Vanishing Boundary Flux on Expanding Domains

Why Boundary Flux Needs Its Own Bound

In an integration-by-parts argument, an integral over a bounded region can produce a boundary term. To use that identity on an unbounded parameter space, one often applies it first on a large bounded region and then lets the region expand. The boundary term must be shown to vanish; it cannot simply be discarded because the boundary is far away. Its size depends both on the field being integrated and on how quickly the boundary area grows.

This issue arises in admissibility arguments when an integrated risk comparison is rewritten using derivatives of a prior density, a weight, or another vector field. The relevant question here is specific: what tail bounds guarantee that the resulting boundary contribution tends to zero?

Definition: Let \(G:\mathbb{R}^d\to\mathbb{R}^d\) be a vector field and let \(\Omega\) be a bounded region with outward unit normal \(n(\theta)\) on its boundary. Its signed boundary flux is \(\int_{\partial\Omega}G(\theta)\cdot n(\theta)\,dS(\theta)\), where \(dS\) denotes surface measure. Its absolute boundary contribution is \(\int_{\partial\Omega}|G(\theta)\cdot n(\theta)|\,dS(\theta)\). The absolute contribution bounds the absolute value of the signed flux.

A Surface-Area Bound on Expanding Balls

Write \(B_R=\{\theta\in\mathbb{R}^d:\|\theta\|<R\}\) and let \(s_{d-1}\) be the surface area of the unit sphere in \(\mathbb{R}^d\). The boundary \(\partial B_R\) has surface area \(s_{d-1}R^{d-1}\). Since the normal vector has unit length, \(|G(\theta)\cdot n(\theta)|\leq \|G(\theta)\|\). These facts give a direct sufficient condition.

Theorem: Suppose \(G:\mathbb{R}^d\to\mathbb{R}^d\) is continuous and, for some constants \(C<\infty\), \(\alpha>d-1\), and \(R_0<\infty\), \(\|G(\theta)\|\leq C(1+\|\theta\|)^{-\alpha}\) whenever \(\|\theta\|\geq R_0\). Then \(\int_{\partial B_R}|G(\theta)\cdot n(\theta)|\,dS(\theta)\to 0\) as \(R\to\infty\). Consequently, the signed flux also tends to zero.

Proof. On \(\partial B_R\), every point has norm \(R\), so for \(R\geq R_0\),

\begin{aligned} \int_{\partial B_R}|G(\theta)\cdot n(\theta)|\,dS(\theta) &\leq \int_{\partial B_R}\|G(\theta)\|\,dS(\theta)\\ &\leq s_{d-1}R^{d-1}C(1+R)^{-\alpha}. \end{aligned}

The last expression is bounded by \(Cs_{d-1}R^{d-1-\alpha}\), which tends to zero because \(\alpha>d-1\). Also, \(\left|\int_{\partial B_R}G\cdot n\,dS\right|\leq\int_{\partial B_R}|G\cdot n|\,dS\), so the signed flux tends to zero as well. \(\square\)

The exponent compares decay of the field with growth of the boundary. The theorem is sufficient, not necessary: signs or cancellations can make a signed flux vanish even when the absolute contribution does not. For an argument that must control a boundary term without relying on cancellation, the absolute bound is the safer quantity.

Worked Example: Finding the Polynomial Decay Threshold

For \(d\geq 2\), consider \(G(\theta)=\theta(1+\|\theta\|^2)^{-\beta/2}\), where \(\beta>0\). Determine directly when its absolute boundary contribution on \(\partial B_R\) tends to zero.

At a point on \(\partial B_R\), the outward normal is \(n(\theta)=\theta/R\). Therefore \(G(\theta)\cdot n(\theta)=R(1+R^2)^{-\beta/2}\), which is nonnegative and constant over the sphere. The absolute contribution is exactly

\int_{\partial B_R}|G\cdot n|\,dS =s_{d-1}R^{d-1}R(1+R^2)^{-\beta/2} =s_{d-1}R^d(1+R^2)^{-\beta/2}.

As \(R\to\infty\), \((1+R^2)^{\beta/2}/R^\beta\to 1\), so this expression is asymptotic to \(s_{d-1}R^{d-\beta}\). It tends to zero if \(\beta>d\), tends to \(s_{d-1}\) if \(\beta=d\), and diverges if \(\beta<d\). In this example, the threshold is exact. The field has norm of order \(R^{1-\beta}\), so the general theorem’s condition \(\alpha=\beta-1>d-1\) gives the same threshold \(\beta>d\).

Other Boundary Shapes and Boundary-Average Bounds

The same reasoning applies beyond balls if the surface area and the distance of the boundary from the origin are controlled. For example, suppose \(\Omega\) is a bounded region with a sufficiently regular boundary, the origin is in its interior, and \(\Omega_R=R\Omega\). Scaling gives \(\operatorname{area}(\partial\Omega_R)=R^{d-1}\operatorname{area}(\partial\Omega)\). Because \(\partial\Omega\) is bounded away from the origin, there is an \(m>0\) such that every \(\theta\in\partial\Omega_R\) has \(\|\theta\|\geq mR\). A bound \(\|G(\theta)\|\leq C(1+\|\theta\|)^{-\alpha}\) on these boundaries therefore gives an absolute flux bound of order \(R^{d-1-\alpha}\), which vanishes when \(\alpha>d-1\).

Sometimes a uniform, pointwise bound on \(G\) is inconvenient or false, while its boundary integral can still be controlled. Hölder’s inequality gives another sufficient test.

Result: Let \(\partial\Omega_R\) have finite surface area \(A_R\), and suppose \(G\) is measurable on that boundary with \(\int_{\partial\Omega_R}\|G\|^p\,dS<\infty\) for some \(p>1\). Then \(\int_{\partial\Omega_R}|G\cdot n|\,dS\leq A_R^{1-1/p}\left(\int_{\partial\Omega_R}\|G\|^p\,dS\right)^{1/p}\). In particular, if \(A_R=O(R^{d-1})\) and \(\int_{\partial\Omega_R}\|G\|^p\,dS=O(R^{-q})\), the absolute contribution tends to zero whenever \(q>(d-1)(p-1)\).

Proof. The unit length of \(n\) implies \(|G\cdot n|\leq\|G\|\). Applying Hölder’s inequality to \(\|G\|\) and the constant function \(1\) on \(\partial\Omega_R\) yields

\int_{\partial\Omega_R}|G\cdot n|\,dS \leq \left(\int_{\partial\Omega_R}\|G\|^p\,dS\right)^{1/p} \left(\int_{\partial\Omega_R}1\,dS\right)^{1-1/p}.

The second factor is \(A_R^{1-1/p}\), proving the inequality. Under the stated order bounds, the right side is \(O\!\left(R^{(d-1)(1-1/p)-q/p}\right)\). Its exponent is negative exactly when \(q>(d-1)(p-1)\), so it tends to zero under that condition. \(\square\)

Worked Example: A Decay Bound on Expanding Squares

In \(\mathbb{R}^2\), let \(\Omega_R=[-R,R]^2\) and let \(G(\theta)=(1+\|\theta\|)^{-2}v\), where \(v\) is a fixed vector with \(\|v\|=3\). Verify that the absolute boundary contribution vanishes.

The square has perimeter \(8R\). Every boundary point has norm at least \(R\), since at least one coordinate has absolute value \(R\). Thus \(\|G(\theta)\|\leq 3(1+R)^{-2}\) everywhere on the boundary. It follows that

\int_{\partial\Omega_R}|G\cdot n|\,dS \leq 8R\cdot 3(1+R)^{-2} =\frac{24R}{(1+R)^2} \leq \frac{24}{R}.

The last bound tends to zero. This verifies vanishing without calculating the signed flux face by face, and illustrates the balance between perimeter growth, of order \(R\), and field decay, of order \(R^{-2}\).

Exponential Tails and a Whole-Space Consequence

Polynomial decay is not the only useful condition. Exponential decay dominates every fixed power of the radius, so it easily absorbs the surface-area factor in any fixed dimension. When the field is the one appearing in an integration-by-parts identity, vanishing flux can also identify the limiting integral of its divergence.

Worked Example: Gaussian-Type Decay in Three Dimensions

Take \(d=3\) and \(G(\theta)=\theta\exp(-\|\theta\|^2/4)\). On \(\partial B_R\), the outward normal is \(\theta/R\), and hence \(G\cdot n=R\exp(-R^2/4)\). Since the sphere has area \(4\pi R^2\), its absolute flux is

\int_{\partial B_R}|G\cdot n|\,dS =4\pi R^3\exp(-R^2/4).

Taking logarithms of the radius-dependent factor gives \(3\log R-R^2/4\), which tends to \(-\infty\); therefore \(R^3\exp(-R^2/4)\to0\). The boundary contribution vanishes. The argument depends on the exponential tail overcoming the factor \(R^3\), not on treating the sphere as having negligible area.

Theorem: Suppose \(G\) is continuously differentiable on \(\mathbb{R}^d\), \(\operatorname{div}G\) is absolutely integrable on \(\mathbb{R}^d\), and \(\int_{\partial B_R}G\cdot n\,dS\to0\) as \(R\to\infty\). Then \(\int_{\mathbb{R}^d}\operatorname{div}G(\theta)\,d\theta=0\).

Proof. The divergence theorem applies on each ball and gives \(\int_{B_R}\operatorname{div}G(\theta)\,d\theta=\int_{\partial B_R}G\cdot n\,dS\). Because \(\operatorname{div}G\) is absolutely integrable, dominated convergence applied to \(\mathbf{1}_{B_R}(\theta)\operatorname{div}G(\theta)\) shows that the left side tends to \(\int_{\mathbb{R}^d}\operatorname{div}G(\theta)\,d\theta\). The assumed boundary-flux limit makes the right side tend to zero. Equating the limits proves the result. \(\square\)

This consequence is useful only when both hypotheses have been checked. Vanishing flux alone does not justify exchanging a limit with a whole-space integral; here absolute integrability supplies the needed justification. In a risk comparison, the field and its divergence come from the particular integration-by-parts identity, so their regularity and integrability must be verified for that application.

Common Pitfall: Pointwise Decay Is Not Enough

A frequent error is to observe that \(\|G(\theta)\|\to0\) as \(\|\theta\|\to\infty\) and conclude that the boundary contribution vanishes. This ignores the growing boundary area. For instance, take \(G(\theta)=\theta(1+\|\theta\|^2)^{-1/2}\) in dimension \(d\geq2\). Its norm tends to \(1\), and its boundary flux is

\int_{\partial B_R}G\cdot n\,dS =s_{d-1}R^{d-1}\frac{R}{\sqrt{1+R^2}} =s_{d-1}\frac{R^d}{\sqrt{1+R^2}}.

This grows like \(s_{d-1}R^{d-1}\), rather than tending to zero. Even a field that tends to zero needs a sufficiently fast rate: the relevant comparison is its decay against the boundary’s area growth. In a Blyth-type risk argument, verify a bound on the actual boundary integral (or its absolute value), not merely a pointwise limit for the integrand.

Key takeaway: A reliable sufficient check is to bound the absolute boundary contribution by boundary area times a uniform bound for the field. On expanding balls in \(d\) dimensions, polynomial field decay faster than \(R^{-(d-1)}\) makes this bound vanish. More refined boundary-integral bounds can be used when a uniform bound is unavailable.

Check Your Understanding

Use the surface-area factor as well as the field’s tail behavior when answering.

On \(\partial B_R\subset\mathbb{R}^4\), a field satisfies \(\|G(\theta)\|\leq C(1+\|\theta\|)^{-3}\). Does the theorem guarantee that its absolute boundary contribution tends to zero? Explain using the powers of \(R\).
For the field in the polynomial-threshold example, what happens to the absolute boundary contribution when \(\beta=d\)?
Suppose \(A_R=O(R^2)\), \(p=2\), and \(\int_{\partial\Omega_R}\|G\|^2\,dS=O(R^{-3})\). Does the boundary \(L^p\) result guarantee vanishing? Check the strict inequality on \(q\).
Why does \(\|G(\theta)\|\to0\) by itself not establish that the boundary flux tends to zero?
In the whole-space divergence result, what hypothesis justifies taking the limit of the integrals over \(B_R\)?