0
$\begingroup$

I have an independent variable $\delta$ and dependent set of statistical classified data $\widehat{\Lambda }(\delta)$. The model I want to fit to this data is: $$ \Lambda (\delta )=Ae^{-k\delta}*f(k) $$ There are many cases but for this example we assume: $$ f(k)=\frac{2}{k^2\sigma ^2}\big(e^{\frac{k^2\sigma ^2}{2}\Delta T}-1\big) $$ Note:$\Delta T$ and $\sigma$ are known variables.

In the white paper I've got this equation there was this equation given for the sum of the squared errors, that needs to be minimized: $$ r(A,k)=\sum_{\delta=1}^{\delta_{max}}\bigg(\log\big(\widehat{\Lambda}(\delta)\big)+k\delta-\log\big(A\big)-\log\big(f(k)\big)\bigg)^2 $$ Now since $f(k)$ is not dependent on $\delta$ I would substitute: $$ u=A*f(k) $$ Which gives you this equation as the model: $$ \Lambda(\delta)=ue^{-k\delta} $$ The function of the squared errors that needs to be minimized for this quasilinear model is : $$ r(A,k)=\sum_{\delta=1}^{\delta_{max}}\bigg(\log\big(\widehat{\Lambda}(\delta)\big)+k\delta-\log\big(u\big)\bigg)^2 $$ Now we have a linear equation we need to do our regression for, which is way easier than any non-linear approach. And with $k$ and $\log(u)$ in hand we can easily solve for $A$

I did some back tests with scipy's optimization functions and I got the same results when I optimized the parameters $A$ and $k$ with the initial model and the Levenberg-Marquardt optimization function, and when I tried fit $\log(u)$ and $k$ with the linear least squares fitting approach by hand and then solved for $A$.

But I want to ask you, the experts, if this is really mathematically justified.

Many thanks in advance.

$\endgroup$
0
$\begingroup$

This same logarithmic transformation is discussed in Least squares fit for exponential function.

To summarize, the logarithmic transform distorts the data. The solution does not find the parameters $A$ and $k$ which minimizes the original equation.

However, one can use the method of least squares in the core of the problem. The trick is to fix $k$ and then the problem is linear. Solve for $A$, the compute $r^{2}$. Scan along $k$ to find the minimum $r^{2}$.

The good news is that the least squares problem takes its simplest form: the average. Now, to define the problem. Consider a sequence of $m$ measurements of the form $\left\{ \delta_{j}, \Lambda_{j} \right\}_{j=1}^{m}$. For fixed $k$ we have the linear system $$ \left[ \begin{array}{c} e^{-k \delta_{1}} f(k) \\ \vdots \\ e^{-k \delta_{m}} f(k) \end{array} \right] % \left[ \begin{array}{c} A \end{array} \right] % = % \left[ \begin{array}{c} \Lambda (\delta_{1}) \\ \vdots \\ \Lambda (\delta_{m}) \end{array} \right]. $$ With the restriction $f(k) \ne 0$, a more recognizable form is $$ \left[ \begin{array}{c} 1 \\ \vdots \\ 1 \end{array} \right] % \left[ \begin{array}{c} A \end{array} \right] % = % \left[ \begin{array}{c} \frac{\Lambda (\delta_{1})} {f(k)} e^{k\delta_{1}} \\ \vdots \\ \frac{\Lambda (\delta_{m})} {f(k)} e^{k\delta_{m}} \end{array} \right]. $$ The least squares solution is the average: $$ A(k) = m^{-1} \sum_{j=1}^{m} \frac{\Lambda (\delta_{j})} {f(k)} e^{k\delta_{j}} $$ To connect this with least squares, note this solution is given by the Moore-Penrose pseudoinverse: $$ A(k) = % \left[ \begin{array}{c} 1 \\ \vdots \\ 1 \end{array} \right]^{\dagger} % \left[ \begin{array}{c} \frac{\Lambda (\delta_{1})} {f(k)} e^{k\delta_{1}} \\ \vdots \\ \frac{\Lambda (\delta_{m})} {f(k)} e^{k\delta_{m}} \end{array} \right] % = m^{-1} \left[\begin{array}{ccc} 1 & \cdots & 1 \end{array}\right] \left[ \begin{array}{c} \frac{\Lambda (\delta_{1})} {f(k)} e^{k\delta_{1}} \\ \vdots \\ \frac{\Lambda (\delta_{m})} {f(k)} e^{k\delta_{m}} \end{array} \right]. $$ Now the target of minimization, the sum of the squares of the residuals, is a function of $k$ $$ r^{2}(k) = \Bigg\lVert \left[ \begin{array}{c} e^{-k \delta_{1}} f(k) \\ \vdots \\ e^{-k \delta_{m}} f(k) \end{array} \right] % \left[ \begin{array}{c} A(k) \end{array} \right] - \left[ \begin{array}{c} \Lambda (\delta_{1}) \\ \vdots \\ \Lambda (\delta_{m}) \end{array} \right] \Bigg\rVert_{2}^{2} % = % \sum_{j=1}^{m} \left( e^{-k \delta_{j}} f(k) A(k) - \Lambda(\delta_{j}) \right)^{2}. $$

The least squares solution for your problem is defined as $$ k_{LS} = \left\{ k\in\mathbb{R} \colon r^{2}(k) \text{ is minimized}\right\}. $$

$\endgroup$

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.