>how to get at u(k)? Can I run time backwards in some way so the H is >stable in reverse time?Flipping an impulse response in time will not change its absolute sum, therefore stability is independent of the "direction" of your time. Emre

Reply by ●September 8, 20092009-09-08

>how to get at u(k)? Can I run time backwards in some way so the H is >stable in reverse time?Flipping an impulse response in time will not change its absolute sum, therefore stability is independent of the "direction" of your time. Emre

Reply by ●September 8, 20092009-09-08

HardySpicer <gyansorova@gmail.com> wrote: < If I have a random signal u(k) and a (known) transfer function H then < y(k)=Hu(k) and if I know H and H is minimum phase then I ca neasily < find u(k). When the subject is Deconvolution, I always recommend Jansson's "Deconvolution of Images and Spectra." I had the first edition from the library some years ago. When I wanted to buy my own, I couldn't find one. It turned out that the second edition was about to come out, so I waited and bought that. You might also find one in a nearby engineering library. -- glen

Reply by ●September 7, 20092009-09-07

HardySpicer wrote:> If I have a random signal u(k) and a (known) transfer function H then > > y(k)=Hu(k) and if I know H and H is minimum phase then I ca neasily > find u(k). > > Suppose H is nonminimum phase eg > > > y(k)=u(k)-2u(k-1) > > how to get at u(k)? Can I run time backwards in some way so the H is > stable in reverse time?What is the goal of this exercise? Vladimir Vassilevsky DSP and Mixed Signal Design Consultant http://www.abvolt.com

Reply by ●September 7, 20092009-09-07

On 7 Sep, 01:56, HardySpicer <gyansor...@gmail.com> wrote:> If I have a random signal u(k) and a (known) transfer function H then > > y(k)=Hu(k) and if I know H and H is minimum phase then I ca neasily > find u(k). > > Suppose H is nonminimum phase eg > > y(k)=u(k)-2u(k-1) > > how to get at u(k)? Can I run time backwards in some way so the H is > stable in reverse time?In general you need to assign a prior probability to u. If the operator H is invertible and the noise is small, this will probably not make too much difference, but for large SNR and non-invertible operators it is essential. illywhacker;

Reply by ●September 7, 20092009-09-07

On 7 Sep, 01:56, HardySpicer <gyansor...@gmail.com> wrote:> If I have a random signal u(k) and a (known) transfer function H then > > y(k)=Hu(k) and if I know H and H is minimum phase then I ca neasily > find u(k). > > Suppose H is nonminimum phase eg > > y(k)=u(k)-2u(k-1) > > how to get at u(k)? Can I run time backwards in some way so the H is > stable in reverse time?Deconvolution is among the hardest quantitative problems in DSP. A minimum phase transfer function H(z) can easily be inverted in frequency domain since 1/H(z) will be causal and stable. A non-minimum phase H(z) is in the worst case impossible to invert: Zeros on the unit circle can not be 'undone'. As for mixed-phase filters, it's anybody's guess how to first of all - in the case of MA models estimated from autocovariance sequences - identify what roots are inside and outside the unit circle, and then merge the causal and anticausal signals to form a useful deconvolved signal. And of course, most of the above are problems one encounter even when the system function H(z) is known. In the case of unknown H(z) you also get all kinds of model mismatch problems, and so on and so forth. Rune

Reply by ●September 6, 20092009-09-06

On Sun, 06 Sep 2009 16:56:29 -0700, HardySpicer wrote:> If I have a random signal u(k) and a (known) transfer function H then > > y(k)=Hu(k) and if I know H and H is minimum phase then I ca neasily find > u(k). > > Suppose H is nonminimum phase eg > > > y(k)=u(k)-2u(k-1) > > how to get at u(k)? Can I run time backwards in some way so the H is > stable in reverse time? > > hardyYes, but. For the case you present the transfer function will be stable in 'reverse time', and in general a transfer function with zeros that are strictly inside and outside of the stability regions can be separated into parts that are stable in forward and reverse time, respectively. But consider the case of a transfer function with a zero or more right on the stability boundary, i.e. y(k) = u(k) + u(k-1), or y(k) = u(k) - 2 b u(k-1) + u(k-2), with |b| <= 1, or y(k) = u(k) + u(k - n), with n > 0. In all of these cases you will have to assume a starting state for u(k), and if you add any noise at all then the inverse transfer function's response to the noise will be infinite. Even if your forward transfer function just has zeros that are relatively _close_ to the stability boundary then you're in the same pickle, as your inverse transfer function may _theoretically_ have a finite noise response, but in practice it may well be too big to be worthwhile. -- www.wescottdesign.com

Reply by ●September 6, 20092009-09-06