Thanks for the comment, Vladimir. Basically Rx_Noise^2 is the integral of PSD over the entire bandwidth, so it is an incomplete description of noise characteristics. To correct an error in the example in my original email, the noise RMS after the averaging filter is Rx_Noise/sqrt(N), not (1/N)*Rx_Noise/sqrt(N). Fangyi From: Dmitriev-Zdorov, Vladimir [mailto:vladimir_dmitriev-zdorov@xxxxxxxxxx] Sent: Monday, May 20, 2013 6:30 PM To: RAO,FANGYI (A-USA,ex1); ibis-macro@xxxxxxxxxxxxx Subject: RE: concern on Rx_Noise Fangyi, Yes, we can study noise propagation only if we know its spectral density. Noise PDF (or sigma for Gaussian case) can serve only as rough characterization of the noise generated on the spot, but not the one that propagates through the channel (or filter). What's the difference compared to Tx_jitter specifying 'Gaussian'? In this case, we don't know spectral density either and typically assume uncorrelated noise. Vladimir ________________________________ From: ibis-macro-bounce@xxxxxxxxxxxxx<mailto:ibis-macro-bounce@xxxxxxxxxxxxx> [ibis-macro-bounce@xxxxxxxxxxxxx] on behalf of fangyi_rao@xxxxxxxxxxx<mailto:fangyi_rao@xxxxxxxxxxx> [fangyi_rao@xxxxxxxxxxx] Sent: Monday, May 20, 2013 6:02 PM To: ibis-macro@xxxxxxxxxxxxx<mailto:ibis-macro@xxxxxxxxxxxxx> Subject: [ibis-macro] concern on Rx_Noise Experts; While thinking about how to handle Rx_Noise in redriver, I came to realized that Rx_Noise's behavior in the downstream channel can't be defined. As described in BIRD 123, Rx_Noise is the RMS of a Gaussian voltage noise, and its unit is Volt. That information is inadequate to determine how the noise is transferred when passing through a filter because to compute the filtered noise one needs to know the input noise Power Spectral Density (PSD) in Volt/sqrt(Hz), but not RMS. Take a simple example and consider a low pass filter that performs a time average on the signal within a window of 1UI. Assume N time points are used per UI. It's well known that the RMS of the average over N random noise samples is (1/N) * Rx_Noise / sqrt(N) where 1/N is the weight of each sample interval. Now the filter output noise depends ambiguously on the number of time points per UI, which is a simulation setting. Am I missing something? Fangyi