Past exam of the mathematics course of the University of Cambridge / 2007 / iii / Paper 80 / 4

Hadamard well-posedness requires existence of a solution for every admissible datum, uniqueness, and continuous dependence of the solution on the datum in the specified normed spaces. An inverse problem is ill-posed if any of these requirements fails. The choice of spaces and norms is part of this assertion: continuity in one norm does not establish continuity in a stronger norm.

For a bounded linear operator $A$, existence requires $y\in \mathrm{ran}A$, and uniqueness requires $\ker A=\{0\}$. Even when both hold for exact data, inversion can be unstable. For example, if $\Vert x_N{\Vert }_X=1$ but $\Vert A{x}_N{\Vert }_Y\to 0$, data $A{x}_N$ approach zero while their solutions do not approach zero. Thus $A^{-1}$ on its range is not continuous. Perturbed data can also leave the range entirely, so an exact solution need not exist after measurement error. These are distinct mechanisms of ill-posedness; the inverse scattering problem below exhibits unstable continuation even on data for which an outgoing continuation exists.

Write $r=|\mathbf{x}|$, $\theta =\mathbf{x}/r$, and choose $R$ so that the sphere $r=R$ strictly encloses the scatterer. Normalize the spherical harmonics to be an orthonormal basis of $L^2(S^2)$, with surface measure on the unit sphere. The outgoing spherical Hankel function is denoted here by $h_n^{(1)}$, corresponding to the printed $H_n^{(1)}$. For fixed $n$, its large-argument behavior isConsequently the coefficient of $Y_n^m$ in the outgoing expansion is asymptotic to $a_n^m{e}^{ikr}/r$: the factors satisfy $i^{n+1}(-i{)}^{n+1}=1$. With the far-field pattern defined by $u_s(r\theta )=e^{ikr}{u}_{\infty }(\theta )/r+O(r^{-2})$, we obtainOne can first project onto each spherical harmonic and then take the large-$r$ limit. This avoids assuming that the fixed-order asymptotic is uniform in $n$. The Sommerfeld radiation condition selects the outgoing spherical Hankel functions rather than their incoming counterparts.

Let $g(\theta )=u_s(R\theta )$. Its coefficients are $g_n^m=d_n{a}_n^m$, where $d_n=k{i}^{n+1}{h}_n^{(1)}(kR)$. Since the field is smooth on a sphere outside the scatterer, $g\in L^2(S^2)$; Parseval's identity givesThe required coefficient condition uses fixed-argument growth of spherical Hankel functions, a different limit from the far-field limit. For fixed $t>0$,For completeness, the series of the ordinary Bessel function gives $J_{-n-1/2}(t)\sim (t/2{)}^{-n-1/2}/\Gamma (1/2-n)$. The identities $y_n(t)=(-1{)}^{n+1}\sqrt{\pi /(2t)}{J}_{-n-1/2}(t)$ and $\Gamma (1/2-n)=(-4{)}^{n}n!\sqrt{\pi }/(2n)!$ give the first formula; the corresponding positive-order series gives the estimate for $j_n$. The successive series corrections at fixed $t$ are $O(1/n)$. Applying Stirling's approximation to the factorial ratio yieldsIn particular there are positive constants giving upper and lower bounds by the last weight for all sufficiently large $n$. The finite trace norm is therefore equivalent to the spherical far-field range conditionwhere the $n=0$ weight is defined to be one. The same conclusion holds at every larger sphere on which the outgoing field is defined.

The printed comparison with $\infty$ must mean a finite sum: a literal non-strict inequality allows a divergent sum. Also, the supplied one-sided big-$O$ bound alone does not imply this necessary condition from a finite trace norm; a lower bound is needed in that direction. The two-sided asymptotic above supplies it. Conversely, its upper bound makes the weighted condition sufficient for an $L^2$ trace at the chosen sphere. This is a condition on exterior-field continuation, not a sufficient condition that the coefficients come from a particular Dirichlet obstacle.

Use the $L^2(S^2)$ data norm. The coefficients of the measurement error are $e_n^m=b_n^m-a_n^m$, so Parseval's identity givesAn unregularized continuation of these measurements to $r=R$ replaces $a_n^m$ by $b_n^m$ in the outgoing spherical harmonic expansion. Whenever this continuation has a finite trace, its error isChoose a measurement error supported on a single normalized spherical harmonic, with $e_N^0=\delta$ and all other coefficients zero. The data error has norm exactly $\delta$, butThis perturbation theory is a finite expansion, so it defines an actual outgoing solution of the Helmholtz equation outside any sphere of positive radius. It is not merely an example of nonconvergent formal coefficients. Taking instead $e_N^0=|d_N{|}^{-1}$ gives data errors tending to zero and trace errors of norm one, which directly proves discontinuity of inversion. General square-summable measurement noise can even violate the spherical far-field range condition, leaving no $L^2$ continuation to the chosen sphere.

For obstacle reconstruction one would continue the field toward the unknown boundary and look for a surface on which $u_i+u_s=0$, the Dirichlet boundary condition. The unbounded spherical far-to-near-field continuation shows why small far-field errors can cause arbitrarily large errors in the field used to estimate that surface. Locally, a differentiable normal displacement $\eta$ of a candidate boundary must obey the linearized condition $\Delta u_s+\eta {\partial }_n(u_i+u_s)=0$. Thus a large continued-field error, or a small normal derivative, defeats an uncontrolled boundary estimate.

There is a necessary qualification to the wording about errors in the scatterer itself. The calculation proves an unbounded field-continuation error. An arbitrary perturbed outgoing field need not correspond to any Dirichlet obstacle, and no admissible class or metric on obstacle shapes is specified. It therefore does not by itself prove an arbitrarily large geometric error in every constrained reconstruction method. Such a claim requires a shape model and an analysis of the boundary-extraction step. The instability displayed here is exactly the mechanism that an unregularized reconstruction must control.

To illustrate Tikhonov regularization precisely, take the unknown to be the trace $g$ on a chosen enclosing sphere of radius $R$. Use Hilbert spaces $X=Y=L^2(S^2)$ and the forward linear operatorAt real positive $kR$, $d_n\ne 0$, because the real Spherical Bessel functions $j_n,y_n$ have a nonzero Wronskian and cannot vanish simultaneously. Moreover $|d_n|\to \infty$. Thus $A$ is an injective compact operator: truncating to finitely many harmonic orders approximates it in operator norm, since the remaining diagonal entries tend uniformly to zero. Its inverse is unbounded. This is a linear field reconstruction; estimating the obstacle from the reconstructed field remains a further nonlinear step.

Given coefficients $b_n^m$ of measured data $y^{\delta }$, minimize the quadratic functionalVarying $g$ in every direction gives $(A^{\ast }A+\alpha I)g_{\alpha }^{\delta }=A^{\ast }{y}^{\delta }$. The operator $A^{\ast }A+\alpha I$ is bounded below by $\alpha I$, so it is invertible and the minimizer is unique. In coefficients the minimization is separable: $A^{\ast }y$ has coefficient $b_n^m/\overline{d_n}$ and $A^{\ast }A$ has multiplier $|d_n{|}^{-2}$. Hence the Tikhonov regularization of spherical far-field continuation isThe implied regularized far-field coefficients are $b_n^m/(1+\alpha |d_n{|}^2)$. This suppresses precisely the high-order coefficients that make backward continuation unstable.

Let $g^{†}$ be the exact trace and $y=A{g}^{†}$, with $\Vert y^{\delta }-y\Vert \le \delta$. The noise-bias decomposition for linear regularization isFor a singular value $s=|d_n{|}^{-1}$, the noise multiplier is $s/(s^2+\alpha )$. Since $s^2+\alpha \ge 2s\sqrt{\alpha }$, its supremum is at most $1/(2\sqrt{\alpha })$. ConsequentlyThis Tikhonov stability bound measures the trace norm, not a distance between obstacle boundaries. The bias coefficient is $\alpha /(|d_n{|}^{-2}+\alpha )$ times the exact coefficient. It tends to zero for each fixed $n,m$, is bounded by one, and $\sum |g_n^{†,m}{|}^2<\infty$. Splitting off a finite sum and bounding its tail therefore proves $B(\alpha )\to 0$ as $\alpha \to 0$.

A convergent regularization parameter choice thus satisfiesIn nondimensionalized variables, $\alpha ={\delta }^p$ with $0<p<2$ is one such rule. Without a smoothness assumption on the exact trace there is no universal bias rate and no universal optimal value of $\alpha$. For a concrete normal-operator source condition, suppose $g^{†}=A^{\ast }Aw$ and $\Vert w\Vert \le C$. Each bias multiplier on $w$ is $\alpha s^2/(s^2+\alpha )\le \alpha$, so $B(\alpha )\le C\alpha$. Minimizing the resulting bound gives, for $C>0$,This optimizes the displayed worst-case bound under that additional source condition, rather than guaranteeing an exact optimum for every datum.

If only the effect of measurement error is minimized, the multiplier bound decreases when $\alpha$ increases. Indeed $\Vert (A^{\ast }A+\alpha I{)}^{-1}{A}^{\ast }\Vert \le \Vert A\Vert /\alpha \to 0$ as $\alpha \to \infty$. But the reconstruction then approaches zero and its bias approaches $\Vert g^{†}\Vert$. Suppressing noise alone does not optimize reconstruction: $\alpha$ must balance noise amplification against regularization bias. A geometrically accurate obstacle estimate additionally needs justified shape constraints and a stable procedure for extracting a boundary from the regularized field.