So you have two datasets, thus sets of (x,y) pairs. Almost always it will be appropriate to log the model. Data that follows models like this often has proportional error structures anyway, thus noise that follows a lognormal distribution, or something close. This is especially important if your data spans more than a power of 10 or so in y. Otherwise, some data points will have too large of an impact on the fit.
So definitely log your data. If the models are:
y1 = a*x1^b
y2 = a*x2^(-b)
Then the logged models will be
log(y1) = log(a) + b*log(x1)
log(y2) = log(a) - b*log(x2)
Note that you wrote the log of the model incorrectly.
I'll assume that x1, y1, x2, y2 are column vectors containing the two data sets.
n1 = length(x1);
n2 = length(x2);
coefs = [ones(n1+n2,1),[log(x1);-log(x2)]]\[log(y1);log(y2];
a = exp(coefs(1));
b = coefs(2);
