3457a DC noise?

[rigrunner]

nn=10 samples may be too few for stable statistics.

What would be considered the minimum number of samples to give stable values?

[3roomlab]

@post#20 @ rigrunner
100mV range? it would appear to be stable @ slow setting (=NPLC10 + moving average 10 = modified NPLC100, real NPLC 100 = NPLC10 + repeating average 10). in NPLC 1, i guess last 2 digits are useless. however if i reference from alex's K2015 logs to compare with my original logs, i have to conclude it seems, this is normal = K2015 have somewhat high noise (ref : https://www.eevblog.com/forum/testgear/dmm-adc-noise-comparison-testing-project/150/ ... post#191)

@post21 @ onlooker
i tested this approach see pic chart @ the same DMM noise thread , see the rows with NPLC 10 vs 2x5 (https://www.eevblog.com/forum/testgear/dmm-adc-noise-comparison-testing-project/?action=dlattach;attach=184637;image)
also note that NPLC 2x5 is taken at 4500 samples/hr. see the averaged p-p noise is also quite similar. with this, i assumed that opening up a larger time window of data allows probability of collecting larger p-p noise. (= i have noisy environment)
p-p noise become more of a probability game, of when the circuit will experience a spike due to external or semiconductor "mini pop-corn" spike? again this is my assumption/observation

both types of NPLC give very similar result imo. maybe better test = measure same source, twice in diff NPLC way, but even then, it is still a different 2 samples. maybe i need to use a DMM in alternate universe  because at same instance it would be impossible to log 2 kinds of samples. or otherwise, mathematically, this will always give same result?

@ post 23 @ Le_bassiste
at a later log, i soldered jump wire inside the DMM input_HI to input_LO before it reaches the mechanical switch (short of cutting the PCB tracks lol). the noise does not seem to show major change. before this, with banana jack. the sockets are wrap with cotton. so with this i deviate back to conclusion, circuit noise > thermal noise.
this is where i got unsatisfied @ 10v range p-p noise, always exceeding 3uV.

imo you should use repeating average, so that 10 new readings is use to produce exactly 1 result. and then any of the 10 readings are not repeated, but fresh 10 readings are taken. moving average mix 1 new reading into previous 9 readings to produce 1 result. so each new result is not independantly averaged from last result.

@ post #25
in the resistor thread, there was a PDF about statistics. 100 samples is what i conclude ... but i end up using more than 1000  (i didnt know then that 100 could be a minimum)

***edit : haha today should be exciting, spare parts are coming in to re-energize the K2015. i be "noisy" with noise logs again

[onlooker]

i tested this approach see pic chart @ the same DMM noise thread...

Thanks for pointing me to this thread. I did see that topic before, but did not go through all the pages. I guess it is because 3457a was not one of its lineup.

What would be considered the minimum number of samples to give stable values?

I think the thread "www.eevblog.com/forum/testgear/dmm-adc-noise-comparison-testing-project"  that 3roomlab referenced had discussions on the number of samples they use to get stable statistics.

For my take, 1st,  one collects, say, 1000 of the nn-samples-averages. Then find the sdev (call it sigma2) of these 1000 nn-averages. Now compare this sdev with the average of the nn-samples sdev (call it sigma1). If sigma2 << sigma1, the nn-samples can be considered enough.

For random noise, the theory would say, for nn=10, sigma2 ~= 0.32 x sigma1. Whether a relative 32% jumpiness in the nn-samples-averages is stable or not is up to interpretation and, yes, is subjective just as many other things in statistics and probability.

Then, all the discussions are assuming one is not tracking systematic fluctuations or "long" term drifting and all the measurements are done under the same environmental conditions.

[rigrunner]

I think the thread "www.eevblog.com/forum/testgear/dmm-adc-noise-comparison-testing-project"  that 3roomlab referenced had discussions on the number of samples they use to get stable statistics.

I've bookmarked the thread. i'll read through it and perhaps add my readings to it.

For my take, 1st,  one collects, say, 1000 of the nn-samples-averages. Then find the sdev (call it sigma2) of these 1000 nn-averages. Now compare this sdev with the average of the nn-samples sdev (call it sigma1). If sigma2 << sigma1, the nn-samples can be considered enough.

For random noise, the theory would say, for nn=10, sigma2 ~= 0.32 x sigma1. Whether a relative 32% jumpiness in the nn-samples-averages is stable or not is up to interpretation and, yes, is subjective just as many other things in statistics and probability.

Then, all the discussions are assuming one is not tracking systematic fluctuations or "long" term drifting and all the measurements are done under the same environmental conditions.

Thanks for that. I'll see if i can automate that via gpib.