Increasing contrast reduces tuning sharpness, intrinsic noise and nonlinear processing.

Each histogram in A,C,E plots the distribution of 148 independent estimates of correlation coefficients computed from entire dataset (>half a million trials); each estimate refers to a specific SNR/contrast level from a specific task in an individual observer. A plots distribution of contrast-dependence (solid bars, red labels) and SNR-dependence (open bars, blue labels) for sharpness of spatial tuning (smooth curves show polynomial fits of degree 6 for visualization purposes only); C,E plot same for internal noise and composite linearity index (see Methods for how these quantities were defined and measured). Ordinate plots number of estimates that fall within the corresponding bin on abscissa. Dependence (plotted on x axis) is the correlation coefficient between the manipulated variable (contrast or SNR) and the measured characteristic (sharpness, internal noise or linearity). Three square insets to panel C show three examples of how internal noise varied as a function of contrast (top inset with x axis labelled ‘contrast’) or SNR (bottom insets with x axis labelled ‘SNR’). Top inset plots internal noise values at four contrast levels (plotted on x axis) and SNR = 3 for one participant in the foveal bright-bar detection task; bottom-left inset plots values at four SNR levels (plotted on x axis) and contrast = 1% for one participant in the mixed-polarity detection task; bottom-right inset plots values at four SNR levels and contrast = 2% for one participant in the peripheral bright-bar detection task. In each inset, red/blue line shows linear fit, light-red/light-blue lines show associated 90% confidence intervals. From each inset, a correlation coefficient is computed and binned into the histograms in C as indicated by arrows. This process is repeated across the entire dataset, to produce the full distributions in A,C and E. When distribution is skewed to the left (e.g. solid bars in A), relevant characteristic (e.g. sharpness in A) is greater at low contrast (solid bars) or SNR (open bars); when distribution is skewed to the right (e.g. solid bars in E), relevant characteristic (e.g. linearity index in E) is greater at high contrast/SNR. Colour-graded red/blue horizontal bars mark confidence intervals for the mean (95/99/99.9% for low/mid/high colour saturation). B,D,F plot distributions from 37 independent estimates (one estimate from each observer and task) for separability of contrast- versus SNR-dependence (estimated using singular value decomposition (SVD); see Methods).