Author Topic: Root Sum Squared Error Analysis  (Read 14617 times)

0 Members and 2 Guests are viewing this topic.

Offline moffy

  • Super Contributor
  • ***
  • Posts: 3091
  • Country: au
Re: Root Sum Squared Error Analysis
« Reply #25 on: May 02, 2025, 10:37:34 pm »
Let me offer an example of something I can't prove, but very strongly suspect. I suspect electrolytic capacitor manufacturers can now control their process to a much higher precision than their standard tolerance. Since they can do that, they save a lot of money on etched foil by building all their capacitors on the low side of nominal. You very rarely, if ever, find one at or above the nominal value. At least this has been my experience with higher values for the last decade or so.
I've noticed that also, some can be slightly higher than nominal but only marginally.
 

Offline bdunham7

  • Super Contributor
  • ***
  • Posts: 9452
  • Country: us
Re: Root Sum Squared Error Analysis
« Reply #26 on: May 02, 2025, 11:56:11 pm »
1/ What distribution curve does the RSS assume for the resistor values?  What justification is there for that answer?

2/ Does the RSS assume some resistors will be outside the specified tolerance? If so, how many?

3/ The RSS calculation produces one number as its answer.  What grounds are there for assuming any particular distribution curve around that value?

When you can answer those fundamental questions we'll be getting somewhere. Otherwise we are vigorously debating something we have no data for: the distribution curve of resistor values. Probably they differ from one manufacturer to another, and probably they differ over time as new manufacturing processes are adopted.

The idea is that if you have two uncorrelated sources of error and both have a normal distribution, then when you combine them (say put two resistors in series) you can take the root of the sum of the squares (RSS) of the two standard deviations and that will be the standard deviation of the error of the combined unit.  Note that these errors should be expressed in absolute terms, not percentages.  You need set up the problem logically, so for example if you are using parallel resistors, you'd convert to Siemens or mhos first.  Also, if you use a number that is a multiple of the std devs, say k=5, as long as you're consistent this multiple will fall out of the math and you'll get a k=5 result.  So assuming this is what we mean:

1/ a normal distribution and there's no justification given for the assumption other than RSS is a math operation that works on std devs.

2/ yes, as explained in a previous post it depends on how many std devs the tolerance stated represents but a normal distribution always implies the possiblity of some samples being beyond any fixed limit

3/ if the inputs are normal distributions, the output will be also--that's just math, but for a mathematical proof you'll have to look it up somewhere.  If you apply RSS to other distributions I don't think you get a useful result.  If you take two rectangular (uniform) distributions and combine them you should get a triangular result, but IDK how you'd even assign a single number to that let alone calculate it. 
« Last Edit: May 03, 2025, 12:23:34 am by bdunham7 »
A 3.5 digit 4.5 digit 5 digit 5.5 digit 6.5 digit 7.5 digit DMM is good enough for most people.
 
The following users thanked this post: SteveThackery

Offline SteveThackery

  • Super Contributor
  • ***
  • Posts: 3463
  • Country: gb
  • 50 year novice
Re: Root Sum Squared Error Analysis
« Reply #27 on: May 03, 2025, 11:59:41 am »
Thanks for a great reply, @bdunham7.

I think I can reasonably claim the following:

Unless you know that the resistor values follow a normal distribution curve, the output from the RSS calculation is not meaningful.

I don't think we can always assume the distribution curve is "normal". If resistors are manufactured to a specific value and target tolerance, and then not tested, then of course it is absolutely reasonable to assume their values should follow a normal distribution curve around the nominal value.

However, I seriously doubt* that resistors could be manufactured to (say) 0.1% tolerance such that no testing is required. I think a test-and-select process would be necessary. Perhaps that is true for wider tolerances - including 1% resistors? I don't know.  It is very difficult to find reliable information.

*I know that "seriously doubt" doesn't make a good argument, but in support of that position I can describe how resistors used to be made. Back in the day, all resistors were tested after manufacture and assigned to tolerance bins. That meant two things:

1/ With the exception of the tightest tolerance band, no other tolerance band had a normal distribution. In fact they were anything but: the distribution curve had a big hole in the middle because those resistors had already been assigned to a tighter tolerance bin.

2/ It really was the case that the assigned tolerance band acted as a "brick wall" (that is, no tails outside the band, like a normal distribution), because those resistors would be assigned to the next wider tolerance bin.

If any element of that tolerance binning is still in use (which I dont know, but strongly suspect is the case), then the assumptions inherent in RSS (a normal distribution) are wrong and RSS does not produce a meaningful result.
« Last Edit: May 03, 2025, 12:05:58 pm by SteveThackery »
If it ain't broke, keep fixing it until it is.
 

Online Siwastaja

  • Super Contributor
  • ***
  • Posts: 11457
  • Country: fi
Re: Root Sum Squared Error Analysis
« Reply #28 on: May 03, 2025, 05:40:47 pm »
Unless you know that the resistor values follow a normal distribution curve, the output from the RSS calculation is not meaningful.

Yes. And more specifically, we know from practical experience (e.g. see above by Victorman222, similar datasets have been posted by many others too) that the opposite is true: that resistor values almost never follow a normal distribution curve, not even close, and even if the shape is that of normal distribution, it would not be centered at the nominal resistance but some unknown mean instead.
 

Offline gf

  • Super Contributor
  • ***
  • Posts: 1838
  • Country: de
Re: Root Sum Squared Error Analysis
« Reply #29 on: May 03, 2025, 06:50:17 pm »
Unless you know that the resistor values follow a normal distribution curve, the output from the RSS calculation is not meaningful.

The RSS calculation applies to any probability distribution, but it tells how the standard deviations of two random variables add up (provided that the samples are uncorrelated):

    stdev(a+b) = sqrt( stdev(a)^2 + stdev(b)^2 )
    or equivalently, variance(a+b) = variance(a) + variance(b)

It does not tell you how "tolerances" add up if "tolerance" is not defined as the standard deviation, but in some other way.
In the special case where the distributions of a and b happen to be normal, the calculation also applies to quantiles.
 

Offline Tation

  • Frequent Contributor
  • **
  • Posts: 316
  • Country: pt
Re: Root Sum Squared Error Analysis
« Reply #30 on: May 04, 2025, 10:17:07 am »
So please answer these questions:

1/ What distribution curve does the RSS assume for the resistor values?  What justification is there for that answer?

2/ Does the RSS assume some resistors will be outside the specified tolerance? If so, how many?

3/ The RSS calculation produces one number as its answer.  What grounds are there for assuming any particular distribution curve around that value?

When you can answer those fundamental questions we'll be getting somewhere. Otherwise we are vigorously debating something we have no data for: the distribution curve of resistor values. Probably they differ from one manufacturer to another, and probably they differ over time as new manufacturing processes are adopted.


The application of the RSS method to other industries (e. g.: mechanics) predates its eventual application to electronics, so such questions were answered many years ago.

The main point of RSS is not assuming that the components distribute normally (although it is a common assumption with theoretical support, more on this later), but that the values of different components are uncorrelated.

E. g.: suppose that 10 pieces of some assembly are mounted, in the final product, stacked. And then they must fit inside some cavity. Say there is a 1 mm tolerance to fit the stack. A worst-case analysis gives 1/10 mm tolerance for each piece. But it is extremely improbable for 10 different pieces, of different shapes and manufactured by different machines, to be selected in a way that all 10 are in the upper edge of their tolerance (here is where no-correlation/independence is expected and assumed). Thus, maybe, we can work with more ample tolerances for the pieces and still fit the cavity, in most cases.

With numbers, working with a tolerance of 3-sigmas (99.73 % of manufactured pieces are inside stated tolerance of 0.1 mm, more on this later), then the probability of all 10 pieces being outside their upper tolerance limit is ~2e-27 %, truly negligible and, surely, all assembled stacks will fit the cavity.

But, if the tolerances of each piece are wider,
0.368 mm instead of 0.1 mm, the analysis (assuming independence and normality) tells us that, still, 99.73 % of the assemblies (3-sigmas) will fit the cavity. This yield may be enough and there is an obvious advantage in manufacturing the pieces with a x3.7 wider tolerance.

But in electronics, scenarios like this are not common. Can anybody show a realistic situation when some components are added and the final combination must fit a given error bound?

Why normality?

Because it is known that the sum of many random, uncorrelated, variables, no matter their distributions, tends to converge to a normal distribution. In the previous example, it is not needed for the dimensions of each piece to distribute normally, but the dimension of the 10-pieces stack will do. In case the dimensions of each piece distribute uniformly, the dimension of the 10-pieces stack will distribute according to the Irwin-Hall distribution with n=10. The figure below compares this distribution with a normal one with same mean and variance (not having a PC here, so asked ChatGPT to generate the plot. Hope it has not screwed it up...). As seen, it is hardly distinguishable from a Gaussian. Thus, the result of RSS, if there is a sufficient number of combined components, can be assumed to be normal. The mathematical foundation for this is the central limit theorem.

Thus, if some industrial process is influenced by many random variables that, in some way, combine additively, the result of such process should resemble a normal distribution. If there is no evidence against this, assuming normality is usually safe. This way, the machine producing resistors will surely produce resistors whose resistance distributes normally. But if, afterwards, measures them, bins them, and also rejects those out of tolerance, then they will no longer distribute normally.

RSS can still be used to compute the tolerance of the combination of various pieces even when they do not distribute normally (but are, still, non-correlated) by means of the so-called "inflation coefficients". Such coefficients multiply each component tolerance (prior to squaring) and depend on how such component value distributes. This coefficient for the case of normal distributions equals 1, and that for uniform distributions is sqrt(3).

Tolerance and sigma

Given than the output of the process uses to be normal, there's a non-zero probability of some manufactured pieces to be out of tolerance. The fraction of products out of tolerance may be expressed in sigmas, meaning how many standard deviations fit inside the tolerance limits. The fractions of out of tolerance components are {31.7, 4.5, 0.27, 6.8e-4, 2e-7} % for {1, 2, 3, 4.5, 6} sigmas. Such out of tolerance components may be rejected by the manufacturer, sold as lower grade parts or sent undetected (or detected) to clients.

« Last Edit: May 04, 2025, 11:10:55 am by Tation »
 
The following users thanked this post: eugene, SteveThackery

Offline bson

  • Supporter
  • ****
  • Posts: 2823
  • Country: us
Re: Root Sum Squared Error Analysis
« Reply #31 on: May 05, 2025, 03:26:31 am »
The RSS calculation applies to any probability distribution, but it tells how the standard deviations of two random variables add up (provided that the samples are uncorrelated):

    stdev(a+b) = sqrt( stdev(a)^2 + stdev(b)^2 )
    or equivalently, variance(a+b) = variance(a) + variance(b)
This is only true if they have the same mean.

Otherwise, measure 100x 100 ohm 1% resistors and 100x 10 ohm 1% resistors, and try to add their standard deviations.  Clearly that's not meaningful.  Combined as a single distribution they would range from ~10 to ~100, which a HUGE standard deviation, and a mean around ~55.  The only way you can combine their std devs is by first recalculating the combined mean, then recalculating their std devs around that mean.  (Which I'm not entirely sure is meaningful either, but I could possibly be convinced oherwise... I mean, meaningful, as in combined mean.  Pun not intended.)
« Last Edit: May 05, 2025, 03:30:31 am by bson »
 

Online ledtester

  • Super Contributor
  • ***
  • Posts: 4270
  • Country: us
Re: Root Sum Squared Error Analysis
« Reply #32 on: May 05, 2025, 04:50:23 am »
Otherwise, measure 100x 100 ohm 1% resistors and 100x 10 ohm 1% resistors, and try to add their standard deviations.  Clearly that's not meaningful.  Combined as a single distribution they would range from ~10 to ~100, which a HUGE standard deviation, and a mean around ~55. ...

What you seem to be describing is a sample space consisting of only 200 values: the 100 "100R 1% resistors"  together with the 100 "10R 1% resistors", and if you take the average value of those 200 values you get something around 55R.

However, we're talking about the space of all possible pairs of resistors, one being a 100R and the other being a 10R for a total of 10,000 possible outcomes. If you take the average value of all of those pairs you get something around 110R.

Quote
The only way you can combine their std devs is by first recalculating the combined mean, then recalculating their std devs around that mean.

This is indeed more like what is going on.
 

Offline bdunham7

  • Super Contributor
  • ***
  • Posts: 9452
  • Country: us
Re: Root Sum Squared Error Analysis
« Reply #33 on: May 05, 2025, 04:56:22 am »
This is only true if they have the same mean.

No, you wouldn't even have to know what the actual values or the mean was to do that calculation.  And the two values of resistors would be separate, they would be your A and B groups.  The (A+B) refers to a unit consisting of one A and one B, in this case in series.
A 3.5 digit 4.5 digit 5 digit 5.5 digit 6.5 digit 7.5 digit DMM is good enough for most people.
 
The following users thanked this post: gf

Offline youngda9

  • Regular Contributor
  • *
  • Posts: 70
  • Country: us
Re: Root Sum Squared Error Analysis
« Reply #34 on: May 05, 2025, 11:31:23 am »
For component WCA, you need to take into account four factors:
- BOL Tolerance
- Temperature Variation
- Age Variation
- Radiation Variation - only needed for space-bound components.

So you have three error sources.
- BOL Tolerance is your 1% that you've chosen.
- Temperature Variation is defined by the Temperature Coefficient of the resistor (100ppm/DegC, 12ppm/DegC, etc.).  You need to calculate the operational temperature range of the resistor in order to factor this into the equaiton.
- Age Variation, you'll have to look this up to find the correct equations...it's been a while and I'm a little rusty with this.
 

Online Siwastaja

  • Super Contributor
  • ***
  • Posts: 11457
  • Country: fi
Re: Root Sum Squared Error Analysis
« Reply #35 on: May 05, 2025, 11:50:08 am »
For component WCA, you need to take into account four factors:
- BOL Tolerance

Please define rare/special acronyms you use. I have been doing this for 25 years and never heard the terms WCA and BOL Tolerance. Even Googling for "BOL Tolerance" does not result a single useful result. WCA is ungooglable as well. This isn't some secret club, write out the words you want to convey.
 
The following users thanked this post: Mickle T.

Offline MisterHeadache

  • Regular Contributor
  • *
  • Posts: 158
  • Country: us
    • Level UP EE Lab
Re: Root Sum Squared Error Analysis
« Reply #36 on: May 05, 2025, 04:30:56 pm »
I'm pretty sure that the RSS calculation is inappropriate for use on a parallel combination of resistances because of another assumption - that the normal effect of the tolerances is just simple arithmetic addition.  In all the literature that I am familiar with (and I do teach introductory tolerance stackups in my company including statistical modeling techniques like RSS), the fundamental mathematics behind how the RSS formula is derived is implicitly based on that.

We also occasionally do Monte Carlo statistical stackups, which are appropriate to use on this problem and might shed some insight.  I ran several 10,000 run simulations of a 100 ohm and a 10 ohm resistor in parallel, each with a tolerance of 1% and each having a centered normal distribution with a standard deviation of 1/3 of their tolerance (those are assumptions behind the scenes of a traditional RSS calcuation by the way). The results spit out consistently the correct result for the resistance combination (9.09... ohms) and a standard deviation that ranged from 0.0272 to 0.0280.  FWIW...
Daryn from Level Up EE Lab
AKA 'MisterHeadache'
 

Offline Tation

  • Frequent Contributor
  • **
  • Posts: 316
  • Country: pt
Re: Root Sum Squared Error Analysis
« Reply #37 on: May 06, 2025, 06:33:01 am »
I'm pretty sure that the RSS calculation is inappropriate for use on a parallel combination of resistances because of another assumption - that the normal effect of the tolerances is just simple arithmetic addition.  In all the literature that I am familiar with (and I do teach introductory tolerance stackups in my company including statistical modeling techniques like RSS), the fundamental mathematics behind how the RSS formula is derived is implicitly based on that.

We also occasionally do Monte Carlo statistical stackups, which are appropriate to use on this problem and might shed some insight.  I ran several 10,000 run simulations of a 100 ohm and a 10 ohm resistor in parallel, each with a tolerance of 1% and each having a centered normal distribution with a standard deviation of 1/3 of their tolerance (those are assumptions behind the scenes of a traditional RSS calcuation by the way). The results spit out consistently the correct result for the resistance combination (9.09... ohms) and a standard deviation that ranged from 0.0272 to 0.0280.  FWIW...

Certainly RSS applies to sums of random variables, thus to series of resistors.

But, if a random variable distributes normally and its standard deviation is much smaller than its mean (that is: its percent tolerance is small), then the inverse of such random variable does also distribute as gaussian, with mean equal to the inverse of the mean of the original variable and maintaining the same ratio of deviation to mean (thus maintaining the percent tolerance).

This makes RSS extensible to parallel combinations of resistors by working with percent tolerances, as the parallel is the inverse of the sum of inverses.

With 100 & 10 Ohm 1 % resistors, the tolerance of their combination (series or parallel) is 0.913 %. In parallel this gives +-83 mOhm and, thus, a deviation of 0.02768 Ohm, agreeing with your Monte Carlo analysis.
 
The following users thanked this post: MisterHeadache

Offline edbaTopic starter

  • Contributor
  • Posts: 32
  • Country: gb
Re: Root Sum Squared Error Analysis
« Reply #38 on: May 06, 2025, 08:32:47 am »
Hi everyone,

Have a read of this article it is very interesting on the subject of tolerance analysis:

https://www.edn.com/for-error-analysis-its-hip-to-be-squared/

Brgds

Edba
 
The following users thanked this post: ledtester, SteveThackery

Offline SteveThackery

  • Super Contributor
  • ***
  • Posts: 3463
  • Country: gb
  • 50 year novice
Re: Root Sum Squared Error Analysis
« Reply #39 on: May 06, 2025, 09:11:01 am »
Hi everyone,

Have a read of this article it is very interesting on the subject of tolerance analysis:

https://www.edn.com/for-error-analysis-its-hip-to-be-squared/


Good arricle - thank you.
If it ain't broke, keep fixing it until it is.
 

Offline edbaTopic starter

  • Contributor
  • Posts: 32
  • Country: gb
Re: Root Sum Squared Error Analysis
« Reply #40 on: May 06, 2025, 10:50:32 am »
It covers the case where components are selected through sorting into bins. A 1% resistor would then have a uniform distribution i.e. equal probability of selecting any value between +/-1%. If you use the RSS technique you multiply the tolerance by 1.73, so the resistor tolerance converts to +/-1.73% resistor. I have done a comparison of normal and uniform distribution below:

Normal
100 Ω + 10 Ω ±√(1 Ω^2) + (0.1 Ω^2) = 110 Ω ±1.0050 Ω

Uniform
100 Ω + 10 Ω ±√(1.73 Ω^2) + (0.173 Ω^2) = 110 Ω ±1.5806 Ω

Brgds

Edba
 

Offline SteveThackery

  • Super Contributor
  • ***
  • Posts: 3463
  • Country: gb
  • 50 year novice
Re: Root Sum Squared Error Analysis
« Reply #41 on: May 06, 2025, 07:56:03 pm »
It covers the case where components are selected through sorting into bins. A 1% resistor would then have a uniform distribution i.e. equal probability of selecting any value between +/-1%.

I'm afraid that's a faulty assumption, too. When resistors are binned you get a big hole in the middle of the distribution because those resistors have already been assigned to the next higher tolerance bin.

Assuming your tightest tolerance is 0.1%, that bin will have a normal distribution, but chopped off at the ends so none of them exceed 0.1% away from the nominal value. The next bin - say 0.2% - will have a hole in the range from -0.1% to +0.1%, so it will look nothing remotely like a normal distribution. The 0.5% bin will have a hole from -0.2% to +0.2%, and so on down the tolerances.

It's not looking good for RSS.
If it ain't broke, keep fixing it until it is.
 

Offline eugene

  • Frequent Contributor
  • **
  • Posts: 602
  • Country: us
Re: Root Sum Squared Error Analysis
« Reply #42 on: May 07, 2025, 02:10:01 pm »
When resistors are binned you get a big hole in the middle of the distribution because those resistors have already been assigned to the next higher tolerance bin.

That's a faulty assumption.
90% of quoted statistics are fictional
 

Offline SteveThackery

  • Super Contributor
  • ***
  • Posts: 3463
  • Country: gb
  • 50 year novice
Re: Root Sum Squared Error Analysis
« Reply #43 on: May 07, 2025, 07:06:09 pm »
When resistors are binned you get a big hole in the middle of the distribution because those resistors have already been assigned to the next higher tolerance bin.

That's a faulty assumption.

It's not an assumption. It's how binning works (but see later): once a resistor is in the 0.1% bin it isn't available to go in the 0.2% bin. If all the 0.1% resistors are taken out of the population requiring sorting, then you know for sure that no 0.1% resistors will be in the 0.2% bin. Hence the hole in the middle of the 0.2% distribution.

Two points: years ago all resistors were sorted into tolerance bins after production. This isnt controversial - it was widely known amongst the EE community. I was taught it in college. However, I don't know know if that is true these days. I think it almost certainly will be for the very tight tolerance parts, in which case you'll inevitably get the two-humped distribution in the bins below the tightest tolerance.  I suspect for wider-tolerance components, modern-day manufacturing is probably accurate enough to not require test-and-bin sorting, but simply to rely on three-sigma (or better) accuracy in production. In that case binning isn't required, and you'll get normal distributions.

Second point: if binning is used, but more resistors end up in the 0.1% bin than are required, then instead of discarding the surplus it would make sense to throw them in with the 0.2% resistors. And so on down the train. In that case you won't get a hole in the 0.2% distribution. It might not end up as a "normal" distribution, but somewhere in between.  Or perhaps even a large peak, depending on how many resistors get "demoted". I've no idea if this is actually done - it's a theoretical possibility.

The problem we face is that it seems to be extremely difficult to get a reliable source of information on whether binning is actually used, and if so how widely. Apart from a few instances reported earlier, the manufacturers don't seem to publish much about this topic.
If it ain't broke, keep fixing it until it is.
 

Offline bdunham7

  • Super Contributor
  • ***
  • Posts: 9452
  • Country: us
Re: Root Sum Squared Error Analysis
« Reply #44 on: May 07, 2025, 08:00:16 pm »
A 3.5 digit 4.5 digit 5 digit 5.5 digit 6.5 digit 7.5 digit DMM is good enough for most people.
 
The following users thanked this post: Mickle T.

Offline SteveThackery

  • Super Contributor
  • ***
  • Posts: 3463
  • Country: gb
  • 50 year novice
Re: Root Sum Squared Error Analysis
« Reply #45 on: May 08, 2025, 12:16:23 am »
Very interesting video - thanks for posting it.

We have one limitation with Dave's demonstration: we cannot know if the Gaussian curve would apply to the next tolerance band out (2%). That's because when resistors are binned, the top bin (tightest tolerance) does, indeed, have a normal distribution. It is only the lower bins that have a hole in the middle.

We don't know what the top level tolerance is on this particular family of resistors. It could be the 1% resistors Dave measured, or maybe it's tighter.

If binning is not used, and the tolerance is achieved purely by tightly controlled manufacturing processes, then we could reasonably expect all the tolerance bands to have a normal distribution.
If it ain't broke, keep fixing it until it is.
 

Offline bdunham7

  • Super Contributor
  • ***
  • Posts: 9452
  • Country: us
Re: Root Sum Squared Error Analysis
« Reply #46 on: May 08, 2025, 04:51:41 am »
That's because when resistors are binned, the top bin (tightest tolerance) does, indeed, have a normal distribution.

That depends on the binning process and probably other factors.  Other people have assumed that binning would result in a more uniform (rectangular) distribution because you are selecting from the center of the peak.  But I don't think that's correct either.  Suppose you take a group of resistors with a more or less uniform distribution over at least the range of +/- 1% and with fat tails beyond 1%--say an std dev of 5% or something.  Your binning selector machine itself has a random error of 0.5% so you tighten up the spec and reject any resistor that is more then 0.5% from nominal.  What does the distribution of the output look like?
A 3.5 digit 4.5 digit 5 digit 5.5 digit 6.5 digit 7.5 digit DMM is good enough for most people.
 

Offline Conrad Hoffman

  • Super Contributor
  • ***
  • Posts: 2307
  • Country: us
    • The Messy Basement
Re: Root Sum Squared Error Analysis
« Reply #47 on: May 08, 2025, 10:13:25 am »
As an aside, I was always impressed that Linear Technology would publish statistical distributions of, I think, noise, maybe other things, in their data sheets.
 

Offline eugene

  • Frequent Contributor
  • **
  • Posts: 602
  • Country: us
Re: Root Sum Squared Error Analysis
« Reply #48 on: May 08, 2025, 02:34:19 pm »
When resistors are binned you get a big hole in the middle of the distribution because those resistors have already been assigned to the next higher tolerance bin.

That's a faulty assumption.

It's not an assumption. It's how binning works

Unless you have specific knowledge about the process, it is an assumption. Some might even call it a wild guess.

Knowledge is often not absolute; we think some things are probably true, we're pretty sure of other things and we're almost completely sure of other things. Recognizing and being cognizant of this is essential to any scientist or engineer. We need to keep straight in our own mind what we know to be true and what we guess to be true.

Some people willfully choose to believe whatever feels right. Let's try to avoid that. It leads to misunderstanding at best and conspiracy theories at worst.
90% of quoted statistics are fictional
 

Offline MisterHeadache

  • Regular Contributor
  • *
  • Posts: 158
  • Country: us
    • Level UP EE Lab
Re: Root Sum Squared Error Analysis
« Reply #49 on: May 08, 2025, 04:08:56 pm »
I ran two more Monte Carlo 10k run simulations.  With the original, here are the parameters:

Simulation #1: Each resistor has a tolerance of +/-1% and each has a centered normal distribution with a standard deviation of 1/3 of their tolerance.

Simulation #2: Each resistor has a tolerance of +/-1% and each has a uniform distribution of their values (meaning any value between the lower and upper tolerance limits has an equal probability of occurring)

Simulation #3. Same as simulation #2 except they have been ‘binned’ such that all values between +/-0.5% have been removed from the population.  So you are left with a uniform distribution with a ‘hole’ in the middle.


Results are as follows:
Simulation #1: As I stated before, you get a gaussian distribution with a standard deviation of approximately between 0.0272 to 0.0280.  See attached image ‘gaussian’ as an example of one 10k trial run.

Simulation #2: You get a ‘mostly uniform’ distribution with a standard deviation of approximately 0.048.  However, I don’t think the standard deviation is a good measure here due to the ‘mostly uniform’ distribution resulting from the combo of just two resistors.  As you increase the number of resistors in the stack, the resulting distribution shifts to be Gaussian (a direct effect of the Central Limit Theorem…)  Anyway, see attached image ‘uniform’ as an example of one 10k trial run.

Simulation #3: This one took a bit of adjustments to my software, but I was able to set it such that for each random trial, there was an equal/uniform chance that a value of 99-99.5 or 100.5-101 would be chosen.  For the 10 ohm resistor, that is 9.9-9.95 or 10.05-10.1.  The standard deviation is approximately 0.063, but you end up with a distinct bimodal distribution.  No surprise, there has to be a hole in the middle, but I was surprised (based on simulation #2) that even with just two resistors we see a gaussian shape in them rather than two mostly uniform shapes.

Anyway, I have definitely :horse: at this point.  And let me paraphrase George Box: “ALL MODELS ARE WRONG, BUT SOME ARE USEFUL.”  So hopefully there is something useful here to the OP question.
Daryn from Level Up EE Lab
AKA 'MisterHeadache'
 


Share me

Digg  Facebook  SlashDot  Delicious  Technorati  Twitter  Google  Yahoo
Smf

 

- - -
-->