Differentiating General Integrals #
by Shuang Zhao
In the previous section, we have discussed the differentiation of 1D Riemann integrals.
In what follows, we discuss the differentiation of a general Lebesgue integral $I(\theta)$ over some bounded and open domain $\domain$ associated with measure $\mu$:
For practical rendering problems, the domain $\domain$ can be:
The surface of the unit sphere
$\sph := \{ \bx \in \real^3 : \| \bx \| = 1 \}$;The union
$\calM$of all object surfaces;The path space (under Veach’s path-integral formulation).
We assume that $f(\bx, \theta)$ is piecewise continuous with a zero-measure extended boundary $\boundary = \partial\domain \cup \Delta\domain$ comprising the domain boundary $\partial\domain$ and all jump discontinuity points denoted as $\Delta\domain$—which are illustrated below as black and orange curves, respectively:
Further, we assume the extended boundary $\boundary$ to be associated with a unit-normal field $\bn$.
That is, for any point $\bx$ on the extended boundary, we have the boundary normal $\bn(\bx)$ specified.
We note that, although this assumption may seem restrictive, it is the case for most, if not all, problems of our interest.
Reynolds Transport Theorem #
In general, when the domain $\domain$ of integration depends on the parameter $\theta$, so does the extended boundary $\boundary$.
In this case, the derivative of Eq. \eqref{eqn:I} with respect to $\theta$ is given by Reynolds transport theorem.
This theorem—which is a generalization of the result from the previous section—states that:
In this equation, the interior integral is obtained by exchanging differentiation and integration operations.
The Boundary Integral #
We now examine the boundary component of Eq. \eqref{eqn:dI_reynolds} more closely.
Domain of Integration #
The boundary integral is over the extended boundary $\boundary(\theta)$.
In practice:
When the domain
$\domain(\theta) = (a(\theta), b(\theta)) \subset \real$of the ordinary integral \eqref{eqn:I} is a 1D interval (with$\mu$being the Borel measure),$\boundary(\theta)$is a discrete set of jump discontinuity points including the interval’s endpoints$a(\theta)$and$b(\theta)$. Further, the boundary integral reduces to the sum over all these discontinuity points—as presented in the full derivative from the previous section.When
$\domain(\theta) \subset \real^2$is a 2D region with$\mu$being the area measure, the extended boundary$\boundary(\theta)$comprises a set of curves (as illustrated in the figure above) with$\dmu$being the curve-length measure.When
$\domain(\theta) \subseteq \sph$is a curved 2D region within the surface$\sph$of the unit sphere and$\mu$the solid-angle measure, the extended boundary$\boundary(\theta)$comprises a set of spherical curves with$\dmu$being the curve-length measure.When
$\domain(\theta) \subset \real^3$is a 3D volume and$\mu$the volume measure, the extended boundary$\boundary(\theta)$becomes 2D surfaces with$\dmu$being the surface-area measure.
Definition of $\Delta f$
#
In Eq. \eqref{eqn:dI_reynolds}, another important component of the boundary integral is $\Delta f(\bx, \theta)$ that captures the difference in $f(\bx, \theta)$ across a discontinuity boundary at $\bx$.
Precisely, let $\bx_{\bot}(\epsilon)$ be a regular curve parameterized by $\epsilon$ (illustrated as the gray dashed line in the figure above) that resides inside the domain $\domain(\theta)$ and satisfies $\bx_{\bot}(0) = \bx$.
When the domain $\domain(\theta)$ is “flat”, for example, we have
where $\bn(\bx)$ is the unit normal of the extended boundary at $\bx$.
Then,
Normal Change Rate $\vel$
#
Another key component of the boundary integral in Eq. \eqref{eqn:dI_reynolds} is the normal change rate term $\vel$ which, at a high level, captures how fast the extended boundary $\boundary(\theta)$ evolves along the normal $\bn$ (with respect to $\theta$).
Precisely, for any $\bx \in \boundary(\theta)$, we have
where “$\cdot$” denotes the dot (or scalar) product of two vectors.
Parameterizing Discontinuity Boundaries #
Evaluating the derivative $\D\bx/\D\theta$ on the right-hand side of Eq. \eqref{eqn:vel} requires the extended boundary $\boundary(\theta)$ to be locally parameterized.
Specifically, for each $\theta$ and $\bx \in \boundary(\theta)$, we assume without loss of generality that $\alphaBnd(\boldsymbol{\xi}, \theta)$ is a differentiable and bijective function satisfying that:
$\alphaBnd(\cdot, \theta)$maps an open ball centered at the origin$\boldsymbol{0}$to an open neighborhood around$\bx$within the extended boundary$\boundary(\theta)$;$\alphaBnd(\boldsymbol{0}, \theta) = \bx$.
Then,
Parameterization Independency #
The derivative $\D\bx/\D\theta$ defined in Eq. \eqref{eqn:vel_raw} is parameterization-dependent.
That is, it depends on the choice of the local parameterization specified by the function $\alphaBnd$.
On the contrary, the normal change rate $\vel(\bx)$ from Eq. \eqref{eqn:vel} is known to be parameterization-independent.
In other words, regardless to the choice of $\alphaBnd$, $\vel(\bx)$ remains identical.