1 INTRODUCTION

Prioritized processing of fearful compared to neutral faces is reflected in behavioral and neural responses, such as event-related potentials (ERPs). Detecting and responding to emotional facial expressions is highly relevant, providing specific perceptual advantages and neuronal responses. Thus, fearful faces exhibit a higher probability of breaking through Continuous Flash Suppression (Yang et al., 2007), decreased visual extinction in brain-damaged patients (Carlson & Reinke, 2008), decreased susceptibility to the attentional blink effect (Fox et al., 2005; Jong et al., 2009), and elicit distinct ERP modulations (for reviews, see Hinojosa et al., 2015; Schindler & Bublatzky, 2020).

One reason for perceptual advantages has recently been found in systematic physical differences between fearful and neutral faces regarding their spatial frequency spectrum. It has been shown that the spectrum of fearful faces matches the human contrast sensitivity function (CSF) better than neutral faces, thus providing more so-called effective contrast (Hedger et al., 2015a, 2015b, 2019). However, the role of the spatial frequency spectrum in modulations of ERPs is not fully understood.

Early ERP components represent distinct stages of face- and emotional expression processing. The occipitally scored P1 component is sometimes found to be enlarged for faces compared to objects (e.g., see Neumann et al., 2011; Thierry et al., 2007), while other studies report no differences (e.g., see Ganis et al., 2012; Schendan & Ganis, 2013), or even the reverse effect (e.g., see Schindler, Tirloni, et al., 2021). Findings on P1 modulations by fearful faces are mixed and variable (see Schindler & Bublatzky, 2020), while P1 increases are also observed for scrambles from fearful faces (Schindler, Wolf, et al., 2021). The N170 is viewed as a structural encoding component and reliably enlarged for faces compared to objects (Eimer, 2011) and fearful compared to neutral expressions (Hinojosa et al., 2015). The subsequent Early Posterior Negativity has been related to early attentional selection (Mühlberger et al., 2009; Schupp et al., 2004; Wieser et al., 2010), and is reliably larger for fearful than for neutral expressions (Mühlberger et al., 2009; Schindler et al., 2019; Smith et al., 2013; Walentowska & Wronka, 2012; Wieser et al., 2012).

The findings listed above outline some variable (P1) and some consistent (N170 and EPN) early ERP amplifications for fearful compared to neutral faces. Based on Hedger and colleagues' behavioral findings (2015a, 2015b, 2019), we examined emotion-specific face frequency information modulations on ERP responses in a previous EEG study (Bruchmann et al., 2020). Here, face images contained the average power spectra of all neutral, of all fearful, or of the averaged neutral and fearful faces. We found enlarged N170 amplitudes to fearful faces but no interaction with spatial frequency. However, interactions of emotional expression and spatial frequencies were observed for the P1 and EPN. For both components, larger effects of facial expression were observed when the spectrum contained neutral instead of fearful frequencies. Notably, for the EPN, fearful and neutral faces did not differ when exhibiting fearful frequencies, while typical emotion effects were found when faces contained average or neutral frequencies (Bruchmann et al., 2020). According to Hedger et al. (2015a, 2015b, 2019), enhanced processing of fearful faces can be partly attributed to their greater effective contrast, possibly due to evolutionary adaptation of visual sensitivity to low-level features of facial expression and/or vice versa, that is, due to adaptation of facial expressions to the human CSF. In general, the human CSF and the spatial frequency spectrum of human faces match rather well (Keil, 2008). However, whereas the human CSF is enhanced for mid-range spatial frequencies (≈1–10 cpd, peaking around 3–5 cpd; De Valois et al., 1974) across all orientations, these frequencies are not evenly distributed across orientations in facial expressions. For example, the frequency spectrum of the mouth region of the human face is dominated by specific spatial frequencies (due to its typical size in natural face-to-face situations) at specific orientations (predominantly horizontal for upright faces).

It has long been known that cells in the primary visual cortex (V1) are tuned in spatial frequency and orientation (Blakemore & Campbell, 1969), but whether the cortical architecture of V1 supports independent processing or not is still debated (Nauhaus et al., 2012; Sirovich & Uglesich, 2004). It is thus possible that evolutionary adaptation of the CSF to facial expressions occurred in an orientation-independent fashion, that is, by increasing the sensitivity for specific frequencies, irrespective of their orientation tuning profile, or to specific combinations of orientation and spatial frequency. Notably, while ERPs of interest here are not assumed to reflect V1 activity directly, V1 activity very likely contributes to subsequent modulations of ERPs.

Finally, orientation and spatial frequency information alone is insufficient for processing visual objects, even when they are two-dimensional gray-scale images of objects. The phase spectrum is the third ingredient necessary to specify an object entirely. The configural information of an image, that is, recognizable image contours, requires the alignment of phases across different orientations and spatial frequencies. By scrambling the phase spectrum (typically by replacing the original phase spectra with spectra of random noise images), one can keep orientation and spatial frequency information unchanged while removing all contours and thus the "identity" of an object.

The present pre-registered study (https://osf.io/7zacf) thus seeks to dissect spectral information of fearful and neutral faces systematically. We do this by manipulating "orientation coherence" (OC) and "phase coherence" (PC) independently. OC is manipulated by either leaving the amplitude spectrum intact or replacing it with the so-called rotational average, which averages the power of a given spatial frequency across all orientations. PC is manipulated by either leaving the phase spectrum intact or replacing it with a randomized phase spectrum. Phenomenologically, OC manipulations barely impact face images, whereas PC manipulations determine whether the stimulus is perceived as a face of random noise (see Figure 1). We hypothesized for the P1 window the main effect of emotion category (fearful vs. neutral) but no effect of PC (faces vs. scrambles). We explored the role of OC (original vs. rotational average spectra) as it is unclear whether orientation and spatial frequency information are already integrated at the time of the P1. For the N170, we proposed that fearful-neutral differences depend on configural information (i.e., recognizable faces) and thus expected an interaction of emotion category and PC, but not between emotion category and OC. For the EPN, we assumed that PC is necessary for emotional effects and may be further enhanced by OC. The latter hypothesis is based on our previous study (Bruchmann et al., 2020), showing the effects of subtle spectrum manipulations on the EPN. This study also indicated that the late positive potential (LPP), which is often analyzed in studies on emotion processing, is not modulated by spectrum manipulations. It was also not modulated by emotional expression, which is common in studies distracting attention from emotional information (daSilva, Crager, Geisler, et al., 2016; daSilva, Crager, & Puce, 2016; Eimer & Holmes, 2007; Rellecke et al., 2011; Schindler et al., 2020). Based on these findings, the EPN, but not the LPP, were included in the present study. Additional to these pre-registered analyses, we explored the ERPs by performing an independent component analysis (ICA) to disentangle potential overlaps of the ERPs of interest.

Stimulus creation process. The illustration shows the process for one example face. After cutting the image to size and applying an oval mask to remove non-facial parts, Fast Fourier Transformation (FFT) is performed to separate the amplitude spectrum (Ampl) from the phase spectrum (Ph) of each image. After manipulating the spectra, inverse Fast Fourier Transformation (iFFT) is applied to obtain faces with intact and randomized phase and orientation coherence, respectively

2 METHOD 2.1 Participants

In total, we examined 46 participants, from which two participants had to be excluded due to lost or missing data, three due to bad EEG data, and one due to failing the inclusion criteria of no reported history of neurological or psychiatric disorders. The resulting 40 participants (28 female and 12 male) fulfilled the registered data sampling plan, for which power calculations (G*Power 3.1.7, Faul et al., 2009) showed a power of >90% to detect medium effects sizes (). Participants gave written informed consent and received 10 Euros per hour for participation. All participants had normal or corrected-to-normal vision and had a mean age of 23.85 years (SD = 3.79). The detailed preregistration can be retrieved in the Open Science Framework (https://osf.io/7zacf) and all raw data and paradigm information in the attached OSF project (https://osf.io/56dhm/).

2.2 Stimuli

The faces were taken from the Radboud Faces database (Langner et al., 2010). We converted the faces into greyscale and cut out each face's oval center, removing facial hair. By this means, we sought to prevent the spatial frequency spectrum from being dominated by expression-irrelevant components. Thirty-six identities (18 male and 18 female) were used, displaying either fearful or neutral expressions, both presented in four different conditions: By means of Fourier analysis, the phase and amplitude spectra of each image were calculated (see Figure 1a). Intact orientation coherence refers to an unchanged amplitude spectrum, that is, the amplitude of each specific spatial frequency-orientation combination remains the same. The rotational average was calculated by averaging all orientations of each spatial frequency, which lay on a quarter circle in the amplitude spectrum. The spectrum thus contains the same power of each spatial frequency. However, the power is no longer concentrated on specific orientations per spatial frequency but evenly distributed across orientations. PC was removed by replacing the phase spectra with the spectra of random noise images.

All frequency manipulations resulted in the same mean luminance per image. Finally, in line with the suggested maximal influence of spatial frequencies on emotional awareness (Hedger et al., 2015a), presented face-pictures exhibited a visual angle of about 6.2° (bizygomatic diameter). Stimuli were presented on a Gamma-corrected display (Iiyama G-Master GB2488HSU) running at 60 Hz with a Michelson contrast of 0.9979 (Lmin = 0.35 cd/m2; Lmax = 327.43 cd/m2). The background was set to medium gray, corresponding to the average image luminance of 163.89 cd/m2.

2.3 Procedure

The experiment was programmed and run with Matlab (Version R2016a; Mathworks Inc., Natick, MA; http://www.mathworks.com), the Psychophysics Toolbox (Version 3.0.15; Brainard, 1997; Kleiner et al., 2007) and the Eyelink Toolbox (Cornelissen et al., 2002). A fixation mark was presented in each trial for a randomized duration between 300 and 700 ms, followed by a face for 50 ms, followed by a blank screen presented for 500 ms before the next fixation mark was presented (see Figure 1c). Participants were instructed to avoid eye-movements and blinks during stimulus presentation. Participants' gaze position was evaluated online with an eye tracker (EyeLink 1000, SR Research Ltd., Mississauga, Canada), stopping the presentation whenever the center was not fixated. Thus, stimulus presentation was paused whenever participants did not direct their gaze at a circular region with a radius of 0.7° around the fixation mark. If a gaze deviation was detected for more than five seconds despite a participant's attempt to fixate the center, the eye-tracker calibration procedure was automatically initiated.

Additionally, participants were instructed to respond to an oddball trial by pressing the space bar. Response feedback was provided for hits (key presses within 1 s after oddball presentation), slow responses (key presses within 1–3 s), and false alarms (key presses outside these windows) through a corresponding text presented for 2 s at screen center. Oddball stimuli consisted of homogenous gray ovals. Sixty trials were presented for each of the six facial expression and frequency conditions within a randomized stimulus stream and a total of 15 oddball target trials. Participants further responded to a demographic questionnaire and the BDI-II and STAI Trait questionnaire (Spielberger et al. 1999; Hautzinger et al. 2009) and a short version of the NEO-FFI (Körner et al., 2008), not relevant for the current research question.

2.4 EEG recording and preprocessing

EEG signals were recorded from 64 BioSemi active electrodes using Biosemis Actiview software (www.biosemi.com). Four additional electrodes measured horizontal and vertical eye-movement. Recording sampling rate was 512 Hz. Offline data were re-referenced to average reference, filtered with a 0.01 Hz (6 dB/oct; forward) high-pass filter, and a 40 Hz low-pass filter (24 dB/oct; zero-phase). Recorded eye-movement were corrected using the automatic eye-artifact correction method implemented in BESA (Ille et al., 2002). A predefined source model was applied to the data, combining three topographies accounting for EOG activities, consisting of horizontal and vertical eye-movement and blinks (HEOG, VEOG, blink) with 12 regional sources modeling the different brain regions. The adaptive artifact correction method then performed a principal component analysis (PCA) for segments where the correlation between data and artifact topography exceeded the HEOG (150 µV) or VEOG (250 µV) thresholds. All PCA components explaining more than the minimum variance were maintained. The recorded data were decomposed using all topographies into a linear combination of brain and artifact activities (Ille et al., 2002). The remaining artifacts were rejected based on an absolute threshold (<120 µV), signal gradient (<75 µV/∂T), and low signal (i.e., the SD of the gradient, >0.01 µV/∂T). On average, 53.98 trials per condition remained after rejection (SD = 5.14, min = 24, max = 60). Noisy EEG sensors were interpolated using a spline interpolation procedure. On average 1.73 sensors were interpolated (SD = 1.85, min = 0, max = 7). A delay of the LCD screen for stimulus presentation of 29 ms, measured with a photodiode, was corrected during epoching. Filtered data were segmented from 100 ms before stimulus onset until 1000 ms after stimulus presentation. Baseline-correction used the 100 ms before stimulus onset.

2.5 EEG data analyses

EEG scalp-data were statistically analyzed with EMEGS (Version 3.0; Peyk et al., 2011) and JASP (Love et al., 2019). Two (emotional expression: fearful vs. neutral faces) × two (orientation coherence: intact vs. rotational average) × two (PC: intact vs. scrambled) repeated measure ANOVAs were set-up to investigate the main effects of emotional expression, OC and PC, as well as their interaction in time windows and electrode clusters of interest. Partial eta-squared () and Cohen's d were calculated to describe effect sizes (Cohen, 1988). When Mauchly's test of Sphericity indicated a violation of sphericity, degrees of freedom were corrected following Greenhouse-Geisser. Time windows of interest were chosen based on similar previous studies (Bruchmann et al., 2020) and visual inspection of the collapsed conditions. We measured ERPs from 80 to 110 ms for the P1, 130 to 170 ms for the N170, and from 240 to 300 ms to investigate EPN effects, deviating in time for the P1 (registered 80–100 ms) and EPN (registered 230–330 ms). For the P1, N170, and EPN time windows, registered symmetrical occipital clusters were examined (P9, P7, PO7, P10, P8, and PO8).

However, even though P1 and N1 components were observed for non-recognizable scrambled stimuli, their amplitude is delayed and broader than those for faces (see Figure 2 below, see also Schindler, Tirloni, et al., 2021). Given these specific temporal processing differences, we aimed to use better-suited methods to compare specific processing stages of emotion, PC, and AC. Thus, we further carried out the following unregistered analysis: we performed ICA at the group level using the group ICA toolbox EEGIFT (V1.0; Eichele et al., 2011). For this purpose, for each subject, single-trial data was sorted by experimental condition. Since the data was already low-pass filtered, no further smoothing was applied. A PCA was applied first to reduce the data to the first 20 principal components sorted by Eigenvalue. Then ICA was performed using the Infomax algorithm and the default settings as implemented in EEGIFT.

Effects of emotion, OC, and PC for the P1, N170, and EPN. (a) ERP waveforms show the time course, averaged over selected sensors of interest (highlighted in (d)). (b) Average amplitudes per component and condition. Error bars display 95% confidence intervals around means. (c) ERP differences between fearful and neutral conditions. The plots contain 95% bootstrap confidence intervals of intra-individual differences. (d) Topographies of fearful-neutral differences in the selected intervals of interest. The corresponding sensors are highlighted in green

3 RESULTS 3.1 Behavior

With regard to behavioral performance, participants had an average hit rate of pHit = 0.955 (SD = 0.085) and a false alarm rate of pFA = 0.0022 (SD = 0.0048). The average reaction time was MRT = 523 ms (SD = 74.32 ms).

3.2 P1

For the P1, there was no main effect of emotion (F(1,39) = 0.03, p = .859, ), and of OC (F(1,39) = 3.58, p = .066, ), but an effect of PC (F(1,39) = 4.10, p = .050, ; see Figure 2), with larger P1 amplitudes for scrambles compared to faces (t(39) = 2.02, p = .050, Cohen's d = 0.320). There were no interactions of any of the factors (Fs < 2.40, ps > .129).

3.3 N170

For the N170, there were main effects of emotion (F(1,39) = 9.80, p = .003, ; see Figure 2), of OC (F(1,39) = 32.84, p < .001, ), and of PC (F(1,39) = 276.94, p < .001, ). There were larger N170 amplitudes for fearful compared to neutral expressions (t(39) = 3.13, p = .003, Cohen's d = 0.495), larger N170 amplitudes for faces compared to scrambles (t(39) = 16.64, p < .001, Cohen's d = 2.631), and larger N170 amplitudes for original compared to rotational average spectra (t(39) = 5.73, p < .001, Cohen's d = 0.906). There were significant interactions of emotional expression and PC (F(1,39) = 50.90, p < .001, ; see Figure 2), and between PC and OC (F(1,39) = 8.71, p = .005, ). Here, fearful faces elicited a larger N170 than neutral faces (t(39) = 7.15, p < .001, Cohen's d = 1.130), while the reversed effect was observed for scrambles: Neutral as compared to fearful scrambles elicited a larger negativity (t(39) = 2.55, p = .013, Cohen's d = 0.403; see Figure 2).

Regarding the interaction between PC and OC, for faces, slightly larger N170 amplitudes were found for original compared to rotated faces (t(39) = 1.99, p = .050, Cohen's d = 0.315), while for scrambles a much larger negativity for original compared to rotated scrambles was found (t(39) = 6.15, p < .001, Cohen's d = 0.972). There were no further interactions of these three factors (Fs < 1.86, ps > .181).

3.4 EPN

For the EPN, there was a main effect of emotion (F(1,39) = 4.36, p = .043, ), and of PC (F(1,39) = 63.12, p < .001, ) but no significant effects of OC (F(1,39) = 3.31, p = .077, ). There were larger EPN amplitudes for fearful compared to neutral expressions (t(39) = 2.09, p = .043, Cohen's d = 0.330) and for faces compared to scrambles (t(39) = 7.95, p < .001, Cohen's d = 1.256). There was a significant interaction of emotional expression and PC (F(1,39) = 4.91, p = .033, ), while all other interactions failed to reach statistical significance (Fs < 2.74, ps > .106). The interaction of emotion and PC, was characterized by larger emotion fearful-neutral differences among faces (t(39) = 3.03, p = .007, Cohen's d = 0.479) than among scrambles (t(39) = 0.50, p = .617, Cohen's d = 0.079; see Figure 2).

3.5 Explorative ICA

The ICA yielded three ICs with very distinct time courses and topographies that resembled the components of interest. Markedly worse signal-to-noise ratios and topographies characterized all other ICs without discernable positive or negative poles. We defined intervals of interest for each IC with a width of 20 ms around positive and negative peaks (see Figure 3). For each IC and interval of interest and we calculated average activations and performed 2 × 2 × 2 repeated-measures ANOVAs to examine the main effects of and emotion (fearful vs. neutral), OC (original vs. rotation average), PC (face vs. scramble), and their respective interactions.

Results of the independent component analysis, showing the effects of emotion, OC, and PC (a) Scalp topographies depict the grand average of each IC. (b) ERP waveforms show the time course over-highlighted sensors. Error bars display 95% confidence intervals around means. (c) Respective difference plots contain 95% bootstrap confidence intervals of intra-individual differences

The first IC was characterized by an occipital positivity, peaking at 120 and 220 ms. These two peaks will be referred to as P1- and late P2-ICs. Note that we did not select the negative peak around 150 ms as its amplitude is close to zero and thus shows a nearly flat topography. For the early P1-IC we observed a main effect of PC (F(1,39) = 49.77, p < .001, ), with scrambles leading to higher amplitudes than faces. The main effect of OC was also significant (F(1,39) = 6.247, p = .017, ), with orientational average spectra (OAS) leading to higher amplitudes than original spectra. For the P2-IC we observed a main effect of emotion (F(1,39) = 9.268, p = .004, ), PC (F(1,39) = 135.4, p < .001, ), and a main effect of OC (F(1,39) = 6.332, p = .016, ). We further observed an interaction of PC × OC (F(1,39) = 8.633, p = .006, ), characterized by larger effects of OC among faces (t(39) = −4.24, p < .001, Cohen's d = −0.670) than among scrambles (t(39) = 0.23, p = .821, Cohen's d = 0.360). Finally, the interaction of PC × emotion was significant (F(1,39) = 6.7, p = .013, ), characterized by larger emotion effects for scrambles (t(39) = 4.0, p < .001, Cohen's d = 0.633) than for faces (t(39) = 0.86, p = .397, Cohen's d = 0.136).

The second IC was characterized by an occipital negativity, with negative peaks at 147 and 262 ms and a positive peak at 196 ms. Please note that negative IC activations reflect reversed polarities, thus the early and late interval of this IC describe occipital positivities and the mid interval an occipital negativity. As shown in Figure 3b, the topography of this IC is highly similar to the first IC but its time course is markedly different. Based on the polarity reversal around 170 ms we refer to this component as the late P1/N170-IC. For the early interval, we observed main effects of emotion (F(1,39) = 9.765, p = .003, ) and of OC (F(1,39) = 13.49, p = .001, ). There were significant two-way-interactions of PC × OC (F(1,39) = 7.687, p = .008, ), characterized by larger OC-effects among faces (t(39) = 4.25, p < .001, Cohen's d = 0.672) than scrambles (t(39) = 0.77, p = .448, Cohen's d = 0.121), and two-way-interactions of OC × emotion (F(1,39) = 5.888, p = .02, ), characterized by larger emotion effects among rotational averages (t(39) = −3.86, p < .001, Cohen's d = −0.611) than among original spectra (t(39) = −0.82, p = .419, Cohen's d = −0.129). Additionally, there was a three-way interaction of PC × OC × emotion (F(1,39) = 8.461, p = .006, ), indicating that the OC × emotion interaction was more pronounced among scrambles (t(39) = 4.88, p < .001, Cohen's d = 0.771) than among faces (t(39) = −0.50, p = .618, Cohen's d = −0.08). The mid interval, featuring an occipital negativity peaking at 196 ms, showed significant main effects of PC (F(1,39) = 22.23, p < .001, ) and of OC (F(1,39) = 7.502, p = .009, ). The late, positively peaking interval showed a main effect of PC (F(1,39) = 13.37, p = .001, ) and an interaction of PC × OC × emotion (F(1,39) = 4.794, p = .035, ), also indicating that the OC × emotion interaction was more pronounced among scrambles (t(39) = 2.52, p = .016, Cohen's d = 0.399) than among faces (t(39) = 0.17, p = .865, Cohen's d = 0.027).

The third IC closely resembled the classical N170, both in its occipito-temporal topography and its pronounced negativity peaking at 182 ms. There were also small additional negative activations (thus occipito-temporal positivities) peaking at 118 and 241 ms. For the earliest peak (118 ms) the main effect of emotion (F(1,39) = 8.21, p = .007, ), as well as the main effect of PC was significant (F(1,39) = 8.63, p = .006, ). The mid interval, peaking at 182 ms, showed a main effect of emotion (F(1,39) = 7.289, p = .01, ) and a main effect of PC (F(1,39) = 241.9, p < .001, ). Additionally, we observed significant interaction between PC and OC (F(1,39) = 8.173, p = .007, ), characterized by positive OC effects among scrambles (t(39) = 2.11, p = .042, Cohen's d = 0.333) and negative OC effects among faces (t(39) = −1.65, p = .106, Cohen's d = −0.261). Finally, there was a significant interaction between PC and emotion (F(1,39) = 29.07, p < .001, ), characterized by larger emotion effects among faces (t(39) = 5.46, p < .001, Cohen's d = 0.863) than among scrambles (t(39) = −1.30, p = .202, Cohen's d = −0.205). In the late interval peaking at 241 ms only the main effect of PC was significant (F(1,39) = 39.13, p < .001, ).

4 DISCUSSION

We tested the effects of orientation and phase coherence (OC and PC) and their influence on ERPs to fearful and neutral faces. This way, we examined if early ERP amplifications to fearful faces are driven by specific properties of their spatial frequency spectrum. The most pronounced effects on occipito-temporal ERPs were caused by PC, that is, by keeping versus removing configural information. This effect was present in all intervals of interest. Faces with emotional expressions elicited the typical amplifications of N170 and EPN amplitudes but did not affect P1 amplitudes. In general, OC had the smallest effect on ERPs, appearing only in interaction with PC for the N170. The explorative ICA overcame problems of differential timing of the ERP components based on PC information and appeared to be more sensitive in finding the effects of our experimental manipulations, including OC. To reconcile the findings of both analyses, we discuss the results in their temporal sequence.

Concerning the P1 component, we found reduced P1 amplitudes for recognizable faces, in line with some recent findings (e.g., see Schindler, Tirloni, et al., 2021; but see Neumann et al., 2011; Schindler, Bruchmann, et al., 2021; Schendan & Ganis, 2013). We recently raised the hypothesis that such P1 effects are based on a sustained positivity for scrambles, which is not interrupted by a sharply peaking N170 component in contrast to intact faces (Schindler, Busch, et al., 2021; Schindler, Tirloni, et al., 2021). Emotion effects were absent and did not interact with PC or OC. The explorative ICA additionally revealed a small boost of the early P1 by intact OC, but did not show any effects of emotion. This observation adds to the mixed findings in the literature (see Schindler & Bublatzky, 2020). However, in studies that detect P1 effects on emotional faces, these effects seem to be strongly driven by low-level visual differences between emotional and neutral faces (e.g., see Schindler, Bruchmann, et al., 2021; Schindler, Wolf, et al., 2021). Thus, our study supports the hypothesis that fearful expressions affect the P1 similarly for intact and scrambled faces, while here the absence of emotion effects for intact faces comes along with the absence of effects for scrambled face-versions.

Concerning the N170 we could validate all predictions, as the N170 was enlarged for faces compared to scrambles (e.g., see Eimer, 2011; Schindler, Bruchmann, et al., 2021; Schindler, Tirloni, et al., 2021). PC also interacted with OC, showing that intact faces with original orientation information elicited the largest N170 response. Further, fearful expressions increased amplitudes (see Hinojosa et al., 2015), and most importantly, interacted with PC, but not OC. For intact faces, fearful faces increased the N170, while this was not the case for scrambles. This indicates that fearful expressions increase visual processing based on intact phase coherence and thus on configural information rather than the intensity relations among spatial frequencies alone. There is no evidence, however, that a low-level association between spatial frequencies and orientation is required.

Of further interest is a reversal of the N170 emotion effect for scrambled faces (somewhat more pronounced with intact OC). This reversal has been observed before (Schindler, Tirloni, et al., 2021), but its meaning is unclear at present. The explorative ICA provides interesting insights into this reversal, as it shows that the IC most closely resembling the N170 (Figure 3c, middle interval) contains no reversed emotion effects for scrambles, in contrast to the P2-IC (Figure 3a, late interval). The reversed effect is thus most pronounced over occipital sensors around 220 ms and not over temporal sensors in the N170 time range. It is the same component that most closely resembles the P1 in its earlier interval (P1-IC). We speculate that the spectral differences of fearful and neutral faces are processed in the absence of configural information in occipital areas, creating a "perceptual hypothesis." Late occipital differences may signal a mismatch detected between this hypothesis and feedback from a configural analysis in the case of scrambled images and none such mismatch in the case of faces (a template-mismatch, see also a recent model for visual search by Eimer, 2014). This tentative explanation is fostered by the observation that the P2 shows no emotion effects for faces, but exclusively the reversed emotion effect for scrambles. A plausible electrophysiological index of this perceptual hypothesis could be the visual C1, which is assumed to be generated in primary visual cortex (Capilla et al., 2016; Di Russo et al., 2001, 2005). The latter is known to be a critical visual area in spatial frequency processing (Henriksson et al., 2008). Furthermore, some studies show that already the C1 can be modulated by emotional facial expressions (Acunzo et al., 2019; Pourtois et al., 2004). In our study, C1 effects could not be tested since stimuli presented at the vertical center of the display typically cause opposing electric current flow in the upper and lower halves of V1 (Clark et al., 1994; Jeffreys & Axford, 1972), preventing the measurement of reliable C1 signals (Capilla et al., 2016). As noted above, this explanation is tentative and future studies could therefore examine the interaction of phase coherence and emotional expression on the visual C1 by presenting stimuli at parafoveal locations, ideally in combination with fMRI measurements or source analytical approaches.

Finally, during the EPN window, we found the expected larger negativity for fearful than for neutral expressions with intact PC (Mühlberger et al., 2009; Smith et al., 2013; Walentowska & Wronka, 2012; Wieser et al., 2012), indicating early attentional selection (Schupp et al., 2004; Wieser et al., 2010). We previously showed that fearful-specific frequency information plays a role during the EPN (Bruchmann et al., 2020). Nevertheless, in contrast to our expectation, the emotion by PC interaction was not further influenced by OC. We conclude that the EPN requires intact configural information for differential emotion processing.

We found emotion effects for the N170 and EPN, which were bound to phase coherence (i.e., recognizable faces) but were unaffected by OC. Thus, we found a different involvement of different low-level properties on emotion effects, showing that emotional information from the N170 onwards depends on intact configural information.

This result further substantiates and extends a previous study on the effects of emotion-specific spatial frequency information in faces, where we manipulated amplitude spectra but kept PC information intact (Bruchmann et al., 2020). It also corresponds to findings on the rapid categorization of complex emotional scenes, showing that successful categorization required intact PC (Rhodes et al., 2019). Interestingly, the distinction between faces and other objects does not necessarily rely on intact PC: phase-scrambled faces have been shown to attract extremely fast saccades (≈100 ms) away from phase-scrambled images of vehicles (Honey et al., 2008). Also, face images "pop out" in an array of non-face images by their amplitude spectrum. They no longer do so when the amplitude spectra are equated (VanRullen, 2006). The studies show that phase-independent information from the amplitude spectrum alone is sufficient for at least some face/non-face differentiations. Differential effects within the same object category appear to require phase information. It should be added, however, that the present study manipulated phase (and orientation) coherence in an all or none fashion. A parametric variation of phase coherence. Rousselet et al. (2008) showed that P1 and N170 vary systematically in latency and amplitude with the degree of phase coherence. Future studies could employ such a parametric variation to dissect when and how much configural information is needed for differential processing of fearful and neutral faces.

Our findings add to specific neuronal sensitivity toward the detection of fearful faces. According to Hedger et al. (2015a, 2015b, 2019), enhanced processing of fearful faces can be partly attributed to their greater effective contrast, possibly due to evolutionary adaptation of visual sensitivity to low-level features of facial expression and/or vice versa, i.e., due to adaptation of facial expressions to the human CSF. From our present results we infer, that the CSF to facial expressions adapted in an orientation-independent fashion by increasing the sensitivity for specific frequencies, irrespective of their orientation tuning profile.

4.1 Conclusion

To concl