How to fit logarithmic curve to data, in the least squares sense?
How to fit logarithmic curve to data, in the least squares sense?
I have simple data of the type $(x,y)$, that is 2D. I need to fit curve of the type: $y = c_1 + c_2\ln(x)$. So I have the $x$'s and the $y$'s but I need to learn the $c_1$ and $c_2$ coefficients.
Thanks
$\endgroup$4 Answers
$\begingroup$One approach is to replace all of your $x_i$ with $X_i=\ln(x_i)$ and then do usual linear regression with the $(X_i,y_i)$.
$\endgroup$ 2 $\begingroup$Your equation is: $a\ln(x) + b = y$. So set up matrices like this with all your data:
$$ \begin{bmatrix} \ln(x_0) & 1 \\ \ln(x_1) & 1 \\ &... \\ \ln(x_n) & 1 \\ \end{bmatrix} \begin{bmatrix} a \\ b \\ \end{bmatrix} = \begin{bmatrix} y_0 \\ y_1 \\ ... \\ y_n \end{bmatrix} $$ In other words:
$$Ax = B$$
Where
$$ A = \begin{bmatrix} \ln(x_0) & 1 \\ \ln(x_1) & 1 \\ &... \\ \ln(x_n) & 1 \\ \end{bmatrix} $$ $$ x = \begin{bmatrix} a \\ b \\ \end{bmatrix} $$ and $$ B = \begin{bmatrix} y_0 \\ y_1 \\ ... \\ y_n \end{bmatrix} $$ Now solve for $x$ which are your coefficients. But since the system is over-determined, you need to use the left pseudo inverse: $A^+ = (A^T A)^{-1} A^T$. So the answer is: $$ \begin{bmatrix} a \\ b \\ \end{bmatrix} = (A^T A)^{-1} A^T B $$
Here is some simple Python code with an example:
import matplotlib.pyplot as plt
import numpy as np
TARGET_A = 2
TARGET_B = 3
N_POINTS = 20
MAX_X = 50
NOISE_A = 0.1
NOISE_B = 0.2
# create random data
xs = [np.random.uniform(MAX_X) for i in range(N_POINTS)]
ys = []
for i in range(N_POINTS): ys.append((TARGET_A + np.random.normal(scale=NOISE_A)) * np.log(xs[i]) + \ TARGET_B + np.random.normal(scale=NOISE_B))
# plot raw data
plt.figure()
plt.scatter(xs, ys, color='b')
# do fit
tmp_A = []
tmp_B = []
for i in range(len(xs)): tmp_A.append([np.log(xs[i]), 1]) tmp_B.append(ys[i])
B = np.matrix(tmp_B).T
A = np.matrix(tmp_A)
fit = (A.T * A).I * A.T * B
errors = B - A * fit
residual = np.linalg.norm(errors)
print "solution:"
print "%f log(x) + %f = y" % (fit[0], fit[1])
print "errors:"
print errors
print "residual:"
print residual
# plot fit
fit_x = range(1, MAX_X)
fit_y = [float(fit[0]) * np.log(x) + float(fit[1]) for x in fit_x]
plt.plot(fit_x, fit_y, color='k')
plt.xlabel('x')
plt.ylabel('y')
plt.show()If such a fit is appropriate then it's done the same way a least-squares fit is done when you have $y=c_1+c_2 w$, but in place of the $w$s you put the values of $\ln x$.
$\endgroup$ $\begingroup$In fact, as long as your functional form is linear in the parameters, you can do a linear least squares fit. You could replace the $\ln x$ with any function, as long as all you care about is the multiplier in front. Section 15.4 of Numerical Recipes, like any other numerical analysis text, will tell you how. Obsolete versions are free.
$\endgroup$
