In the sample size $n=2$ case when transforming $\{x_1, x_2\}$ to $\{\bar{x}, s\}$ (where $X_1, X_2 \overset{iid}{\sim} IG(\mu, \lambda)$, $\bar{X}=\frac{\sum_i^2 X_i}{n}$, and $S=\sum_i^2 (\frac{1}{X_i}-\frac{1}{\bar{X}}$), Schwarz and Samanta (1991) write:

I understand the motivation for restricting the transformation to this half-plane, but I'm at a loss for how to find this Jacobian myself. I understand how to find Jacobians for one-to-one transformations, but I'd appreciate a push in the right direction for cases like this.


For simplicity write $x=x_1$ and $y=x_2,$ so that

$$2\bar x = x+y,\quad s = \frac{1}{x} + \frac{1}{y} - \frac{4}{x+y}.\tag{*}$$

Their Jacobian $J(\bar x, s)$ can be computed by comparing the differential elements in the two coordinate systems

$$\mathrm{d} \bar x \wedge \mathrm{d} s = \frac{1}{2}(\mathrm{d} x + \mathrm{d} y) \wedge \left(-\frac{\mathrm{d} x}{x^2} - \frac{\mathrm{d} y}{y^2} + \frac{4(\mathrm{d} x + \mathrm{d} y)}{(x+y)^2}\right)=\frac{1}{2} \left(\frac{1}{x^2}-\frac{1}{y^2}\right)\mathrm{d} x\wedge \mathrm{d} y$$

(These differential forms and the wedge product are described at Construction of Dirichlet distribution with Gamma distribution.)

It remains only to divide both sides by $(1/x^2 - 1/y^2)/2$ and then re-express the left hand side entirely in terms of $\bar x$ and $s.$ Elementary algebra (see below for details) gives

$$\frac{1}{x^2}-\frac{1}{y^2} = \frac{s^{1/2}(2 + s \bar{x})^{3/2}}{\bar{x}^{3/2}},$$


$$\mathrm{d} x\wedge \mathrm{d} y = \frac{1}{\frac{1}{2} \left(\frac{1}{x^2}-\frac{1}{y^2}\right)}\mathrm{d} \bar{x} \wedge \mathrm{d} s =\frac{2\bar{x}^{3/2}}{\sqrt{s}(2 + s \bar{x})^{3/2}}\,\mathrm{d} \bar x \wedge \mathrm{d} s = J(\bar x, s) \,\mathrm{d} \bar x \wedge \mathrm{d} s.$$

Thus, when converting an integral in terms of $(x,y)$ to one in terms of $(\bar x, s)$, the differential element on the left side must be replaced by the differential element on the right; the absolute value of the factor $J(\bar x, s)$ that is introduced is the Jacobian.

Here is one way to do the algebra. Multiplying $s$ by $2\bar x$ to clear the fraction in $(*)$ gives

$$2\bar{x} s = \frac{(y-x)^2}{xy}.$$

Adding $4$ to that produces

$$2\bar{x} s + 4 = \frac{(x+y)^2}{xy} = \frac{(2\bar x)^2}{xy},$$

which upon dividing both sides by $(2\bar x)^2$ establishes

$$\frac{1}{xy} = \frac{2\bar{x} s + 4}{(2\bar x)^2}.$$

For $y\gt x \gt 0$ we may combine the three preceding results to obtain

$$\left(\frac{1}{x^2} - \frac{1}{y^2}\right)^2 = \frac{(y-x)^2(x+y)^2}{(xy)^4} = (2\bar{x}s)(2\bar{x}s + 4)\left(\frac{2\bar{x} s + 4}{(2\bar x)^2}\right)^2.$$

Collecting identical terms in the fraction and taking square roots gives

$$\frac{1}{x^2} - \frac{1}{y^2} = \left(\frac{s(2\bar{x} s + 4)^3}{(2\bar x)^3}\right)^{1/2}$$

which easily simplifies to the expression in the question.

| cite | improve this answer | |

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.