Dismiss Notice
Join Physics Forums Today!
The friendliest, high quality science and math community on the planet! Everyone who loves science is here!

Total error in a measurement.

  1. Sep 8, 2008 #1
    I have a question that has been bugging me lately. How is it that you determine the total error of a measurement?

    For example, if we are trying to measure the flow rate of water coming out of a hose. We let the water flowing through the hose fill a graduated cylinder and measure the time it takes to do it. So we would have:

    Flow rate = Volume / Time

    However lets say that we need to know the error of this measurement. Would we say that the total error is:

    Error = Ev*dQ/dv + Et*dQ/dt (d's are partial derivatives)

    Q = function for flow rate
    Ev = max error from volume measurement
    Et = max error of time measurement
    Error = total error

    So the formula would ultimately be:

    Error = Et*-V/t^2 + Ev*1/t

    Would this be correct? Haven't done this in a while and its just not making sense to me?
    Last edited: Sep 8, 2008
  2. jcsd
  3. Sep 8, 2008 #2
    I think you take the *partial* derivatives of the function wrt each variable, and multiply each pd by the error, then sum them up.

    So in your example where F = V/t

    you have err = ev d(V/t)/dV + et d(V/t)/dt (where these d's are partials)

    You have to be careful to keep track of the units, sometimes errors are given as percent of scale and sometimes in absolute units.
  4. Sep 8, 2008 #3
    Thanks for the reply. Thats what I meant to say, I should probably correct that. I always get confused about this formula for some reason it just doesn't make much "physical" sense to me.
  5. Sep 8, 2008 #4
    The physical meaning is this - the partial derivative of f(x,y) wrt x is how much f changes for a given change in x. Now consider ex, the error in x, as that "given change in x". Then the change in f for the change in x is ex times partial of f wrt x.
  6. Sep 8, 2008 #5

    Andy Resnick

    User Avatar
    Science Advisor
    Education Advisor

    John Taylor has written an excellent book "An Introduction to Error Analysis", and it's worth reading.

    Under most conditions- errors in one quantity (say cylinder radius) are independent of errors in another quantity (say time to fill)- then the different errors add in quadrature. Another importnat assumption is that repeated measurements form a Gaussian distribution. Then, given a function of several variables q=F(x,y,v,...t) then:

    [tex]\delta q = \sqrt{(\frac{\partial q}{\partial x} \delta x)^{2}+(\frac{\partial q}{\partial y} \delta y)^{2}+(\frac{\partial q}{\partial v} \delta v)^{2}+...(\frac{\partial q}{\partial t} \delta t)^{2}}[/tex]

    or variants thereof. The formula your wrote is also true, but it's an extremum value.
  7. Sep 8, 2008 #6
    sorry for any confusion - you are quite right about summing inquadrature (a.k.a, root-sum-square or RSS). This is probably where the OP has trouble with the "physical meaning." I remember when I was trying to get to understand the meaning of the RSS formula, I found Taylor's book and it helped me out. Now that was a few years ago and I can't remember how he explained it...And my copy of the book is not at hand...
  8. Sep 8, 2008 #7


    User Avatar
    Science Advisor
    Gold Member

    Error estimation is very tricky and there are often no single "correct" estimate. Nowadays Monte Carlo calculations are used in more advanced applications but even then there are problems; mainly related to how we interpret probability (i.e. what does the limits actually mean).
    From what I understand even GUM* is a bit messy in that regard, mixing Baysian and functionalistic approaches.

    *GUM=Guide for measurement uncertainty, a publication from ISO, IEC and a few other standardization organizations.
  9. Sep 8, 2008 #8
    I believe your equation for the error is correct, apart from absolute values!

    I think it's supposed to be:

    [tex]dQ = \left| \frac{\partial Q}{\partial V} \right| \, dV + \left| \frac{\partial Q}{\partial t} \right| \, dt = \frac{1}{t} \, dV + \frac{V}{t^2} \, dt[/tex]
    (Assuming t and V are both always positive as would be the obvious case with time and volume)

    Note the + V/t^2..! It doesn't make sense to have your error become less (or even negative!) if some partial derivative happens to be negative...
  10. Sep 9, 2008 #9
    Thanks for all the replies. I am now a little more knowledgeable about error analysis. I'll try to get to that book someday but right now I just don't have the time due to other books that I have to read.
Share this great discussion with others via Reddit, Google+, Twitter, or Facebook