Dear Andrew, Your assertion is right on target that the RSS (square-root-of-sum-of-squares) approximation for the risetime of a cascaded linear system is ... just an approximation. I can fill in some of the mathematics behind the imperfections in this approximation. If you are not mathematically inclined, you might want to step out for a quick cup of tea, and then check back in at the end to see the cases where RSS doesn't work very well. The RSS risetime approximation stems from a general property of the convolution operator (when cascading linear systems their impulse responses convolve, so the properties of the convolution operator therefore determine how the complete system behaves): [A] When impulse responses convolve, their variances add v[overall] = v[1] + v[2] + ... v[n] This property applies to any impulse response that is strictly positive, meaning that the associated step response is monotonic. It's provable mathematically from first principles using the definitions of variance and the convolution operator (although I will admit I am loath to reconstruct the proof here). The convolution property relates to the RSS approximation through the following chain of logic (High-Speed Digital Design, p. 400): (1) Variance is the square of standard deviation (2) The standard deviation of an impulse response is proportional to its width (3) The width of an impulse response is proportional to the rise time of its corresponding step response This chain of reasoning suggests that the rise time of a step response is proportional to the standard deviation of its associated impulse response. Combining this conclusion with [A] you can derive RSS: trise[overall]^2 = trise[1]^2 + trise[2]^2 ... + trise[n]^2 where the operator "^2" means "squared" To see what goes wrong in the chain of reasoning 1-3, let's get specific. Suppose you have two systems that you propose to cascade. Looking at the step responses, measure the 10-90% rise times t1 and t2 (or 20-80%, or whatever interests you). Now take the derivative of each step response (these will be the impulse responses). If the step responses are monotonic this gives you two strictly-positive impulse responses. Scale the impulse responses so the total area under each impulse response is unity (i.e., these are the derivatives of unit steps). Measure the variance of each pulse, v1 and v2, and take the square root of the variances to get the standard deviations, s1 and s2. If the impulse responses are strictly positive it will always be true that v[overall] = v[1] + v[2] and also s[overall]^2 = s[1]^2 + s[2]^2 What goes wrong has to do with steps (2) and (3). The relation between standard deviation and "pulse width" has to do with how you define that width. For example, suppose you define the impulse-response width at the points comprising 10% and 90%, respectively, of the total area under the impulse response (corresponding to a 10-90% risetime definition). The relation between the standard deviation width measurement and the 10-90% width measurement now depends on the exact shape of the impulse response curve. Looking at the table on p. 401 of my book, a single-pole RC relaxation with a standard deviation of 1.00 produces a 10-90% risetime of 0.877. A gaussian impulse response with the same standard deviation of 1.00 produces a 10-90% risetime of 1.02. Same standard deviation, different 10-90% risetimes. If all three of your waveforms (the first system, the second system, and the cascade) have the same impulse-response shape, then all three will have the same relation between standard deviation and risetime, and the RSS formula works perfectly. This happens when the two sub-system components have a Gaussian response (albeit with different standard deviations). The result in this case always turns out to be Gaussian, and so the RSS formula applies exactly. A gaussian impulse response is the ONLY shape for which the impulse response of the cascade has the same shape as the impulse response of the subsystems. If the two subsystems are not Gaussian, then the shape of the output impulse response won't precisely match the shape of the impulse responses of the two subsystems (even if the two subsystems are identical), and the RSS approximation becomes no longer perfect. RSS is still somewhat useful, though. You can see in my example showing the numbers for an RC relaxation and a gaussian impulse response the extent of the imperfection in the relation between standard deviation and rise time. Thinking about cables, you can make a simlar argument in the frequency domain. When you cascade two sections of cable it doubles (in dB) the attenuation at all frequencies. Finding the new system risetime is related to asking this question, "Starting from the -3dB point on a single cable, how far do I have to back down in frequency to get to the 3dB point in the new cascaded system?" That question is exactly the same as this: "Starting from the -3dB point on a single cable, how far do I have to back down in frequency to get to the -1.5dB point on the same single cable?" Obviously, the answer to that has something to do with the shape of the frequency response of the cable. If the cable response is gaussian [H(f) = exp(-k*f*f)] then the answer is that you back down by 1/sqrt(2) , which corresponds precisely the RSS formula. CASES WHERE RSS DOESN'T WORK VERY WELL If the cable response is dominated by the dielectric-effect mechanism (as are many long pcb traces) then the response looks like [H(f)=exp(-k*f)], for which a doubling of the trace length requires that you back down by a factor of two in frequency. If the cable response is dominated by the skin-effect mechanism (as are many coaxial cables) then the response looks like [H(f)=exp(-k*sqrt(f))], for which a doubling of the cable length requires that you back down by a factor of four in frequency. It used to be that when working with oscilloscopes the responses of the probe and vertical amplifier were very nearly gaussian, and so the RSS formula worked beautifully. In modern scopes with bandwidths above 4 GHz it seems our scope manufacturers have taken to using digital means of "flattening" the response so it's no longer gaussian and the old RSS formula doesn't necessarily work any more. Agilent just did a good presentation of this at DesignCon. NOTE TO ANDREW, the imperfections in the RSS approximation for cables aren't due to the fact that the cable sections interact, it's just that the frequency response isn't exactly gaussian. You could hook buffer amplifiers between the sections and get the same result. Best regards, Dr. Howard Johnson, Signal Consulting Inc., tel +1 509-997-0505, howiej@xxxxxxxxxx http:\\sigcon.com -- High-Speed Digital Design articles, books, tools, and seminars -----Original Message----- From: si-list-bounce@xxxxxxxxxxxxx [mailto:si-list-bounce@xxxxxxxxxxxxx]On Behalf Of Ingraham, Andrew Sent: Monday, February 24, 2003 12:40 PM To: si-list@xxxxxxxxxxxxx Subject: [SI-LIST] Re: Rise-time of Cascaded Lossy T- Lines The RSS approximation of risetimes of cascaded elements is just that ... an approximation. It seems to work pretty well in most cases, but it is not exact. One assumption that I believe goes into this approximation, is that each cascaded section is effectively isolated from the others. In other words, connecting component B to component A doesn't affect the response of component A itself. The frequency response of the whole system would be the product of each individual section. But that is not the case when cascading passive elements like lossy t-lines without repeaters. I think this is where the assumption breaks down. > I've been reading Mr. Eric Bogatin's "Lossy Transmission Lines: Plain > and Simple" paper and have a question, which I hope those of you who > have also read it could answer (The same reasoning I present here also > follows, to my understanding, from Mr. Howard Johnson's "Black Magic" > book, App. B): > > > > On page 21 of his paper, Mr. Bogatin presents the > Root-of-Sum-of-Squares > equation for calculating the rise-time at the end of a lossy > transmission line, as a function of the driver rise-time and the 3-dB > rise-time of the interconnect. > > > > It seems from this paper as if one can cascade several interconnects > and > get the total effective rise-time of the whole interconnect by using > the > equation as follows: Take the individual 3dB-rise-times of the > individual interconnects and take the square-root of their sum of > squares. > > > > But it seems there is a problem with this use of the equation. One can > take a long interconnect, and calculate its effective rise-time using > 2 > methods: as an un-divided interconnect, or sub-divide it to (e.g.) 2 > identical halves, and do the arithmetic again. Using the 2nd method > one > gets a result which is (1/square-root of 2) of the "undivided" > calculation rise-time. > > > > This is a strange result, so I must be missing something, but what? > > > > Thanks for anyone who can help > > > > Itzhak Hirshtal > > Elta Systems > > Israel > > > > > > > ------------------------------------------------------------ ------ To unsubscribe from si-list: si-list-request@xxxxxxxxxxxxx with 'unsubscribe' in the Subject field or to administer your membership from a web page, go to: //www.freelists.org/webpage/si-list For help: si-list-request@xxxxxxxxxxxxx with 'help' in the Subject field List archives are viewable at: //www.freelists.org/archives/si-list or at our remote archives: http://groups.yahoo.com/group/si-list/messages Old (prior to June 6, 2001) list archives are viewable at: http://www.qsl.net/wb6tpu -- Binary/unsupported file stripped by Ecartis -- -- Type: application/ms-tnef -- File: winmail.dat ------------------------------------------------------------------ To unsubscribe from si-list: si-list-request@xxxxxxxxxxxxx with 'unsubscribe' in the Subject field or to administer your membership from a web page, go to: //www.freelists.org/webpage/si-list For help: si-list-request@xxxxxxxxxxxxx with 'help' in the Subject field List archives are viewable at: //www.freelists.org/archives/si-list or at our remote archives: http://groups.yahoo.com/group/si-list/messages Old (prior to June 6, 2001) list archives are viewable at: http://www.qsl.net/wb6tpu