Could it be that the formula for computing the chi² distance between histograms as pointed out here is wrong?
If one divides the squared difference of (h1 and h2) by h1 the distance is not symmetrical anymore which is not logical.
Here is a paper that defines the chi square distance between histograms as:
1/2 * sum over i of (h1_i - h2_i)^2 / (h1_i + h2_i)
which is symmetric and makes more sense as a distance measure