Remote temperature sensing in 2D and the Bergman kernel Steven R. Bell, Leah McNabb Abstract . We explore the problem of estimating the steady state temperature in a two-dimensional domain at a point knowing the temperature to high order at another point. We find connections to the Bergman kernel of the domain, Runge’s theorem, and approximate null quadrature identities. 1991 Mathematics Subject Classification 30C40; 31A35 To celebrate the life and work of Marek Jarnicki 1. Introduction Suppose that \Omega is a bounded domain in the complex plane and u is a real valued harmonic function on \Omega (which we can think of as representing the steady-state temperature in a \Omega -shaped metal plate). Suppose further that u is bounded by M>0 on \Omega , i.e., |u(z)|0 , is it possible to find a positive integer N and real coefficients c_{nm} for 0\leq n,m\leq N such that \left|u(b)-\sum_{0\leq n,m\leq N}c_{nm}\frac{\partial^{n+m}u}{\partial^{n}x\,\partial^{m}y}(a)\right|<\epsilon? We emphasize that N and the coefficients c_{nm} might depend on a , b , \epsilon , and M , but not on u . We like to think of this problem as follows. Suppose we know that the (steady-state) temperature in the two-dimensional state of Indiana is bounded by 130^{\circ} . If we can measure the temperature in Indianapolis to a high enough degree of accuracy to be able to compute finitely many partial derivatives, can we estimate the temperature in West Lafayette to within one-half degree? We remark that one never needs to use difference quotients to compute partial derivatives of a harmonic function. One can always reduce them to an integral of the function times a smooth function on a circle about the point of interest (or smooth simple closed curve containing the point) via the Poisson integral formula. We will investigate a method to solve the sensing problem that involves the Bergman kernel and “approximate null quadrature identities” for holomorphic functions that will lead to some interesting connections to Runge’s theorem and approximation problems for holomorphic functions. 2. Proof of concept We will always assume that \Omega is a bounded domain in the complex plane and that a and b are distinct points in \Omega . Let K(z,w) denote the Bergman kernel for \Omega and write K_{a}(z)=K(z,a) . Also write K_{a}^{0}(z)=K_{a}(z) and K_{a}^{m}(z)=\left.\frac{\partial^{m}}{\partial\bar{w}^{m}}K(z,w)\right|_{w=a}. If h is a function in the Bergman space, then h(z)=\langle h,K_{z}\rangle, the inner product being the L^{2} inner product on \Omega with respect to area measure. Furthermore, h^{(m)}(z)=\langle h,K_{z}^{m}\rangle. Hence, if a function in the Bergman space is orthogonal to K_{a}^{m} for m=0,1,2,\dots , then it vanishes to infinite order at a and is consequently identically zero on \Omega . This implies that the complex linear span of the functions \{K_{a}^{m}(z)\,:\,m=0,1,2,\dots\} is dense in the Bergman space. Therefore, given \epsilon>0 , there is an element in this span that is within \epsilon of K_{b} in L^{2} norm. Say \lambda:=K_{b}-\sum_{m=0}^{N}c_{m}K_{a}^{m} has L^{2} norm less than \epsilon . Then, (2.1) \langle h,\lambda\rangle=h(b)-\sum_{m=0}^{N}{\bar{c}}_{m}h^{(m)}(a) for functions h in the Bergman space, and we can estimate (2.2) \left|h(b)-\sum_{m=0}^{N}{\bar{c}}_{m}h^{(m)}(a)\right|\leq\epsilon\|h\| where the norm is the L^{2} norm. Consequently, if |h|1 . Hence, h^{\prime}=u_{x}-iu_{y}\qquad\text{and}\qquad h^{(m)}=\frac{\partial^{m}u}{\partial x^{m}}-i\frac{\partial^{m}u}{\partial x^{m-1}\,\partial y} if m>1 . Assume that an approximate null quadrature identity of the form ( 2.2 ) holds for our function h in the Bergman space associated to \Omega_{1} . Write c_{m}=A_{m}+iB_{m} in terms of real and imaginary parts. Now note that the real part of \displaystyle h(b)- \displaystyle\sum_{m=0}^{N}{\bar{c}}_{m}h^{(m)}(a)= \displaystyle h(b)-(A_{0}-iB_{0})u(a)-\sum_{m=1}^{N}(A_{m}-iB_{m})\left[\frac{\partial^{m}u}{\partial x^{m}}-i\frac{\partial^{m}u}{\partial x^{m-1}\,\partial y}\right] can be expressed as u(b) minus a real linear combination of u and some of its partial derivatives of order less than or equal to N at a . To be precise, the real part is u(b)-A_{0}u(a)-A_{1}u_{x}(a)+B_{1}u_{y}(a)-\sum_{m=2}^{N}\left[A_{m}\frac{\partial^{m}u}{\partial x^{m}}-B_{m}\frac{\partial^{m}u}{\partial x^{m-1}\,\partial y}\right](a), which is bounded in absolute value by \epsilon\|h\| , where the norm is the L^{2} norm on \Omega_{1} . We now claim that \|h\| is less than or equal to a positive constant times the L^{2} norm of u on the larger set \Omega . Since clearly |u|\leq|h| on \Omega_{1} , we are reduced to estimating |v| . Estimating the harmonic conjugate of a given harmonic function on a subset of a domain makes a good exercise in a graduate course in complex analysis. One approach could be to use the line integral formula for v to show that we can bound v on \Omega_{1} by estimating first partial derivatives of u on a slightly larger relatively compact open set \Omega_{2} of \Omega that contains the closure of \Omega_{1} . To do this, we might cover the closure of \Omega_{1} by finitely many small discs of radius d/2 where d is less than the distance of \Omega_{1} to the boundary of \Omega_{2} . The value of the first partial derivatives of u inside the discs of radius d/2 can be expressed as integrals of u against smooth functions on the circles of radius d bounding the discs by differentiating under the integral in the Poisson formula. This completes the proof that N and appropriate constants exist in the temperature sensing problem. We now turn to describing a constructive proof of the holomorphic sensing problem, which would be the basis of a constructive proof of the harmonic sensing problem using the techniques of the paragraph above. 3. A more constructive approach The Hilbert space arguments in §2 are purely existential. In this section, we describe a method to explicitly determine the order N and to compute the coefficients in an identity like ( 2.2 ) for holomorphic functions. The method is based on the observation that the basic problem is rather straightforward in the unit disc when a=0 and b\neq 0 . Indeed, since the Bergman kernel of the unit disc is K(z,w)=\pi^{-1}(1-z\,\bar{w})^{-2} , an easy calculation yields K_{0}(z)=\pi^{-1} and K_{0}^{m}(z)=\pi^{-1}(m+1)!z^{m} . Since K_{b}(z)=\pi^{-1}(1-z\,\bar{b})^{-2} is holomorphic on the disc of radius 1/|b| , we may approximate K_{b}(z) uniformly on any disc D_{r}(0) about the origin of radius r with 10 extends holomorphically past the boundary of \widetilde{\Omega} . Let B=f(b) and let F:=f^{-1} . Let K(z,w) continue to denote the Bergman kernel of the unit disc and let \widetilde{K}(z,w) denote the Bergman kernel of \widetilde{\Omega} . We will need the transformation formula for the Bergman kernels in the form (3.1) f^{\prime}(z)K(f(z),w)=\widetilde{K}(z,F(w))\,\overline{F^{\prime}(w)}. Note that setting w=B in this formula yields f^{\prime}(z)K_{B}(f(z))=\widetilde{K}_{b}(z)\,\overline{F^{\prime}(B)} , and since F^{\prime}(B)=1/f^{\prime}(b) , we have (3.2) \overline{f^{\prime}(b)}\,f^{\prime}(z)K_{B}(f(z))=\widetilde{K}_{b}(z). We can differentiate equation ( 3.1 ) m -times with respect to \bar{w} and set w=0 to also obtain (3.3) f^{\prime}(z)K_{0}^{m}(f(z))=\frac{\partial^{m}}{\partial\bar{w}^{m}}\left[\widetilde{K}(z,F(w))\,\overline{F^{\prime}(w)}\right]_{w=0}, which is an explicit complex linear combination of elements \widetilde{K}_{a}^{n} with 0\leq n\leq m . Observe that by making \Lambda(z):=K_{B}(z)-\sum_{m=0}^{N}c_{m}K_{0}^{m}(z) sufficiently uniformly small on a disc D_{r}(0) with r>1 where r is sufficiently close to 1 , since f^{\prime} extends holomorphically past the boundary, f^{\prime}(z)\Lambda(f(z)) can also be made uniformly small on an open set containing the closure of \widetilde{\Omega} . We can now multiply f^{\prime}(z)\Lambda(f(z)) by \overline{f^{\prime}(b)} and use the transformation formulas ( 3.2 ) and ( 3.3 ) to obtain our goal of finding an explicit combination of the form \widetilde{K}_{b}(z)-\sum_{m=0}^{N}\tilde{c}_{m}{\widetilde{K}}_{a}^{m}(z) that is uniformly small on an open subset of \Omega containing the closure of \widetilde{\Omega} . 4. Connections to Runge’s theorem We will now show that certain Runge approximation problems involving rational functions without poles on the exterior of a domain that tend to zero at infinity are closely connected to Bergman kernel approximation in the interior. Since inversion by a mapping of the form 1/(z-a) where a is a point in the domain transforms rational functions without poles on the exterior to rational functions without poles inside a bounded domain, it will also be seen that rational approximation problems in a domain are equivalent to Bergman approximation problems. The connection will be exhibited via the following theorem from [ 1 , p. 90] , which has just been waiting for a proper application. Suppose \Omega is a bounded simply connected domain bounded by a simple closed C^{\infty} -smooth curve \gamma and h is a holomorphic function on \Omega in L^{2}(\Omega) . Define a holomorphic function H(w) on the outside of \Omega via \overline{H(w)}=\iint_{z\in\Omega}\frac{h(z)}{\bar{z}-\bar{w}}\ dA_{z}, where dA_{z} represents Lebesgue area measure in the z variable. Notice that H and all its derivatives tend to zero at infinity. Theorem 4.1 . Given a positive integer N , there is a positive integer M and a constant C>0 such that, if the C^{M} -norm \|H\|^{M} of H on the outside of \Omega is finite, then h is in C^{N}(\Omega) and the C^{N} -norm \|h\|_{N} satisfies \|h\|_{N}\leq C\|H\|^{M}. Consequently, h(z) extends C^{\infty} -smoothly to the boundary of \Omega from the inside if and only if H(w) extends C^{\infty} -smoothly to the boundary of \Omega from the outside. This theorem gives us a direct connection between the problem of finding an approximate quadrature identity of the form ( 2.2 ) and the problem of approximating (z-b)^{-1} by polynomials in (z-a)^{-1} on the exterior of \Omega . Indeed, if \Omega is a simply connected domain with smooth boundary that contains a , then h^{(m)}(a)=\left when h is in the Bergman space associated with \Omega . In particular, if we let h(z)=(w-z)^{-1} where w is a point outside of the closure of \Omega , then \frac{m!}{(w-a)^{m+1}}=\left and the dual nature of our two problems via Theorem 4.1 is clear. A fascinating possibility is that the way we have suggested obtaining an approximate quadrature identity using the Bergman kernel might correspond to a superior way of doing the classical Runge approximation of (z-b)^{-1} by polynomials in (z-a)^{-1} on the exterior of \Omega . Runge’s original proof of one of his famous theorems used an explicit technique to approximate 1/(z-b) by a polynomial in 1/(z-a) when b\neq a on a large open set away from a curve \gamma connecting b to a . The argument involved gradually moving a point along \gamma to get uniform approximations to 1/(z-b) on the open set given by the set of points that are more than a distance of \delta>0 from the curve. The classic fact is explained nicely in the textbook [ 6 , pp. 63-64] . We summarize it here. Let \gamma be a curve connecting b to a parametrized by z(t) , \beta\leq t\leq\alpha and let \delta>0 be less than |a-b| . Let U_{\delta} denote the (open) set of points that are a distance greater than \delta away from the curve \gamma . We now describe step one of the construction. Since \delta is less than the distance from b to a , there is a point a_{1}=z(t_{1}) on \gamma that is a distance \delta from b with t_{1} as small as possible. We may approximate the function (z-b)^{-1} by finitely many terms of a Laurent expansion centered at a_{1} uniformly, say within \epsilon>0 , outside a disc of radius 2\delta about a_{1} . Let P_{1}(z) denote the polynomial in (z-a_{1})^{-1} that represents the finitely many terms in the Laurent expansion. The next step is to replace b by a_{1} , \beta by t_{1} , and the function (z-b)^{-1} by P_{1}(z) in step one to obtain a point a_{2} along the curve and a polynomial P_{2}(z) in (z-a_{2})^{-1} that is uniformly close, within \epsilon , to P_{1}(z) outside a disc of radius 2\delta about a_{2} . Note that P_{2}(z) is within 2\epsilon of (z-b)^{-1} outside the intersection of the previous sets on which the approximations were made. This process can be repeated until the distance from the N -th point a_{N} to a is less than or equal to \delta . In the last step, we let a=a_{N+1} and approximate P_{N}(z) outside the disc of radius 2\delta about a by a polynomial in (z-a)^{-1} . We can now replace \epsilon by \epsilon/(N+1) to obtain a uniform approximation within \epsilon of (z-b)^{-1} by a polynomial in (z-a)^{-1} in the open set U_{2\delta} . We remark that this idea of Runge could have been the basis for a solution to the remote sensing problem. Indeed if \Omega is a simply connected domain bounded by a smooth curve which compactly contains the set of points that are within a distance \delta of the curve \gamma , and if the modulus of \frac{1}{z-b}-\sum_{j=1}^{N}\frac{A_{j}}{(z-a)^{j}} is bounded by \epsilon on U_{\delta} , then integrating this function against a holomorphic function h\,dz around the boundary of \Omega and using the higher order Cauchy integral formulas yields an approximate one-point quadrature identity for h that can be bounded by a constant times \epsilon times the length of the boundary curve times the supremum of |h| on the boundary. Then the same ideas used in § 2 to get a solution to the sensing problem for harmonic functions can be applied to the holomorphic identity to solve the harmonic problem. Runge’s approximation method is rather crude and likely not an efficient way to solve the problem, whereas the approximation scheme on the unit disc via Taylor approximations to a rational function with poles outside the unit circle is more efficient. The method on the unit disc pulls back to a method on a simply connected domain via the Riemann map. It seems possible that these ideas might find connections to the work of L. N. Trefethen ( [ 7 , 8 ] ) on rational approximation and numerical analytic continuation. 5. Further musings on the subject The techniques of § 2 - 3 are not sensitive to the original region being simply connected. Indeed, the first step was to shrink the domain down down to a simply connected domain that contains both points a and b . An intriguing idea is to design “probe domains” tailored to solve the temperature sensing problem where the Riemann maps involved are explicit and perhaps easier to compute with. Given a possibly unbounded and non-simply connected domain \Omega and points a and b in \Omega , let \gamma be a smooth curve in \Omega such that a and b are both points interior to the curve, i.e., in the curve, but not at either endpoint. Next, if \gamma is parametrized by a smooth complex function z(t) , 0\leq t\leq 1 , approximate the real and imaginary parts of z(t) by polynomials and make minor adjustments to get a complex valued polynomial P(t) that agrees with z(t) at the two image points a and b that stays close to the original curve and stays in \Omega . We may also assume that P^{\prime}(t) is nonvanishing on [0,1] . Let \Gamma denote the curve in \Omega parametrized by P(t) . We next let t go complex so that P(\tau) maps a (perhaps thin) open simply connected set U about [0,1] in the complex plane one-to-one onto a snake-like domain about the curve \Gamma that is compactly contained in \Omega . Let E be an ellipse (or rectangle, if preferred) that contains [0,1] , is symmetric about the x -axis, and is a subset of U . The conformal map from the unit disc onto E is well known. By composing next with P(z) we get a conformal map from the unit disc to a “probe domain” in \Omega that contains a and b . It is easy to arrange that this map sends the origin to a . We may now use the techniques of § 3 to solve the holomorphic and the heat sensing problems. It is interesting to note that, in the realm of bounded domains \Omega with C^{\infty} -smooth boundary, it is possible to approximate K_{b} by a finite complex linear combination of the functions K_{a}^{m} in the topology of C^{\infty}(\overline{\Omega}) (see [ 2 ] ). Coming up with better constructive ways to do this might lead to better constructive ways to solve the sensing problems. It is reasonable to consider the problems of this paper in several complex variables. Indeed, the topic is closely connected to the subject of “one-point quadrature domains,” and it has been shown that many of the results in one variable on these domains generalize to several complex variables (see [ 3 ] and [ 4 ] ). The soft analysis in the “proof of concept” section § 2 yields that there exists a solution to the holomorphic sensing problem in several variables. The second author’s proof that one-point quadrature domains in {\mathbb{C}} are also one point quadrature domains for harmonic functions carries over to several complex variables to show that one-point quadrature domains for holomorphic functions in several variables are also one point quadrature domains for pluriharmonic functions. Following the same reasoning we did in § 2 and restricting to a simply connected subregion containing the two points yields that we can solve the sensing problem for pluriharmonic functions (which have nothing to do with temperature problems). The fact that there is no Riemann mapping theorem in several complex variables means alternative means to a more constructive approximation problem would have to be found.