If y = conv(x,w) What is the derivative of y with respect to w ? Here y is the convolution between x and w, all are one dimensional. Thanks This message was sent using the Comp.DSP web interface on www.DSPRelated.com
Derivative of convolution
Started by ●April 1, 2005
Reply by ●April 1, 20052005-04-01
"reju_vg" <reju_vg@yahoo.com> writes:> If y = conv(x,w) > What is the derivative of y with respect to w ? > Here y is the convolution between x and w, all are one dimensional. > > ThanksIf y = f(w(t)), then by the chain rule for differentiation, dy / dt = (dy / dw) * (dw / dt). Then just solve for dy / dw: dy / dw = (dy / dt) / (dw / dt). Now that I've answered this homework question for you, why don't you think twice next time about bypassing the opportunity to think out your own solution? -- Randy Yates Sony Ericsson Mobile Communications Research Triangle Park, NC, USA randy.yates@sonyericsson.com, 919-472-1124
Reply by ●April 1, 20052005-04-01
Hello Reju, Before you can have a derivative, I have to ask, are y,x,w discrete time sequences? If they are, then a derivative doesn't make sense unless you assume some kind of intepolation function is being used. In this case write each of x and w as convolution sums (I.e., discrete convolution with the interpolating function - a sinc() perhaps). Thus you now have two continuous functions x_hat and w_hat. Put these two functions in a convolution integral and then find the derivative with respect to w_hat. Finally sample the result to get back to a discrete sequence. Is that what you needed? Clay
Reply by ●April 1, 20052005-04-01
Clay wrote:> Hello Reju, > > Before you can have a derivative, I have to ask, are y,x,w discrete > time sequences? If they are, then a derivative doesn't make sense > unless you assume some kind of intepolation function is being used.A time derivative makes no sense but wrt the coefficients, yes, they do make sense. How do yo derive the LMS algorithm without being able to take a derivative wrt the filter coefficients?> > In this case write each of x and w as convolution sums (I.e., discrete > convolution with the interpolating function - a sinc() perhaps). Thus > you now have two continuous functions x_hat and w_hat. Put these two > functions in a convolution integral and then find the derivative with > respect to w_hat. Finally sample the result to get back to a discrete > sequence. Is that what you needed? > > Clay >
Reply by ●April 1, 20052005-04-01
Stan Pawlukiewicz wrote:> Clay wrote: > > Hello Reju, > > > > Before you can have a derivative, I have to ask, are y,x,w discrete > > time sequences? If they are, then a derivative doesn't make sense > > unless you assume some kind of intepolation function is being used. > > A time derivative makes no sense but wrt the coefficients, yes, theydo> make sense. How do yo derive the LMS algorithm without being able to> take a derivative wrt the filter coefficients? > > > >Stan, In the LMS case there is a continuous error function. Perhaps Reju can yield moro information about his problem. I.e., is everything discrete or continuous? Since little info was given, I have to guess at what he has and wants. Clay
Reply by ●April 4, 20052005-04-04
>Hi All,Thank you for the reply. Here the signals are discrete time sequence. To clarify the situation let me write part of my problem here. I want to differentiate the equation E{(x1-y2*w21)(x2-y1*w12) with respect to w21. Here E{.}is the expectation operator and y2*w21 means the convolution between y2 and w21. x1, x2, y1, and y2 are the discrete time speech signals and w12 and w21 are the impulse responses of a room. Thanks Reju>Stan Pawlukiewicz wrote: >> Clay wrote: >> > Hello Reju, >> > >> > Before you can have a derivative, I have to ask, are y,x,w discrete >> > time sequences? If they are, then a derivative doesn't make sense >> > unless you assume some kind of intepolation function is being used. >> >> A time derivative makes no sense but wrt the coefficients, yes, they >do >> make sense. How do yo derive the LMS algorithm without being able to > >> take a derivative wrt the filter coefficients? >> > >> > > >Stan, > >In the LMS case there is a continuous error function. Perhaps Reju can >yield moro information about his problem. I.e., is everything discrete >or continuous? Since little info was given, I have to guess at what he >has and wants. > >Clay > >This message was sent using the Comp.DSP web interface on www.DSPRelated.com