I have a sequence of $N$ strictly positive real values $y_n$. They form some kind of peak; for simplicity, let's assume $f(x, \mu) = A \exp^{-(x-\mu)^2}$ is the shape, with $A$ and $\mu$ real (in the end, I want it to work for a peak of fixed but arbitrary shape though). There is some additive gaussian-distributed noise in the $y_n$, i.e. they don't fit $f(x, \mu)$ exactly. My objective is to determine $A$ and $\mu$ according to the optimium given by the least-squares method, i.e. the values minimizing $\chi^2 = \sum_{n=1}^N \left(y_n - f(x_n, \mu)\right)^2$.

Of course, the intuitive approach is to use e.g. the Levenberg-Marquardt algorithm and iterartively determine those parameters.

However, for some classes of problems (esp. those linear in their parameters) explicit solutions do exist. I am wondering whether it is possible to find an explicit solution for this kind of problem as well.

What I tried so far is this: you can calculate the cumulative sum of the $y_n$. The resulting functional relation $g(x_n) = \sum_{j=1}^n y_n$ is strictly monotonic, and thus can be inverted. Looking at the inverted data $g^{-1}$, a change in $\mu$ now simply causes a constant offset. For this problem, an explicit solution of the least-squares problem can easily be derived. This works, but it has ugly properties in practice (e.g. additive noise causes a systematic shift in the resulting parameter distribution, etc).

Any ideas how one could do better?


Consider simplify the problem by removing the linear term. The posts Least squares method of a nonlinear function and Least Square Approximation for Exponential Functions show examples. In your sample case this would be $$ \hat{A}(\mu) = \frac{\sum y_{k} \exp\left(-(x_{k}-\mu)^{2}\right)} {\sum \exp\left(-2(x_{k}-\mu)^{2}\right)}. $$

Now the problem is a one-dimensional minimization with favorable monotonicity.


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.