Why count inverse images?
A measurement can lose information. Knowing does not tell us whether was positive or negative. The probability of a small interval around a positive value of therefore comes from two separated intervals of possible values. We must add both contributions. In two dimensions a fold can likewise give several preimages; the Jacobian adjusts area on each branch, but does not count the branches for us.
A density is defined only up to changes on sets of Lebesgue measure zero. Thus a formula that is correct away from branch boundaries can still determine the entire distribution, provided those boundaries carry no probability mass. This qualification is useful at the origin in polar coordinates, where an angle is not uniquely defined.
A finite-branch rule
Theorem 28.1.1
(Adding the inverse branches)
Suppose has density on . Up to a set of -probability zero, partition into finitely many disjoint open regions . On each, suppose is a continuously differentiable bijection onto an open region , with continuously differentiable inverse . Then a density of is, for almost every ,
A branch that does not contain contributes zero. The formula includes transformed supports; it does not claim that all branches exist everywhere.
Proof
(Partition first, then change variables)
Let be a measurable set. The discarded set has -probability zero, and the regions are disjoint, so is the sum over of . On the map is a bijection onto with inverse , hence . The one-to-one change-of-variables theorem, applied on each branch, gives
The sum is finite, so the right side equals with as displayed. Since this holds for every measurable , is a density of .
Proof X-ray
Why the branch contributions add
The one idea is to sort the input by the branch it lies in, rather than to sort each output by its preimages. On the input side the pieces are disjoint, so their probabilities add without double counting. The qualification “up to a set of probability zero” is used exactly here: it lets the proof discard branch boundaries, such as the point for , where is not locally invertible.
Injectivity of on each and the continuously differentiable inverse are used only to apply the one-to-one change of variables to . On the output side the images may overlap, which is why several terms can be present at the same , and why a branch whose image misses contributes nothing. The inverse determinant measures input volume per unit output volume; it is not an extra probability assigned to a branch.
Worked example 28.1.2
(Squaring a uniform variable)
Let be uniform on and put . For , both and lie in the original interval; for only does. Each inverse derivative has absolute value . Hence
The values at can be chosen arbitrarily. The two integrals are and , respectively. For instance , because the original interval corresponding to is , of length two. The density diverges near zero, but that does not give positive probability.
Counterexample mode
One branch is not always enough
Using only the positive square root would give total mass in the example, and would incorrectly give zero density for . Conversely, doubling one branch over the whole range also fails: the positive branch ceases to exist at . Branch counts and transformed support must be checked together.
The same rule handles an input density on the whole line, where the fold point lies in the interior of the support.
Worked example 28.1.3
(The square of a standard normal variable)
Let , with density , and put . Take and ; the omitted point has probability zero. Both branches map onto , with inverses and inverse derivatives of absolute value . Since is even,
and for . As a check, the substitution gives , so integrates to one. The edge case is the fold point: vanishes at , which is where the inverse derivative blows up. The density is unbounded near but still integrable, because the single excluded point carries no probability.
Folding a joint density
Worked example 28.1.4
(Absolute value with a changing support)
Let be uniform on the triangle , ; its area is one, so the joint density is one there. Set , . For and , both inverse points and lie in the triangle. Each inverse determinant has absolute value one. Thus
and it is zero outside this triangle, apart from arbitrary boundary values. Integrating gives for . Given at such a value, the conditional density of is on . Although the joint density is constant, and are not independent: the allowed range of changes with .
Polar coordinates and the standard normal pair
Let be independent variables. Their joint density is . Away from the origin define and choose the angle ; the excluded negative horizontal ray and origin have probability zero. The inverse map is
with determinant . Therefore
The factor is the area scaling: a narrow angular sector is wider at larger radius. The radial factor integrates to one by ; the angular factor integrates to one on . Their factorization on the product support proves that and are independent. In particular,
Concept lensGeometric
A random radius with an independent direction
The factorization gives a second precise description of the same random point. Instead of two independent coordinates, choose a radius with and, independently, a direction uniform on , and set , . Reading the computation backwards, the inverse determinant cancels the factor , so this pair has density : it is the standard normal pair.
This description makes rotation invariance visible. A rotation by replaces by taken modulo , which is again uniform and independent of , so the rotated pair has the same law. Every normalized density of the form has a uniform direction independent of its radius, by the same factorization . Independent coordinates alone do not give this: two independent uniform variables on are not rotationally symmetric.
The Jacobian factor and the shape of the support are the two places where polar computations most often go wrong.
Common mistakes
Common mistake
(Leaving out the area factor )
Writing the joint polar density of the normal pair as omits the area factor . The result is not even a density on : its integral is . Likewise is not standard normal: it is nonnegative and has density on .
Common mistake
(Reading independence from a formula on a non-product support)
If is uniform on the square , its polar density is , a function of times a constant. Yet and are dependent, because the support is not a product set. For we have , so . Hence , while (points near have radius near ) and . A factorized formula proves independence only on a product support.
Practice with solutions
Quick check
(Let be uniform on and . Find and .)
Decide how many inverse branches lie in for each .
Solution · Answer
Both roots lie in for , and each contributes . Their sum is on ; elsewhere the density is zero apart from endpoints. Thus , which agrees with .
Quick check
(In the triangular example, find and .)
Check that the region lies inside the support of .
Solution · Answer
The region , is a triangle of area , and it lies entirely inside the support because . Multiplying by the density two gives . Also .
Quick check
(For independent standard normals , find and the conditional probability .)
Use the radial tail and the independence of and .
Solution · Answer
. Since and are independent and is uniform on , ; the conditioning event has positive probability.
Quick check
(Let have density for , and set . Find for .)
Solve and evaluate at both roots.
Solution · Answer
The roots are , each in . Since at both roots, the branch rule gives , using . With , . The derivative vanishes at , but this single point has probability zero and the formula is needed only almost everywhere.