Lets say I have a polynomial $ax^4 +bx^3 +cx^2 +dx +e$ and the uncertainties on each coefficient. Now I need to calculate the tangent at some points as well as some areas under this curve. How would I find the uncertainties on those results?

My initial idea was to simply differentiate or integrate it and then use the partial derivatives method on the obtained expression(ignoring the error on "x" itself). ex. $slope = 4ax^3 + 3bx^2 +2cx +d$

$$\sigma_{slope}^2 = (4x^3)^2*\sigma_a^2 + (3x^2)^2*\sigma_b^2 + (2x)^2*\sigma_c^2 + \sigma_d^2$$

But higher x will make this error much higher which doesn't seem logical to me. I couldn't find the answer online and my book doesn't cover it.

Whats the right way to propagate uncertainties in such cases?


1 Answer 1


This is exactly the right approach if the errors in the coefficients are uncorrelated.

You would indeed expect the error of the slope, as an absolute value, to be larger when x is larger. An easy way to start to think about this is to consider a simple parabola: $ax^2$ with some error in a. In this case we find that

$\sigma_{slope} = 2x\sigma_a$

We know that at $x=0$ the slope is exactly zero, so the error should be zero. At some finite value of x we expect a finite slope and a finite error. In this simple case, the fractional error in the slope is constant, which to me makes intuitive sense:

$\frac{\sigma_{slope}}{slope} = \frac{\sigma_a}{a}$

Now if the errors in the coefficients are correlated, a more complete approach is necessary. As an example, if the polynomial coefficients were determined from fitting a set of measured data, the errors in the coefficients would be expected to be highly correlated.

In this case the formula for the propagation of error for a function $y = f(a,b)$ is:

$\sigma_y$ = $\left|\frac{\partial f}{\partial a}\right|^2\sigma_a^2 + \left|\frac{\partial f}{\partial b}\right|^2\sigma_b^2 + 2\frac{\partial f}{\partial a}\frac{\partial f}{\partial b}\small{\mathrm{COV}}_{ab}$

where $\small{\mathrm{COV}}_{ab}$ is the covariance between $a$ and $b$. This can be expanded to many parameters by including every cross term, which becomes quite long to write out. In your specific example you will have 10 cross terms to consider.

  • $\begingroup$ I still find it a little bit hard to understand. If the errors on my original data are the same across some interval of x, why does the fitted polynomial always returns higher errors on higher x? The fit itself should not be "worse" for higher x. $\endgroup$ Nov 13, 2013 at 5:30
  • $\begingroup$ @user1830663 why not? If my error is $1$, then $(3+1)^2-(3-1)^2=12$, but $(10+1)^2-(10-1)^2=40$. A given amount of error in the initial measurement amounts to different amounts of uncertainties in the final measurement because consecutive squares aren't the same distance from each other. $\endgroup$ Nov 13, 2013 at 13:05
  • 1
    $\begingroup$ I think I see the source of your confusion, you are starting from a set of measurements ($y_i$), with some constant error ($\sigma_{y}$), and fitting a polynomial to the data. As part of the fitting procedure you determine the errors in the fitting coefficients. You are then trying to determine the error in $y_{fit}$ or $dy_{fit}/dx$ from these coefficients. What is happening here is that the errors in the coefficients are not independent. To correctly determine the error in $y_{fit}$ you also need to include the covariance terms. $\endgroup$
    – amicitas
    Nov 14, 2013 at 5:15

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.