Are there any DSP benefits of choosing a random seed for an MLS?

What about Kasami sequences? Could these also be used to determine the 
impulse response of a linear system?

Nicholas Kinar wrote:


> Does the Gold code have similar properties to the FHT, in that the SNR 
> is high?  How would I process a Gold code to obtain the frequency 
> response of the system?

1. What are you trying to accomplish? What is the goal?

2. Keep in mind that the impulse response of a room will be entirely 
different if you shift a speaker or a mike by 1 ft or so.

3. Why do you need to mess with MLS or Gold codes? Any random-like 
sequence would do, as well as a frequency sweep or a multitone signal.

4. Aside from the direct measurement, there is plenty of methods to get 
the impulse response.

VLV

> 
> MLS is fine if your device-under-test (DUT) is perfectly LTI.
> 
> it turns out that non-linearities in the DUT cause the apparent
> impulse response to have spurious spikes that are at fixed locations
> in the impulse response (which means you can rerun the test average
> until the cows come home and those spurious spikes ain't going away).
> i discuss this briefly at the end of http://www.dspguru.com/info/tutor/mls2.htm
> .

That's a great tutorial, Robert.  I've learned more from that tutorial 
than many papers dealing with DSP in acoustics.

> 
> different MLS's generated by different primitive polynomials will put
> those spikes at different places in the impulse response.  you can
> pick out 3 (or a bigger odd number) different primitive polynomials,
> run your MLS test 3 different times with the 3 different polynomials,
> line up your 3 different impulse responses on top of each other and,
> point-by-point, pick out the median for each sample.  it's unlikely,
> for the same nonlinearity, that the 3 different tests will spike at
> exactly the same place.  this idea comes from or is motivated by a
> paper from an interesting person, Paul Kovitz, whom i haven't heard
> much from for a decade.
> 
> r b-j

The notion of using different primitive polynomials is a great idea, 
Robert.  Hats off to you for posting it here.

I am assuming that the different primitive polynomials all have the same 
(2^N - 1) length. Would I generate each primitive polynomial using 
different feedback taps?

What is the rationale behind picking the median for each sample?

Hi Vladimir,

Thank you for your response.

> 
> 1. What are you trying to accomplish? What is the goal?
> 

To determine the impulse response of an acoustic space using an embedded 
processing system.

> 2. Keep in mind that the impulse response of a room will be entirely 
> different if you shift a speaker or a mike by 1 ft or so.

Of course!  This is good advice.

> 
> 3. Why do you need to mess with MLS or Gold codes? Any random-like 
> sequence would do, as well as a frequency sweep or a multitone signal.
> 

Injecting MLS into a LTI system and determining the impulse response has 
   extremely high SNR.

http://purebits.com/mlsteo.html

The time taken to use an MLS is less than that of a frequency sweep. 
This is the reason why it is popular in acoustics signal processing 
(perhaps not as popular in mainstream DSP).

The MLS has extremely high SNR, compared to other types of stimuli.

http://www.libinst.com/mlsmeas.htm


> 4. Aside from the direct measurement, there is plenty of methods to get 
> the impulse response.


Yes, I've tried frequency sweeps and other types of broadband noise. 
However, in acoustics, MLS measurements are considered to be better due 
to higher SNR.

I am just wondering if there are any benefits of using a random seed.

Nicholas Kinar wrote:
> 
>>
>> MLS is fine if your device-under-test (DUT) is perfectly LTI.
>>
>> it turns out that non-linearities in the DUT cause the apparent
>> impulse response to have spurious spikes that are at fixed locations
>> in the impulse response (which means you can rerun the test average
>> until the cows come home and those spurious spikes ain't going away).
>> i discuss this briefly at the end of 
>> http://www.dspguru.com/info/tutor/mls2.htm
>> .
> 
> That's a great tutorial, Robert.  I've learned more from that tutorial 
> than many papers dealing with DSP in acoustics.
> 
>>
>> different MLS's generated by different primitive polynomials will put
>> those spikes at different places in the impulse response.  you can
>> pick out 3 (or a bigger odd number) different primitive polynomials,
>> run your MLS test 3 different times with the 3 different polynomials,
>> line up your 3 different impulse responses on top of each other and,
>> point-by-point, pick out the median for each sample.  it's unlikely,
>> for the same nonlinearity, that the 3 different tests will spike at
>> exactly the same place.  this idea comes from or is motivated by a
>> paper from an interesting person, Paul Kovitz, whom i haven't heard
>> much from for a decade.
>>
>> r b-j
> 
> The notion of using different primitive polynomials is a great idea, 
> Robert.  Hats off to you for posting it here.
> 
> I am assuming that the different primitive polynomials all have the same 
> (2^N - 1) length. Would I generate each primitive polynomial using 
> different feedback taps?

The same length isn't very important. If the same length, then different 
taps. Not all lengths have more than one MLS tap sets.

> What is the rationale behind picking the median for each sample?

If one measurement is different and the two others are the same, the 
median discards the oddball. (I bet you knes that!)

Jerry
-- 
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������

> 
> The same length isn't very important. If the same length, then different 
> taps. Not all lengths have more than one MLS tap sets.

Okay, this makes sense.

> 
>> What is the rationale behind picking the median for each sample?
> 
> If one measurement is different and the two others are the same, the 
> median discards the oddball. (I bet you knes that!)
> 

Oh of course.  Where's my brain this morning?

On Aug 25, 12:51&#4294967295;pm, Nicholas Kinar <n.ki...@usask.ca> wrote:
> Nicholas Kinar wrote:
>
> > Thanks for your response, Lasse. &#4294967295;Yes, I agree that Gold codes have very
> > little overlap, but I don't know if they could be used in lieu of an
> > MLS. &#4294967295;More research would be required, I guess.
>


Maximum-length sequences have (nearly) ideal periodic autocorrelation
functions, and this fact is exploited in the technique used to
determine
the impulse response of a system.  Gold sequences and Kasami
sequences have non-ideal autocorrelation functions with several small
peaks in the out-of-phase autocorrelation.  Using them to determine
the
impulse response effectively gives the impulse response *plus* several
ghost copies of the impulse response, as if we were looking at the
impulse response transmitted over a multipath channel with several
taps.  It might be possible to separate the wheat from the chaff in
such
cases.  Whether the effort is worthwhile or not may require more
research.

If two different maximum-length sequences are used at the same time,
then
each measurement of the impulse response will contain ghost images
due to the crosscorrelation between the two sequences.  And no, it is
*not* possible for a pair of sequences to have ideal autocorrelation
and
zero crosscorrelation.

>I believe that the Gold code has a length of (2^N + 1)

Gold codes have length 2^N - 1, same as the maximum-length
sequences.  There are 2^N + 1 different Gold sequences, two
of which happen to be maximum-length sequences.

Hope this helps.

Dilip Sarwate

> 
> Maximum-length sequences have (nearly) ideal periodic autocorrelation
> functions, and this fact is exploited in the technique used to
> determine
> the impulse response of a system.  Gold sequences and Kasami
> sequences have non-ideal autocorrelation functions with several small
> peaks in the out-of-phase autocorrelation.  Using them to determine
> the
> impulse response effectively gives the impulse response *plus* several
> ghost copies of the impulse response, as if we were looking at the
> impulse response transmitted over a multipath channel with several
> taps.  It might be possible to separate the wheat from the chaff in
> such
> cases.  Whether the effort is worthwhile or not may require more
> research.
> 

I'll check it out, thanks, Dilip.  It is interesting to know about the 
nearly ideal periodic autocorrelation functions.

> 
> Gold codes have length 2^N - 1, same as the maximum-length
> sequences.  There are 2^N + 1 different Gold sequences, two
> of which happen to be maximum-length sequences.
> 
> Hope this helps.
> 
> Dilip Sarwate
> 

Yes it does.  This newsgroup always helps to put me on the right track.