When a datasheet lists minimum/typical/maximum

[cosmicray]

Should the distribution (across a random sampling) be a bell curve, or something closer to a flat line ?

The datasheet in question: https://www.ti.com/lit/ds/symlink/lm317l.pdf

The variable is Reference voltage. Pretty much all the manufacturer's datasheets, for that part, show the minimum as 1.2, typical 1.25, and maximum as 1.3. Do the actual parts vary across that range, or do they congregate around 1.25 with some parts being above or below ?

The parts I'm working with are not TI, they are generic bag of 100 TO-92 for $8. I seem to be seeing a scattering of Vref values. They seem to work OK as 2-terminal current limiters, but most require manual tweaking as 3-terminal voltage regulators. When I read the datasheet, and the various sample circuit diagrams, I noticed the variable resistor R2, and wondered why that was drawn like that. Now I'm beginning to understand why.

[PGPG]

The parts I'm working with are not TI, they are generic bag of 100 TO-92 for $8.

The world is full of fakes that do not meet the parameters of the original at all. I have read that even military factories are not able to completely protect their supply chains from fakes that may even have the markings of the originals.
So it is hard to say what you should expect.

[Benta]

Ideally it would be a bell curve, but that's only over an extremely large sample taken over time.
in practice it won't.
A wafer batch will be quite uniform, but whether it's 1.21, 1.25 or 1,29 V is impossible to say. But they 'll be very much alike.
I once saw an extreme example where a very big customer demanded selected parts at the very end of the tolerance band (millions of parts). Suddenly a large part of the tolerance spectrum was missing on the market.

[tggzzz]

Should the distribution (across a random sampling) be a bell curve, or something closer to a flat line ?

In general it could be something else. Classic example is op amps with a bimodal input offset voltage - because the manufacturer selected the better components and sold them as a more expensive variant.

Of course you also have to consider the population being measured.
It could be that all devices from one batch are similar - but different to those from another batch.
Or that all devices will exhibit similar behaviour when an external variable changes, e.g. PSU voltage or temperature.

[TimFox]

In the uncommon case where the production value is totally uncontrolled (over some range) and the resulting components are sorted into bins for different nominal values and tolerance, the distribution over the units sold would be a "uniform distribution" between test limits.
However, many production systems for, say, discrete components are controlled to concentrate the results around the nominal value.
A factory engineer for film capacitors once told me their target for accuracy when winding the films before test, and it was much tighter than their specified tolerance.

[Phoenix]

[quote author=TimFox link=topic=445927.msg5723441#msg5723441 date=1732305044.
A factory engineer for film capacitors once told me their target for accuracy when winding the films before test, and it was much tighter than their specified tolerance.
[/quote]

It's my experience with film caps these days is that a +10%, - 15% tolerance will come in quite repeatedly at around -10%. They know how to control it well but why add the extra material to get it any better when it meets spec.

[CatalinaWOW]

As demonstrated by prior answers the reality is that it can be almost anything.  Most will be some variation on, or sampling of a normal distribution (bell curve).  About the only thing you can be sure of is that the typical value is almost meaningless.  It may have been the design goal, or the production mean at the time of the data sheets publication, or the mean of the sample parts the application engineer had on hand or something else entirely.

Not uncommonly on older designs the production tolerances have tightened up enough that the bell curve is very narrow, far smaller than the data sheets limits.  But the mean could be anywhere within the upper and lower tolerances.   Unless the vendor thinks a much higher price can be charged for these tighter tolerance parts no change will be made on the data sheet so that future process shifts won't force an update

[mtwieg]

There's no basis for assuming any kind of distribution unless the manufacturer provides a histogram (which might not be representative of distributions for single lots, but rather for many different lots combined).

[PCB.Wiz]

The variable is Reference voltage. Pretty much all the manufacturer's datasheets, for that part, show the minimum as 1.2, typical 1.25, and maximum as 1.3. Do the actual parts vary across that range, or do they congregate around 1.25 with some parts being above or below ?

Of course 'it depends'.
The better parts will give you a histogram, that is a guide from a batch of samples.

Failing that, the other detail you can check, is does the vendor offer a 'better part' ?

It is common with voltage references to have 2% 1% & 0.5% specs, (less common in voltage regulators)
In those cases, someone buying a 2% part can expect the curve to be worse than flat, they may have a hole in the middle with very few 0.5% parts.

Alternatively, if the vendor is very good with their quality, they may make everything 0.5% and they simply label 2% parts because that is on purchasing officers spread sheets.

In the real world : if it matters test it, or buy a part that has the required better specs.

What variances were you seeing ?  If they were excessive, are you sure the part was stable ?

[WillTurner]

Some years ago I worked through the mathematics to try to understand engineering tolerances from a statistical viewpoint. I think the "uniform distribution" is a "rectangular" distribution. There are simple transforms between rectangular and gaussian. I'm sure there is someone much more knowledgeable out there who can help with that approach.
  However, given that the device in question is a bag lot from a questionable (and cheap-ass) source, it would be optimistic to expect any kind of conformance to the data sheet. Dave did a video on measuring the statistical distribution of a bunch  of resistors which would be found through a search. If it was worth the time, and the components were that worthy, then measuring the statistical distribution of a sample would give you some traction.
  But IMHO it's not worth the time. Buy components from reputable sources, and expect conformance to the (wide) tolerances in the data sheet. If something is out of specification, either you broke it, or something is seriously wrong. If you really need the close tolerance part, then pay for the better component.

[cosmicray]

But IMHO it's not worth the time. Buy components from reputable sources, and expect conformance to the (wide) tolerances in the data sheet. If something is out of specification, either you broke it, or something is seriously wrong. If you really need the close tolerance part, then pay for the better component.

So I checked Digikey, which should classify as a reputable source. I see many variants of LM317L (numerous from TI). Carefully checking the most recent TI datasheet, what distinguishes one from another seems to be (a) the package, (b) operating temperature range, and (c) how they are physically packaged (tape, ammo box, loose, etc). I'm not seeing any obvious sign of the parts being qualified by reference voltage. It may be possible to obtain that qualification, but it likely requires talking directly to TI (or whomever) and having to meet a minimum order for this level of binning.

For many parts, and certain parameters, the ability to buy better parts does exist. For the (relatively common) LM317L what I'm seeing leads me to conclude that the datasheet is what you get. The parts that I purchased could have been first quality from a Chinese producer, or they could have been floor sweepings. That I'm able to tweak my circuit, using a 3mm pot, to obtain the desired result (and the TI datasheet clearly shows a variable resistor in the circuit examples) makes me suspect that this part just has some variability (notwithstanding special orders for binned parts).

Various people ( to all of them ) are suggesting that there may be some form of a product objective. Anything that falls outside of that objective obviously should never see the light of day, and be sent to that great semiconductor final resting place. I'm suspicious that somewhere in the process is a QA department that monitors the reject percentages for signs of something wong. It may be that the current processes are so good, that rejects almost never happen. But the Vref range does exist, and it has to be adjusted for when it goes one way or the other.

Peace

[WillTurner]

I see many variants of LM317L (numerous from TI). ...  I'm not seeing any obvious sign of the parts being qualified by reference voltage.

So if you have the genuine part, and it operates with its reference voltage anywhere within the voltage tolerance band over it's temperature range (and other specified parameters), then it is in spec. Expecting the part to operate as a precision voltage reference is like trying to turn a pigs ear into a silk purse.   

That I'm able to tweak my circuit, using a 3mm pot, to obtain the desired result (and the TI datasheet clearly shows a variable resistor in the circuit examples) makes me suspect that this part just has some variability (notwithstanding special orders for binned parts).

Yes, the voltage output can be tweaked to bring it to the typical specification. There are costs in doing that.

Various people ( to all of them ) are suggesting that there may be some form of a product objective. Anything that falls outside of that objective obviously should never see the light of day, and be sent to that great semiconductor final resting place. I'm suspicious that somewhere in the process is a QA department that monitors the reject percentages for signs of something [wrong].

I certainly hope there is a QA process that is monitoring production. QA is looking for statistical shifts in production that indicate not only product failure, but shifts in the manufacturing process that would indicate a longitudinal problem. (At least that's how I think it works!) 

But the Vref range does exist, and it has to be adjusted for when it goes one way or the other.

Manual adjustment is one solution, but like I said it has costs.
If you want precision, you could go and hang out with the volt-nuts in the metrology section for a while. They will quickly disavow you of the notion of using variable resistor trimmers to get the precison output you seek!

Edit: minor grammar.

[donlisms]

It would certainly seem that a trimmer is a bad idea on the surface; they’re kinda messy in several ways.

The usual technique for trimming resistance (or a voltage drop, or current) to high precision is to put the trimmer in a part of the circuit where it has less influence, and a smaller effective range - generally in parallel with a much smaller resistor.

For example, suppose I want to precisely trim about 100 ohms (more or less).  If I put a 10k ohm pot in parallel across a 100 resistor (and assume that I should only use .7 of its range because things work out better that way), I have a range 100||10k to 100||3k, which is 99.0099 ohms to 96.7149 ohms, or 2.2357 ohms total - a reasonably fine adjustment range compared to the nominal 100 ohms.

Pots don’t realistically have infinite resolution. My expectation for trimmers and pots is to resolve about 100 “steps” per revolution, or 1000 steps for a 10-turn.  So my 100 ohm trim resolves to 2.2357 milliohms.

But the fun part is that the characteristics of the pot are reduced by a ballpark factor of roughly 100 in this case, so if its temperature coefficient is 250ppm, it effectively becomes 2.5ppm.

A lot of high-end metrology gear uses such techniques, and have done so for well over a hundred years.  (How do you balance a Wheatstone bridge to 0.00001 ohms, with switched decades?)

Use a reasonably good pot, to be sure, but perhaps you don’t have to worry quite so much about its harmful effects as it seems at first. 

[Smokey]

Didn't Dave make a video where he measured a bunch of resistors from the same bandoleer?  I forgot what the distribution looked like.

[cosmicray]

It would certainly seem that a trimmer is a bad idea on the surface; they’re kinda messy in several ways.

The usual technique for trimming resistance (or a voltage drop, or current) to high precision is to put the trimmer in a part of the circuit where it has less influence, and a smaller effective range - generally in parallel with a much smaller resistor.

Stet. That is exactly how the circuit is designed, two fixed 1% resistors in parallel with the trim pot. Actual pot p/n is EVN5ESX50B15 (of which I obtained a partial reel a few years back at a hamfest). Data sheet https://industrial.panasonic.com/cdbs/www-data/pdf/AOI0000/AOI0000C63.pdf

The intent of my question was not to search for precision parts. My intent was to ask about the distribution of Vref across common everyday parts. That the generated Vref varies across parts in a selection appears to be a given. Without asking for binning, the simple solution seems to be trimming the resistance to the negative rail (as shown in the data sheet). That there will be part variance seems to be obvious. So far I've seen a variety of answers as to why and, notwithstanding doing an in depth survey of the parts I have, it has to be designed for and with no expected distribution. Perhaps the variance is part of why/how this regulator came into existence in the first place.

What may have led me to a different initial assumption was the chart titled Reference Voltage Temperature Stability (using the National datasheet for the moment) where it is suggested that the Vref is going to be 1.25v, and this is how it varies based on temperature. Obviously that is not what to expect from every part, and is more likely a percentage from the Vref of a particular sampled part. Which says that should the Vref be close to one of the limits, temperature might actually push it slightly out of limit.

[cosmicray]

Didn't Dave make a video where he measured a bunch of resistors from the same bandoleer?  I forgot what the distribution looked like.

Possibly some combination of

Gaussian resistors redux. from 2011.

https://www.eevblog.com/forum/blog/gaussian-resistors-redux/

Gaussian Resistors also from 2011

https://www.eevblog.com/forum/blog/gaussian-resistors/