0

I measure a capacitor \$C_1\$ with an instrument that has a tolerance of measurement \$\pm 5\%.\$ I do the same with another cap \$C_2\$. Then I put \$C_1\$ and \$C_2\$ in parallel (values add up) and measure both of them with the same instrument. Did I just dropped the tolerance to \$\pm 2.5\%\$? Can I compute (manually calculate) this way a new value of both \$C_1\$ and \$C_2\$ with a higher precision? (\$2.5\%\$)

I've just done some math and it seems unlikely. It's counterintuitive! (sad face). So here's the thing. When I ADD on paper the values of \$C_1\$ and \$C_2\$, their tolerances don't add up, contrary to what I thought. It still remains \$5\%\$. Then when I MEASURE the same capacitors in parallel, I get the MEASURED value of the sum, with the same tolerance \$5\%\$ as the one CALCULATED by hand on paper. So I win nothing. I get nothing.

However, my intuition tells me that all this is wrong. My intuition tells me that I CAN INCREASE THE PRECISION OF MEASUREMENT by multiple configurations of same two capacitors.

Hear me out. All the theory books tells us, that in order to increase the precision of a measured thing, you measure it repeatedly and average out the results. The more measurements, the higher the precision, the closer you get to the REAL VALUE OF THAT THING. That is in theory.

However, in practice, as usual, things are slightly different. I want to increase the precision of measurement of 2 capacitors with a digital \$LC\$ meter model LC100-S which has a tolerance of every measurement within \$5\%\$ of the REAL VALUE (\$+\$ or \$-5\%\$).

If I measure the same capacitor 10 times, I get the same value each and every time. So you can throw the theory books out the window on this one. Contrary to what the math tells us, my intuition tells me I CAN INCREASE THE PRECISION OF MEASUREMENT and decrease the tolerance from \$5\%\$ to a much lower value. If I measure the two capacitors \$C_1\$ and \$C_2\$ in parallel and in series, I can determine their values and average those values with the values measured individually. Thus I get a much closer value to the TRUE REAL value. WILL IT WORK? How do I do it?

Daniele Tampieri
  • 3,161
  • 4
  • 14
  • 29
Niculae George Razvan
  • 87
  • 6
  • The "accuracy" of measurement does not change. Only the "probability" of measuring the "good" value (but nothing says that it is the "good") increases when the number of measurements increases... according to the distribution of a Gaussian curve. – Antonio51 Mar 19 '22 at 08:25
  • I don't think the probability of MEASURED VALUES with a digital LC meter is a normal distribution (Gaussian curve) centered on the REAL VALUE. I think it's a different distribution where all values have same probability. I think it's something to do with how the ADC work and sampling and all that. – Niculae George Razvan Mar 19 '22 at 08:29
  • Have you measured the same "capacitor" at different "times", over a long period? The digital LC meter was "cold" or "hot"? Does it drift with "time" (yes)? The "precision" of a LCR meter is not only a "%" but also with a "number" of the "count". – Antonio51 Mar 19 '22 at 08:35
  • 1
    NB: you are right about the fact that it can not be the REAL VALUE. But it a "Gaussian" curve, because the ADC has a "Gaussian" curve... – Antonio51 Mar 19 '22 at 08:40
  • All measurement are done with LC meter cold. I never leave it on for too long. It never runs hot. I turn it on, measure and then turn it off. There is some drift with time, but it's slow and small over days and weeks. More likely is the capacitor that drifts than LC meter, but could be both. I'm interested in a short, quick measurement that doesn't take days to complete. It's impractical to get a high precision measurement if it takes days to complete in order to average out multiple measurements. – Niculae George Razvan Mar 19 '22 at 08:41
  • Perhaps you are right about the Gaussian curve of ADC inside the LC meter, I'm no expert, but surely this doesn't invalidate my work so far. See my answer. Until someone proves me wrong, I managed to reduce the tolerance and increase the precision of the measurement of my two caps. – Niculae George Razvan Mar 19 '22 at 08:44
  • 1
    https://everynanocounts.com/2013/07/09/noise-effect-on-the-distribution-of-adc-output-codes/ – Antonio51 Mar 19 '22 at 08:49
  • 1
    Note also that measurements must be made (for "good" comparison) at the same frequency. Otherwise, these are less "meaningful" and there are some "others" errors not taken into account ... "Measuring" is really a difficult "art" ... – Antonio51 Mar 19 '22 at 10:57
  • 2
    *All the theory books tells us, that in order to increase the precision of a measured thing, you measure it repeatedly and average out the results.* - WRONG!!! – Andy aka Mar 19 '22 at 11:00
  • Good point Antonio51. Forgot about that. This particular LC meter I'm using always seem to change the frequency at which the measurement is done. It has a button to display the used frequency. So I should make all measurements in a sequence quickly one after the other, making sure the frequency is the same(I have no control over it). – Niculae George Razvan Mar 19 '22 at 11:00
  • 1
    Oh come on Andy aka! As I said to someone else around here, Don't just contradict. Prove me wrong. Show me why I'm wrong. If you don't have time for that, don't bother commenting at all. You are just adding more confusion to the problem. I am here on this site to eliminate that, not to add more. Andy aka, you think you are helpful with your comment? Think again! – Niculae George Razvan Mar 19 '22 at 11:03
  • 1
    When one measures repeatedly some "thing", we obtain the "most" probable value of that "thing" (according to Gauss curve) ... but only if the "measurement system" is really "reliable" and well "calibrated"? – Antonio51 Mar 19 '22 at 11:10
  • True antonio51, there are some fundamental assumptions that (are/must) generally accepted before such a statement is made. If not, then such a statement does not hold. Note I used the term "all", not "always". I didn't generalized it. The golden rule: There are exceptions to every rule (including this one: self-referral) meaning there are a series of transcendental rules that have no exceptions whatsoever that are beyond the scope of the problem at hand. – Niculae George Razvan Mar 19 '22 at 11:15

4 Answers4

5

Did I just dropped the tolerance to ±2.5%?

Nope. Let's say the caps are 1 uF, just to have a number to deal with.

5% of 1 uF is .05 uF, and you know that that is the maximum error for each cap.

If you measure in parallel, your maximum error is 5% of 2 uF, or 0.1 uF. If each cap were high by .05 uF, the parallel combo will be high by 0.1 uF, and the meter won't help.

WhatRoughBeast
  • 59,978
  • 2
  • 37
  • 97
  • you probably meant "If each cap were high by .05uF, the parallel combo measurement could be as high as 2.1uF, so such a parallel measurement is meaningless". I agree. Did you read my answer to my question? I did found a workaround the issue. – Niculae George Razvan Mar 19 '22 at 14:01
4

Further edit: in order to answer the observations of Antonio51 and the one of the Asker, I've added some notes on the real structure of the reading precision of instruments (which is independent from the instrument being analog or digital) and on the LC100-S LC meter.

The answer
The calculations are formally impeccable: however the conclusions are flawed by a (quite common) misconception related to the meaning of the (so called) percentage error. Percentage error is only a means to express the absolute error in a more compact way: precisely, when I say that an instrument as an error of \$\pm 5\%\$ I mean it has an absolute error of \$\pm 5\%\$ of its full scale value(s). Thus, if your \$LC\$ meter is set to measure a full scale maximum value of say \$10\text{nF}\$ and you measure a capacity \$C_1=7\text{nF}\$, it is not correct to say that the value of \$C_2\$ lies in the interval \$(7\pm 0.35)\text{nF}\$. The \$\pm 5\%\$ error really means $$ C_1 \in (7\pm5\% \mathbf{fs})\text{nF}=(7\pm0.5)\text{nF}\label{1}\tag{1} $$ where \$\pm5\%=0.5\text{nF}\$ is the absolute error of the given instrument for that full scale value.

In sum, the increase of precision of the measured value in your reasoning is due to the fact that a constant \$\pm 5\%\$ error on the measured value(s) is assumed, which is not true.

Notes

  • The complete structure of reading precision, for an analog or digital mesuring instrument. The above formula comes from the following, exact one $$ v_\text{true} \in \big(v_\text{read}\pm e_\% \mathbf{fs}\pm 1\text{lsf}\big), \label{2}\tag{1'} $$ where

    • \$v_\text{true}\$ is the "true" value of the quantity to be measured.
    • \$v_\text{read}\$ is the value given by the instrument,
    • \$e_\% \mathbf{fs}\$ is the absolute error, or the precision of the instrument,
    • \$1\text{lsf}\equiv 1\$ unit on the least significant figure is the reading error i.e. the one done when you read the value on the display of the instrument. It can be the single graduation in the graded index of an analog instrument, or a "\$1\$" one on the least significant digit of your digital instrument.

    In any case, the reading error is never a constant percent of the read value: values near to the low side of the range are measured with a significantly worser precision than values near to the high side.

  • The applicability of formulas \eqref{1} and \eqref{2} to a digital instrument. Answering to Niculae's comment, let's see why these formulas are respectively nearly exact and exact by analyzing their application to the LC100-S, which thus is no exception to the rule defined by formula \eqref{2}. This nice instrument offers a reading with four significant digits and an autorange function. This means that, for the case of a capacitance measurement, you have for example the following ranges automatically switched $$ \begin{array}{|c|c|} \style{font-family:inherit}{\text{Reading}} & \style{font-family:inherit}{\text{Full scale }(\mathbf{fs})\text{ range}} \\\hline 5.00\mu\text{F} & 10.00\mu\text{F} \\\hline 7.00\text{pF}. & 10.00\text{pF} \end{array} $$ In the first case we have $$ C_\text{true} \in \big(5.00\pm 0.1\pm 0.01\big)\mu\text{F}= \big(5.00\pm 0.11\big)\mu\text{F}\simeq \big(5.00\pm 0.1)\mu\text{F}, $$ while in the second case we have $$ C_\text{true} \in \big(7.00\pm 0.1\pm 0.01\big)\text{pF}= \big(7.00\pm 0.11\big)\text{pF}\simeq \big(7.00\pm 0.1)\text{pF}, $$ Thus, equation \eqref{1} and \eqref{2} are respectively nearly exact and exact also for digital instruments, in particular for the LC100-S: therefore the objection contained in my answer remains valid.

Daniele Tampieri
  • 3,161
  • 4
  • 14
  • 29
  • 2
    +1. The "precision" of an LC meter is not only a "%" :-) ,but also with a "number" of the "count", because it is a "digital" instrument. – Antonio51 Mar 19 '22 at 09:39
  • I see. However, I don't know about other LC meters, but this LC100-S has only one scale value for capacitance and two for inductances. The scale for capacitance is from 0.01pF to 10uF. If I put these values in your equation, it's ludicrous. The Tolerance is actually from 1% to 5%, which I didn't knew exactly why some producers mention that, so I used the greatest 5% which gave me the greatest error. Now I see it's more complicated than that. As I approach the end of the scale with my measured cap, let's say 9uF, the error is almost 1%, but a cap of 100pF has more like 5% error of measurement. – Niculae George Razvan Mar 19 '22 at 09:52
  • The equation you gave is true for analog instruments. Precision of digital measurement instruments works slightly different. – Niculae George Razvan Mar 19 '22 at 10:00
  • @Antonio51 I'll add something more to my answer in order to precise the exact form of the measure error, for both dialog and digital instruments, following your observation and the following related note of the answer – Daniele Tampieri Mar 19 '22 at 12:40
  • @NiculaeGeorgeRazvan I'll adding something more on the complete structure of the error, as I wrote to Antonio51 above. – Daniele Tampieri Mar 19 '22 at 12:46
  • 1
    Thank you Daniele Tampieri for you answer. It took me some time to figure out everything you said. I understand now. From my view the problem is solved and the question is closed. – Niculae George Razvan Mar 19 '22 at 20:28
  • @NiculaeGeorgeRazvan thank you: I am glad to have been of some help. – Daniele Tampieri Mar 19 '22 at 20:30
1

Assuming you have perfect caps and the error is only on the measurement side.

You measure (C1 + C1err) on C1 and (C2 + C2err) on C2. When you add those, you add errors. So the worst it gets is 1.05*C1 + 1.05*C2 = 1.05*(C1+C2). Obviously, errors can be of the opposite sign, so you result will get more accurate since the errors partially cancel each other out. But it may just as well be in one direction.

The point is, the error is on the measurement side. For your measuring tool, there are no 2 caps of 5nF. There is one cap of 10nF. There are no 10 caps of 5nF. There is one cap of 50nF. Measuring those will give exactly identical results, since, again, it's your tool that sees 50nF wrong (for example, shows 48nF for the real value of 50nF). Your tool doesn't know how many caps there are, it only sees total capacitance and nothing else. It doesn't know what it's measuring.

Now, if you have 10 tools each with 5% error and you measure with them, then the distribution of the error will indeed be tighter, but the maxima and minima will still be 5%.

Things change if you talk about part tolerance - measuring multiple caps is more likely to be close to true value, especially if you use more and more caps, but we are still talking about probabilities. If you imagine error has normal distribution around true value, then more and more caps will make the middle more and more likely (at least as long as caps are of similar order of magnitude, bigger cap obvious has greater individual influence on error, I'm simplifying for the sake of making it more intuitive). But minima and maxima add and will still be +-5%, there is still a chance that all your measurements will be all 5% off to one direction and therefore so will the total error as well.

If you want to get true cap value, it is actually a good idea to measure many caps of the same value in parallel, the chance that you get some weird batch with all of them off in one direction will already be next to zero (assuming your tool is much more precise and the error comes predominantly from parts' tolerance). It doesn't take too much to get probabilities of that comparable to piano falling on you in the street. It's like tossing a coin 1,000,000 times. You can get 900,000 to 100,000 in theory, but you will probably get something like 50.1 to 49.9.

Ilya
  • 3,478
  • 7
  • 33
  • Completely agree. Some of your answer repeated what I said in different terms and words. I ask for forgiveness if I didn't used proper terminology. English is not my first language. Thank you for your support. If you can, please revise my work so far (my answer to my question) – Niculae George Razvan Mar 19 '22 at 08:54
  • 2
    You make there false assumption of true value being 5% away from measurement (if real value is 100pF, you measure 95pF, then real value is more than 5% off). And you talk in terms of specific example with specific values, which is wrong. Your logic pretty much collapses there. Any kind of specific example is like taking a sample from the distribution - you will most likely get close to true value result, but you absolutely can not generalize from it. That's why statistics is so intuitively difficult (among other reasons). You have to stop fixating on values and think in probability functions – Ilya Mar 19 '22 at 09:27
  • Also, in real life, your inductor for measurement probably has some serious tolerance like 10% (often 20%), and common ceramic capacitors also have some 10% tolerance. So the whole discussion is more of an ideal model with controlled parameters, in real life it gets even more complicated with even more unknowns with more probabilities of various parameters all interacting together. – Ilya Mar 19 '22 at 09:30
  • You said: if real value is 100pF, you measure 95pF, then real value is more than 5% off. I say: FALSE. Let Measured value be M and true real value V. M always falls within 5% of V, which means always (M>=0.95*V and M<1.05*V) OR (M>0.95*V and M<=1.05*V) but also V is "kind of" within 5% of M, although not all values +/-5% of M could be V (intervals don't coincide they are always shifted, only exception being when M=V). So (0.95M=V and 1.05M>V) – Niculae George Razvan Mar 19 '22 at 09:56
  • You say I make 3 false assumptions, but you only mention 2. true value(V) being 5% away from measurement(M) which is not false as you say(check prev comment), and 2nd assumption you say it's false is my generalisation. You forget that this is only one example.I didn't generalized anything, YET. More work needs to be done before that can even be approached.Perhaps the 3rd false assumption you mention regards the Tolerance of the capacitor. This can be calculated after the measurement, but has nothing to do with it and doesn't interact with it.I don't see how this omission is a false assumption. – Niculae George Razvan Mar 19 '22 at 10:11
  • As for generalization, I could simply code a program to take all kinds of values, give specific input parameters in order to stress-test my solution. Output some neat graphs and plots to satisfy some self-righteous statisticians who think in terms of probability functions around here. – Niculae George Razvan Mar 19 '22 at 10:26
  • I never said "three" false assumptions. I just said "false assumptions" (the word is "there", not "three"). "You said: if real value is 100pF, you measure 95pF, then real value is more than 5% off. I say: FALSE." this reply of your is plain wrong in its entirety. There is no "kind of" right. It's maths. It's either yes, or no. If B is within 5% of A, it does NOT mean that A is within 5% of B. Period. End of story. – Ilya Mar 19 '22 at 10:40
  • Also, you absolutely do have to take tolerance into account. Imagine you have 5% tolerance and 5% measurement error. 100pF becomes real 95pF, when if you have 5% off during measurement, it becomes 90.25pF. And yes, we can assume for the sake of ideal model that there is no tolerance. Just pointing out that things all interact with one another and edge cases can get really ugly. And since every measurement is a pick from probability distribution, you might get perfect measurement, but you can also get an outlier. It's important to remember where every error comes from. – Ilya Mar 19 '22 at 10:43
  • It's not the end of the story just because I don't know how to express myself correctly. I thought am here to learn, but it turns out nobody is helpful today. Did you even bother to check the inequalities given? If you don't have time to put your mind to it and try to understand me and my improper terminology and words, then don't bother answering at all. – Niculae George Razvan Mar 19 '22 at 10:44
  • @NiculaeGeorgeRazvan I'm genuinely trying to help and explain why your reasoning is wrong. I suggest you take a step back on your reasoning and start over on a completely abstract level. No values. Only abstract distributions. Here is a distribution of measurement for one point. Here is for another. Is there a different error for every measurement? Is there the same relative error per measuring tool for all measurements? It will make things more intuitively understandable. Also, don't forget that while you can have better *probability* of getting better measurement, tolerance is edge case. – Ilya Mar 19 '22 at 10:49
  • You said>100pF becomes real 95pF, when if you have 5% off during measurement, it becomes 90.25pF. 100pf doesn't matter. That's just a marking on the cap to show what the value should be, but it rarely is. The 100pF is just the tip of the gaussian curve of the normal distribution with edges 5%, 10% (based on cap markings). How much off-center a specific cap is, is irrelevant to it's measurement. That's exactly what I want to determine by the measurement. I can't use that in my initial measurement because of circular logic fallacy. – Niculae George Razvan Mar 19 '22 at 10:51
  • By the time I make the generalization form (with no thanks to anybody around here), you will forget about me or this post. Submerged in deep frustration I say thanks for your demoralizing no-help. – Niculae George Razvan Mar 19 '22 at 10:56
  • @NiculaeGeorgeRazvan then you need to set clear parameters, such as: does every measurement have different error or does your tool have fixed error that is reproduced in every measurement? And yes, you'll have to search for real value C, which is not the value nominal Cnom. In such a case, we can indeed drop Cnom altogether out of discussion. Then you do indeed have only measurement tolerance as the only source of error. Still, your 5% logic is mathematically wrong. If A = 0.95B, B = 1.0526A, if A=1.05B, B= 0.952A. Math is strict. There is no "kinda". Again, think in probabilities. – Ilya Mar 19 '22 at 10:59
  • From many measurements I've done to same capacitor over time, I have the impression the LC meter doesn't have a fixed error (which is more common for analog measurement instruments), and seem to indicate measured values that lie on a normal distribution around the unknown real value. Yes math is strict, and your equalities are correct, but I don't have equalities, I have multiple system of inequalities, which is a huge difference. Note the "and/or" operators. Where is an "and" both inequalities (in the sistem) must be true AT THE SAME TIME. MATH IS STRICT. – Niculae George Razvan Mar 19 '22 at 11:11
0

I got somewhere with this. Here's my progress so far. Let's start with this example problem:

Two capacitors \$C_1\$ and \$C_2\$ are given. We are asked to measure their values with maximum precision possible. We are given an LC Meter which has a tolerance of 5% for every measured component. We are given the TRUE REAL values of our caps as \$C_1\$=7nF and \$C_2\$=5nF, but we don't know these values, and we are asked to show on paper how through math, we have calculated(based on measurements) new values for \$C_1\$ and \$C_2\$ which are much closer to their REAL TRUE values.

Problem appears simple, we just measure the caps and we are done, right? Not so fast. This is surprisingly much more complex than that. We are asked to increase the precision of our measurements.

At first we can determine the 5% interval based on their real values. For \$C_1\$ the measurement MUST BE within 6.65-7.35 and for \$C_2\$ the measurement MUST FALL within 4.75-5.25

If we calculate the sum for the minimum values is 6.65+4.75=11.4 which is 5% lower than the TRUE SUM(12) and if we calculate the sum for the maximum values we get 7.35+5.25=12.6 which is 5% higher than the TRUE SUM(which is 12). We thus have proven that if we measure a parallel connection (with 5%) we are no closer or farther from what we can just calculate from their individual measured values. We haven't increased any precision nor decreased any tolerance. We aren't going anywhere with this, or are we?

We start by measuring \$C_1\$. The LC meter reads 6.83nF and for \$C_2\$ it reads 5.15nF. We then put them in parallel and measure 12.1nF Interesting...

We have to think this in reverse. Instead of thinking of our measured value to fall within 5% of the TRUE REAL VALUE (left and right 5%) we think of the REAL VALUE to be within 5% of the measured value (which is true if you think about it). This reversed-thinking is very helpful to proceed further with our calculations. We now can determine that OUR REAL VALUES for \$C_1\$ must be within 6.4885-7.1715 and for \$C_2\$ must be within 4.8925-5.4075.

We then calculate the sum of \$C_1\$ and \$C_2\$=11.98 and we average it with the one we measured 12.1 and we get 12.04 This new sum immediately reduces our intervals, cutting from the extreme ends of them, since not all values added respect this new sum of 12.04

Please feel free to correct me if I'm wrong, but the second interval for \$C_2\$ (4.8925-5.4075) "restrains" the first \$C_1\$ interval from (6.4885-7.1715) to (6.6325-7.1475) which does not seem a lot, but trust me IT IS!!! Put in another way, this new restrained, reduced or cropped interval means our NEW VALUE FOR \$C_1\$ (the middle of the new interval) is 6.89nF \$\pm 3.7\%\$ !!! Amazing! The value for \$C_2\$ remains (in this example) unchanged.

This is extraordinary jaw dropping(at least for me) improvement in measurement accuracy !!! And we can do this just by 3 measurements \$C_1\$, \$C_2\$ and then measure them in parallel. All this calculations can be done in the software for an eventual future build digital LC meter! Amazing!

I have yet to complete similar calculations for a series connection of the same two capacitors.

What am I doing wrong?

Niculae George Razvan
  • 87
  • 6
  • 1
    "We have to think this in reverse." "(which is true if you think about it)" no it's not. – TypeIA Mar 19 '22 at 08:03
  • Please elaborate. Don't just contradict. Prove me wrong. Think of a measured value to be at the very end of an interval. 5% is vice-versa for both values (real and measured). The measured value if 5% lower than the real value, and the real value if 5% higher than the measured value. Sure the intervals are not the same, but it doesn't matter. What matters is that the REAL value will lie within 5% of the measured value. – Niculae George Razvan Mar 19 '22 at 08:06
  • If real value is 5 and I measure 4.75 (we are at the lowest edge of 5% interval), then 5% lower than 4.75 is 4.5125 which is way to off and way to low to be true. But it doesn't matter. What matters is that 5% higher than measured (4.75) is 5 which is true real value. The real value will ALWAYS be within 5% of the measured value and vice-versa. – Niculae George Razvan Mar 19 '22 at 08:17
  • 1
    `M * 1.05 > V` and `M > V * 0.95` are not the same. 1 / 0.95 is 1.0526... – TypeIA Mar 19 '22 at 09:14
  • I think I see. The problem is with how I used my words. Forgive me if I'm lacking proper terminology. Let Measured value be M and true real value V. M always falls within 5% of V (which means always M>=0.95*V and M<1.05*V OR M>0.95*V and M<=1.05*V) but also V is "kind of" within 5% of M, although not all values +/-5% of M could be V (as I stated previously) (intervals don't coincide they are always shifted, only exception being when M=V). So 0.95M=V and 1.05M>V – Niculae George Razvan Mar 19 '22 at 09:36