something working, with hystersis for comparator and summing input with feedback

circuit-20251130-2020.svg

lowering RC to just 1kohm + 1nF results in the RC charging going all the way to either rail, which is good because that means that every PWM pulse has an identical symmetric charge and discharge profile, thus the total integration of that signal would equal the total integration of a PWM. That is good for consistency, so not sensitive to RC values and so can digitally reconstruct exactly.

Note, a simple IIR digital filter can act as a capacitive RC block to remove DC: https://en.wikipedia.org/wiki/High-pass_filter#Discrete-time_realization

so to decimate the delta PWM signal, apply that formula y[n]=x[n]-x[n-1]+a*y[n-1] and continuously accumulate that, and dump every nysquist sampling rate sum as the delta for that nyquist sample. If can get the oscillation to be around 96kHz, then a simple boxcar averager will null that oscillation when downsampling to 48kHz.

Also note during analog reconstruction, a simple notch filter will null around that oscillation.

hypothetical notch filter:

Might want to combine notch with general lowpass and maybe notch higher overtones too.

nice no jaggedy integration