRPs) from the post-perceptual time window of interest, thereby preserving the interpretive specificity of the SPN as an index of symmetry detection.

After reading the instructions, participants pressed the spacebar to begin and then completed a practice block designed to familiarize them with the brief, irregular stimuli and with the delayed, trial-wise response mapping. The practice block replicated the main task in all respects: the same fixation jitter (850–1,150 ms), the same 20 ms stimulus duration, a 1 s blank interval after stimulus offset, the same post-offset display of the key mapping, and the same centrally presented octagons. Only two brief instructional elements were added during practice. First, immediate accuracy feedback was presented for 1 s after each response (“Correct” in green; “Incorrect” in red). Second, during the 20 ms stimulus, a thin red oblique guideline was momentarily superimposed along one potential axis of symmetry; on each practice trial its orientation was 45° or 135°, selected pseudorandomly. Because the guideline orientation and the shape’s geometry (including any elongation that defined an object-centered reference axis) varied from trial to trial, the guideline always coincided with one of the two possible symmetry axes but only sometimes aligned with the object’s intrinsic reference frame.

The practice block comprised 80 trials. Because most participants were naïve to EEG, this block was essential to ensure stable task performance and proper recording behavior (e.g., minimizing blinks and muscle tension). While a slightly shorter practice might have sufficed, this dose remains within the same order of magnitude as familiarization phases commonly used in SPN studies, where practice typically consists of a few dozen trials (sometimes repeatable to criterion) and includes trial-level feedback to stabilize performance before the no-feedback main task (e.g., Karakashevska et al., 2024; Rampone et al., 2022).

In the main experiment, participants completed 120 pseudorandomized trials. No feedback or axis cues were presented; between trials the screen contained only a fixation point. No backward masking followed the 20 ms stimulus, ensuring that EEG responses during the SPN window were driven solely by the brief shape onset.

Electrophysiological data were recorded with a 64-channel actiCAP arranged according to the extended 10–20 international system. Preprocessing was fully scripted and executed in EEGLAB v2024.2 (Delorme et al., 2011) running in MATLAB R2023a. Continuous data were resampled to 256 Hz and then band-pass filtered from 0.5 to 40 Hz using zero-phase Hamming-windowed sinc FIR filters to minimize phase distortion. To attenuate mains-related line noise, CleanLine was applied in multi-taper mode with adaptive sinusoid regression at 50 and 100 Hz (Mullen, 2012).

Automated continuous-data cleaning proceeded with the clean_rawdata/ASR framework (Kothe et al., 2019) using the following a-priori criteria applied uniformly across participants: FlatlineCriterion = 10 s, ChannelCriterion = 0.80 (correlation-based channel outlier detection), LineNoiseCriterion = 4.5, High-pass range for ASR calibration = 0.25–0.75 Hz, BurstCriterion = 30, WindowCriterion = 0.35 with WindowCriterionTolerances = [−Inf, 10], and BurstRejection = ‘off’. Channels flagged at this stage were removed for decomposition and later restored by interpolation.

Independent Component Analysis (ICA) was performed with the Picard algorithm, after PCA reduction to the number of retained channels to ensure full rank. Components were classified with ICLabel (v1.3) and automatically rejected when their class probabilities exceeded the following thresholds: Muscle ≥ 0.95, Eye ≥ 0.80, Heart ≥ 0.80, Line noise ≥ 0.95, Channel noise ≥ 0.95, and Other ≥ 0.95; no component was rejected based on “Brain” probability (Pion-Tonachini et al., 2019).

After ICA-based artifact removal, previously removed channels were re-interpolated to the original 64-channel montage using spherical splines. Data were then re-referenced to the common average computed over scalp EEG channels only.

Continuous EEG was epoched from −500 to +1,000 ms relative to stimulus onset for asymmetric (“A”) and symmetric (“S”) trials, and a baseline correction was applied using the −200 to 0 ms interval. In keeping with best practice for stimulus-locked ERPs, only correct trials were retained for ERP averaging: behavioral accuracy codes were aligned to epochs, and incorrect trials were removed automatically after epoching. Remaining epochs then underwent a fully automated rejection pass based on scalp channels only (EOG excluded): trials were rejected if they exceeded ±120 μV within −200 to 600 ms, exhibited a linear trend with slope > 75 μV and R² > 0.50, or showed kurtosis > 7 SD at either the local- or global-channel level.

Across the 24 participants processed with this automated pipeline, the mean number of channels removed by the continuous cleaning stage and subsequently re-interpolated was 5.25 per participant (range: 3–9). The number of independent components rejected by ICLabel-guided automation ranged from 1 to 8, with a mean of 4.08 components per participant. Behavioral exclusion of incorrect responses removed on average 24.04 trials per participant from the initial 120, after which the automated epoch-quality screen flagged and removed a further 2.46 trials on average (≈ 2.5% of available epochs; range: 0–6). The final number of retained epochs per participant was therefore 93.5 on average (SD = 13.52; range: 57–111). Condition-wise, participants contributed on average 44.88 asymmetric trials (range: 28–55) and 48.63 symmetric trials (range: 25–57), which were adequate for stable ERP estimation in the predefined occipito-parietal regions of interest, as demonstrated by a reliability analysis of SPN amplitudes as a function of per-participant epoch count reported in the Results section. All preprocessing steps, parameters, and decision rules were applied uniformly by the scripted workflow, ensuring full reproducibility and eliminating rater-dependent variability from the pipeline.

To identify spatiotemporal clusters reflecting sustained symmetry-related responses, we performed a non-parametric, within-subject cluster-based permutation test. For each participant, trials were first averaged within condition to obtain subject-level time courses at every electrode. Condition differences were then evaluated with paired (two-tailed) mass-univariate tests computed at each electrode–timepoint.

Spatiotemporal clusters were defined as contiguous sets of electrode–time samples whose univariate tests exceeded a cluster-forming threshold of p < 0.05. Spatial adjacency was based on inter-electrode distances, treating sensors separated by < 35 mm as neighbors, yielding on average ~3.3 neighbors per electrode; as expected for peripheral sites, TP9 and TP10 had no valid neighbors. This neighborhood structure balances anatomical specificity with the expected spatial spread of scalp potentials in high-density EEG (Maris and Oostenveld, 2007). Family-wise error was controlled at the cluster level using a Monte Carlo permutation procedure (5,000 random label-swaps). On each iteration, condition labels were exchanged within participants to simulate the null hypothesis while preserving the within-subject design. For every permuted dataset, the maximum cluster mass (sum of test statistics over each spatiotemporally contiguous cluster) was recorded to form the empirical null distribution of maximal effects. Observed clusters were deemed significant if their mass fell in the extreme 5% of this distribution (two-sided, cluster-level α = 0.05).

This approach produces binary spatiotemporal masks marking samples that belong to statistically reliable clusters and supports conservative, data-driven inference about the timing and scalp distribution of the sustained posterior negativity without imposing a priori constraints on latency or topography.

We quantified symmetry-related activity as the SPN, defined as the mean ERP amplitude difference between symmetric and asymmetric trials (S − A) computed over posterior electrodes within a specified time window. In the data-driven analysis, both the electrode set and the latency window were taken directly from the significant spatiotemporal cluster identified by the permutation test (see above). For each participant and condition, we first averaged the ERP within this window at each electrode in the cluster, yielding one mean per electrode per condition. We then computed electrode-wise difference scores (S − A) and averaged these across the cluster electrodes to obtain a single SPN value per participant.

For completeness and comparability with prior work, we also computed a canonical SPN using a literature-based posterior ROI (O1, O2, Oz, PO3, PO4, PO7, PO8) and the conventional 250–600 ms window. The same steps were applied: per-condition averaging within the canonical window at each ROI electrode, electrode-wise S − A differencing, and averaging of these differences across ROI electrodes. Both the cluster-defined and canonical SPN estimates were carried forward, with the former serving as the primary, data-driven definition and the latter as a robustness check.

The non-parametric, trial-level cluster permutation analysis revealed a single statistically significant negative spatiotemporal cluster, consistent with an SPN pattern in which responses to symmetric stimuli were more negative than to asymmetric stimuli. The cluster extended from 309 to 508 ms post-stimulus (duration ≈ 199 ms), encompassed 25 posterior occipito-parietal electrodes (P7, P3, Pz, P4, P8, O1, O2, Oz, CP2, PO3, PO4, CP5, CP6, TP7, CPz, CP3, CP4, TP8, P5, P1, P2, P6, PO7, POz, PO8), and reached a maximal spatial extent of 14 contributing electrodes at 383 ms. The cluster-level probability, corrected for multiple comparisons, was p = 0.002. No positive cluster (Symmetric > Asymmetric) survived correction (see Figure 3).

To estimate SPN amplitude at the individual level, we computed each participant’s mean ERP difference (Symmetric − Asymmetric) within the spatiotemporal extent of the negative cluster window (309–508 ms), which was defined by the cluster-based permutation test rather than a priori assumptions (FieldTrip cluster p = 0.002). As shown in Figure 4A, the mean SPN across participants (N = 23) was −0.446 μV (median = −0.397 μV). The interquartile range spanned −0.669 to −0.036 μV, indicating that most individuals showed a clear negativity with relatively few null or reversed effects. Maximal negativity within the window occurred at ≈387 ms (≈ − 0.636 μV). A one-tailed one-sample t-test confirmed that the SPN was less than zero, t (22) = −3.390, p = 0.001; the one-sided 95% CI was (−∞, −0.173] μV and the two-sided 95% CI was [−0.719, −0.173] μV. The corresponding within-subject effect size was d_z = −0.71 (medium to large).

To characterize the scalp distribution, we located the time of maximal negativity within the same window by averaging the SPN difference wave across participants and electrodes; the peak occurred at ≈387 ms. The resulting topography (Figure 4C) shows a posterior distribution with strongest negativity over occipital–parietal sites. For reference, Figure 4B displays the grand-average ROI waveforms for Symmetric (red), Asymmetric (blue), and the SPN difference (green) with 95% confidence bands.

We asked whether the SPN was stronger over one hemisphere by computing two related lateralization indices (LI) within the cluster-defined window (309–508 ms) and using the same left- and right-hemisphere electrode sets.

Condition-based LI. For each participant and for each condition (Symmetric, Asymmetric), we averaged ERP amplitudes across the cluster electrodes and time points. We then computed the right–minus–left difference within each condition and combined the two conditions into a single, normalized index by dividing that difference by the sum of the absolute left and right means. This index reflects any hemispheric bias present in the overall waveforms, whether or not it is specifically tied to symmetry.

SPN-based LI. For each participant we first formed the SPN on each hemisphere (SPN_L = Symmetric minus Asymmetric on the left; SPN_R = Symmetric minus Asymmetric on the right). We then took the right–minus–left difference of these SPNs and normalized it by the sum of their absolute values. This index isolates lateralization of the symmetry effect itself, ignoring activity common to both conditions.

The SPN-based indices were near zero [M = 0.127, SD = 0.689, 95% CI (−0.171, 0.425), t (22) = 0.89, p = 0.385, d_z = 0.18]. A Bayesian one-sample test also favored the null (BF10 = 0.31; roughly 3:1 evidence for no lateralization). The condition-based index led to the same conclusion (t (45) = −0.21, p = 0.832, BF10 = 0.16).

The sustained posterior negativity elicited by 20-ms symmetric stimuli is essentially bilateral: the SPN does not show a reliable left–right asymmetry, and the raw ERPs likewise do not reveal a systematic hemispheric bias.

An important question is whether participants who exhibit a stronger neural response to symmetry (as indexed by the SPN) are also those who perform better in the behavioral task or adopt a more conservative or liberal decision strategy. To address this, we computed correlations between each participant’s SPN amplitude and their signal-detection metrics: sensitivity (d′) and decision criterion (c). SPN amplitude was defined as the mean difference between Symmetric and Asymmetric trials (S − A), averaged across the data-driven posterior NEG cluster (25 electrodes) and the cluster-defined time window (309–508 ms), which captures the sustained posterior negativity elicited by symmetry.

For perceptual sensitivity (d′), the Pearson correlation with SPN amplitude was very small (r = 0.062) and not statistically significant, with a 95% confidence interval (CI) that ranged from a moderately negative to a moderately positive association (−0.359 to 0.463). This interval includes zero, indicating that we cannot rule out the possibility of no association at all. The corresponding Bayesian correlation analysis produced a Bayes Factor of BF10 = 0.46, meaning the observed data were about 2.2 times more likely under the null hypothesis of no correlation (BF01 ≈ 2.2). The 95% Highest Density Interval (HDI) for the population correlation ρ was [−0.320, 0.403], again including zero and suggesting substantial uncertainty about the true relationship.

For the decision criterion (c), which reflects response bias, the correlation was slightly larger but still non-significant [r = 0.160, 95% CI (−0.270, 0.536), p = 0.467]. Participants who showed a stronger (i.e., more negative) SPN did not consistently adopt a more conservative or liberal response tendency. The Bayesian analysis yielded BF10 = 0.56 (BF01 ≈ 1.8), which provides only anecdotal evidence favoring the null hypothesis. The 95% HDI for ρ was [−0.249, 0.465], indicating that correlations in either direction remain plausible given the present sample size.

In this dataset, the neural index of symmetry processing (SPN) appears largely independent of both discrimination performance and decision bias. The convergence of frequentist non-significance, confidence intervals that include zero, and Bayesian evidence favoring the null suggests that any genuine brain–behavior association—if present—is at most small to moderate. With N = 23 the achieved power to detect a medium correlation (ρ = 0.30) is modest (≈ 0.28), and the breadth of both confidence and credibility intervals counsels caution against treating the absence of an effect as definitive. Future studies should therefore employ larger samples to narrow these intervals and yield a more precise estimate of any putative association (see Figure 5).

Within the posterior ROI (O1, O2, Oz, PO3, PO4, PO7, PO8; 250–600 ms), the SPN was reliably negative. Across participants (post-exclusion N = 23), the mean SPN was −0.507 μV (median = −0.221 μV), with an interquartile range from −0.919 to 0.154 μV—indicating that most individuals showed a clear negativity, with few null or reversed effects. A one-tailed one-sample t-test confirmed that the SPN was less than zero, t (22) = −2.83, p = 0.005, yielding a one-sided 95% CI of (−∞, −0.199] μV [two-sided 95% CI: (−0.879, −0.135) μV]. The corresponding within-subject effect size was d_z = −0.59, in the medium range.

To determine the minimum number of epochs per condition needed for a reliable canonical SPN, we performed a fixed-cohort subsampling analysis on epoch-level ROI means (post-exclusion N = 23). We evaluated five candidate counts per condition, m ∈ {5, 10, 15, 20, 25}. For each m, we ran 1,000 permutations. Within each permutation—and within each participant—we sampled without replacement m symmetric and m asymmetric epochs, computed the participant’s SPN as mean(S) − mean(A), and tested the across-participant mean against zero with a one-tailed one-sample t-test (H₁: SPN < 0). “Reliability” was defined as the proportion of permutations with p < 0.05; 95% confidence intervals were exact Clopper–Pearson intervals. In parallel, for each m we summarized the mean SPN across permutations and its 2.5–97.5% permutation band.

Subsampling was balanced and used a fixed-cohort: every included participant contributed the same number of epochs per condition. Accordingly, the largest testable m is capped by the participant with the fewest balanced usable epochs (here, 25). In the present dataset, available counts were typically much higher (per-participant medians ≈ 45 for A and 51 for S; minima 28 and 25). The analysis is therefore conservative: it asks how many trials per condition are minimally sufficient for reliable detection in this cohort while preventing a few high-trial participants from disproportionately driving the result. Reliability increased monotonically with m: 30.5% (5), 45.9% (10), 64.0% (15), 79.4% (20), and 89.1% [25; exact 95% CI ≈ (0.87, 0.91)]. At m = 25, the SPN was significant in ≈89% of permutations. Across permutations, the mean SPN remained stably negative (≈ − 0.52 to −0.53 μV), indicating that improved reliability with larger m reflects reduced sampling variability rather than changes in effect magnitude.