Introduction
Boundary integral methods reduce elliptic problems with boundary or interface conditions to layer potentials on lower-dimensional boundaries. However, the kernels inherit the singularity of the fundamental solution, and standard quadratures lose accuracy at target points on or close to the boundary. Evaluation directly on the curve (the singular case) is central to boundary integral equations [21 , 29 ] . Evaluation near the curve (the nearly singular case) arises in interacting drops or vesicles in Stokes flow [37 , 42 ] , and in grid-based methods requiring values at Cartesian points adjacent to an interface [5 , 31 , 32 , 33 ] . Here we derive high-order regularized kernels for two-dimensional layer potentials with regularization error O ( δ p ) O(\delta^p) O ( δ p ) (p = 3 , 5 , 7 p = 3,5,7 p = 3 , 5 , 7 ), allowing accurate evaluation by the periodic trapezoidal rule without specialized near-singularity quadratures. We treat harmonic single- and double-layer potentials, the Stokeslet, stresslet, and the pressure in Stokes flow. The formulas hold uniformly near and on the boundary.
The nearly singular regime is challenging because, for an analytic periodic integrand, the convergence rate of the trapezoidal rule is governed by the distance from the real axis to the nearest singularity in the complex parameter plane [48 ] . For layer potentials, this distance is proportional to the physical distance ∣ b ∣ |b| ∣ b ∣ from the target to the boundary [3 ] . Consequently, accuracy deteriorates as ∣ b ∣ / h → 0 |b|/h \to0 ∣ b ∣/ h → 0 , where h h h is the mesh size, even when the geometry and data are analytic. To overcome this loss of accuracy, we regularize the integral by replacing the singular kernel with a smooth one that depends on a smoothing radius δ \delta δ . In three dimensions, Beale and collaborators replace 1 / r 1/r 1/ r by erf ( r / δ ) / r \operatorname{erf}(r/\delta)/r erf ( r / δ ) / r [7 , 11 , 14 , 46 , 47 ] and either extrapolate in δ \delta δ [12 ] or modify the smoothing factor to obtain error O ( δ p ) O(\delta^p) O ( δ p ) with p = 3 , 5 , 7 p=3,5,7 p = 3 , 5 , 7 [13 ] . In two dimensions the fundamental solution is logarithmic, and Gaussian regularization adds a term to log r \log r log r rather than multiplying by a cutoff factor. Smoothed logarithmic kernels are standard in vortex methods for area integrals [10 ] and vortex sheet desingularization [2 , 28 ] , and were used for boundary potentials by Beale and Lai [8 ] . Using the gradient of the regularized kernel, which involves the exponential integral E 1 E_1 E 1 , they treated double layers and tangential derivatives of single layers. After subtracting the density at the closest point, their double-layer error was O ( δ 2 ) O(\delta^2) O ( δ 2 ) , and the single-layer error was O ( δ ) O(\delta) O ( δ ) . They eliminated these leading terms with computed corrections requiring derivatives of the density and parametrization, and corrected the discretization error for δ \delta δ proportional to h h h . For the two-dimensional logarithmic kernel, however, existing kernel regularizations have not achieved higher-order accuracy uniformly near the boundary without derivatives of the density. Here we extend the approach of [13 ] by modifying the E 1 E_1 E 1 kernel to achieve uniform regularization error O ( δ p ) O(\delta^p) O ( δ p ) for p = 3 , 5 , 7 p=3,5,7 p = 3 , 5 , 7 . Higher order permits a larger ratio δ / h \delta/h δ / h , which governs the trapezoidal discretization error. If δ / h \delta/h δ / h is fixed as h → 0 h \to0 h → 0 , the discretization error decreases only slowly (O ( h ) O(h) O ( h ) in [8 ] , corrected to O ( h 2 ) O(h^2) O ( h 2 ) ). We therefore set δ = c h q \delta= ch^q δ = c h q with q < 1 q<1 q < 1 , ensuring the discretization error decays faster than the regularization error and yielding convergence O ( h p q ) O(h^{pq}) O ( h pq ) ; for example, p = 7 p=7 p = 7 and q = 5 / 7 q=5/7 q = 5/7 give O ( h 5 ) O(h^5) O ( h 5 ) . With δ \delta δ proportional to h h h , we observe O ( h p ) O(h^p) O ( h p ) over the range where discretization error is negligible.
We illustrate the approach with the single-layer potential
S f ( y ) = ∫ Γ G ( ∣ x − y ∣ ) f ( x ) d s x , G ( r ) = 1 2 π log r , Sf(y)=\int_{\Gamma}G(|x-y|)f(x)\,\mathrm{d}s_x,\qquad G(r)=\frac{1}{2\pi}\log r, S f ( y ) = ∫ Γ G ( ∣ x − y ∣ ) f ( x ) d s x , G ( r ) = 2 π 1 log r ,
on a smooth closed curve Γ \Gamma Γ with given density f f f . The regularized version is
S δ ( 1 ) f ( y ) = ∫ Γ G δ ( 1 ) ( ∣ x − y ∣ ) f ( x ) d s x , G δ ( 1 ) ( r ) = G ( r ) + 1 4 π E 1 ( r 2 / δ 2 ) . S_{\delta}^{(1)}f(y)=\int_{\Gamma}G_{\delta}^{(1)}(|x-y|)f(x)\,\mathrm{d}s_x,\qquad G_{\delta}^{(1)}(r)=G(r)+\frac{1}{4\pi}E_1(r^2/\delta^2). S δ ( 1 ) f ( y ) = ∫ Γ G δ ( 1 ) ( ∣ x − y ∣ ) f ( x ) d s x , G δ ( 1 ) ( r ) = G ( r ) + 4 π 1 E 1 ( r 2 / δ 2 ) .
Here E 1 ( ξ ) = ∫ ξ ∞ v − 1 e − v d v E_1(\xi)=\int_{\xi}^{\infty}v^{-1}e^{-v}\,\mathrm{d}v E 1 ( ξ ) = ∫ ξ ∞ v − 1 e − v d v is the exponential integral. The function G δ ( 1 ) G_{\delta}^{(1)} G δ ( 1 ) is smooth. If y y y is near Γ \Gamma Γ , we can write y = x 0 + b n y=x_0+b n y = x 0 + bn , where x 0 x_0 x 0 is the closest point to y y y on Γ \Gamma Γ , n n n is the outward normal at x 0 x_0 x 0 , and b b b is the signed distance. With λ = b / δ \lambda=b/\delta λ = b / δ , we show in § 4 that
S δ ( 1 ) f ( y ) = S f ( y ) + 1 4 π [ w 0 δ I 0 ( 1 ) ( λ ) + w 2 δ 3 I 2 ( 1 ) ( λ ) + w 4 δ 5 I 4 ( 1 ) ( λ ) ] + O ( δ 7 ) , S_{\delta}^{(1)}f(y)=Sf(y)+\frac{1}{4\pi}\left[w_0\delta I_0^{(1)}(\lambda)+w_2\delta^3 I_2^{(1)}(\lambda)+w_4\delta^5 I_4^{(1)}(\lambda)\right]+O(\delta^7), S δ ( 1 ) f ( y ) = S f ( y ) + 4 π 1 [ w 0 δ I 0 ( 1 ) ( λ ) + w 2 δ 3 I 2 ( 1 ) ( λ ) + w 4 δ 5 I 4 ( 1 ) ( λ ) ] + O ( δ 7 ) ,
uniformly for y y y within a fixed multiple of δ \delta δ of Γ \Gamma Γ . Here the I n ( 1 ) I_n^{(1)} I n ( 1 ) are moments that are known explicitly, and the Taylor coefficients w n w_n w n depend on the curve, the density, and y y y , but not on δ \delta δ .
To make these moments vanish, we add scaled derivatives of E 1 E_1 E 1 , where E 1 [ k ] E_1^{[k]} E 1 [ k ] denotes the k k k -th derivative,
G δ ( p ) ( r ; b ) = G ( r ) + 1 4 π s ( p ) ( r 2 δ 2 ; λ ) , s ( p ) ( ξ ; λ ) = E 1 ( ξ ) + ∑ k = 1 m a k ( p ) ( λ ) ξ k E 1 [ k ] ( ξ ) , (1) G_{\delta}^{(p)}(r;b)=G(r)+\frac{1}{4\pi}s^{(p)}\left(\frac{r^2}{\delta^2};\lambda\right),\qquad s^{(p)}(\xi;\lambda)=E_1(\xi)+\sum_{k=1}^{m}a_k^{(p)}(\lambda)\xi^k E_1^{[k]}(\xi), \tag*{(1)} G δ ( p ) ( r ; b ) = G ( r ) + 4 π 1 s ( p ) ( δ 2 r 2 ; λ ) , s ( p ) ( ξ ; λ ) = E 1 ( ξ ) + k = 1 ∑ m a k ( p ) ( λ ) ξ k E 1 [ k ] ( ξ ) , (1)
with m = ( p − 1 ) / 2 m=(p-1)/2 m = ( p − 1 ) /2 coefficients determined by the conditions that the first m m m moments vanish; the resulting regularization error is O ( δ p ) O(\delta^p) O ( δ p ) . Integration by parts relates the double-layer and Stokes kernels to the single layer, so that they use the same coefficients a k ( p ) ( λ ) a_k^{(p)}(\lambda) a k ( p ) ( λ ) , together with an explicit local correction at x 0 x_0 x 0 for one remaining moment in the Stokes kernels. The pressure is written in terms of Laplace single- and double-layer potentials and uses the same kernels.
We test the method on Laplace and Stokes problems in several geometries and verify the predicted orders of accuracy for target points on and near the boundary. Following Mayo [31 , 32 ] , Beale and Lai [8 ] , and Beale and Tlupova [13 ] , we also use the regularized integrals at grid points near the curve to compute solutions on the entire Cartesian grid with a fast Poisson solver.
The method belongs to a family of approaches that analyze the local error of a smoothed or trapezoidal approximation and remove its leading terms. For evaluation on a curve, corrected trapezoidal rules for logarithmic and other singularities [1 , 26 , 43 , 51 ] and the product quadrature of Kress [29 ] achieve high order. Near the curve, related methods include the local correction of [8 ] , the asymptotic approximations of [15 ] , and the corrected trapezoidal rules for near-singular integrals of [35 ] . In three dimensions, related constructions include the extrapolated regularization of [12 ] , its fast-summation variant [44 ] , the high-order smoothing functions of [13 ] and [19 ] , and corrected trapezoidal rules for near-singular integrals on surfaces [36 ] ; Richardson extrapolation has also been applied to regularized Stokeslets [20 ] . Conditions under which discretization corrections can be neglected are analyzed in [6 ] .
Other high-order approaches treat the local singularity directly. In two dimensions, Helsing and Ojala [24 ] evaluate layer potentials close to the curve by panel-based product integration, an approach extended to drops in Stokes flow in [37 ] , and Barnett and coauthors construct quadratures that remain accurate close to the curve for Laplace, Helmholtz, and Stokes potentials [3 , 4 ] . Helsing’s higher-order singularity subtraction [23 ] expands singular and nearly singular kernels locally and integrates the resulting terms by high-order product integration on parametrized surface patches; quadrature by expansion [27 ] evaluates the potential from an expansion about a point away from the boundary; harmonic density interpolation [39 ] replaces the density locally by a function whose layer potential is known analytically, and [18 ] extends density interpolation to other elliptic operators. Values near a surface can also be interpolated from points farther away [52 ] , and quadrature by fundamental solutions [45 ] provides a kernel-independent alternative for simple objects. The method of regularized Stokeslets [16 , 17 ] also replaces the singular kernel by a smooth one, but it uses a fixed smoothed source and does not cancel moments of the error; its regularization error can be reduced for rigid objects in three dimensions [34 ] .
A practical advantage of the method is its simplicity. Once the regularized kernels are defined, potentials are evaluated directly with the periodic trapezoidal rule on a fixed set of boundary nodes for all targets, provided δ / h \delta/h δ / h is sufficiently large. The regularized kernel differs appreciably from the singular kernel only within a neighborhood of width O ( δ ) O(\delta) O ( δ ) around the target. § 2 presents the regularized Laplace kernels and uniform error estimates, and § 3 treats the Stokes kernels and pressure. § 4 derives the asymptotic expansions, proves the error estimates, and bounds the discretization error. Numerical tests are given in § 5 and § 8 , conclusions in § 9 , and Appendix § 10 gives stable formulas for evaluating the kernels and coefficients without cancellation.
Regularized kernels for the Laplace equation
The single layer and the E 1 E_1 E 1 kernel
Let Γ ⊂ R 2 \Gamma\subset\mathbb{R}^2 Γ ⊂ R 2 be a smooth, simple, closed curve, oriented counterclockwise, with outward unit normal n n n , unit tangent τ \tau τ , arclength s s s , and signed curvature κ \kappa κ , defined by τ s = − κ n \tau_s=-\kappa n τ s = − κn and n s = κ τ n_s=\kappa\tau n s = κ τ , so that κ > 0 \kappa>0 κ > 0 on a convex curve. Write Ω \Omega Ω for the interior domain. For a target y y y in a fixed, sufficiently narrow tubular neighborhood of Γ \Gamma Γ where the closest boundary point is unique, let x 0 = x 0 ( y ) x_0=x_0(y) x 0 = x 0 ( y ) be that closest point and write
y = x 0 + b n 0 , n 0 = n ( x 0 ) , λ = b δ . (2) y = x_0 + b n_0,\qquad n_0 = n(x_0),\qquad\lambda= \frac{b}{\delta}. \tag*{(2)} y = x 0 + b n 0 , n 0 = n ( x 0 ) , λ = δ b . (2)
Here δ \delta δ is the smoothing radius, and λ \lambda λ is the scaled signed distance from Γ \Gamma Γ to the target, with b > 0 b>0 b > 0 outside Ω \Omega Ω .
The source point x x x lies on Γ \Gamma Γ , and y y y is the target point where the potential is evaluated. We set r = x − y = ( r 1 , r 2 ) \mathbf{r}=x-y=(r_1,r_2) r = x − y = ( r 1 , r 2 ) and r = ∣ r ∣ r=\lvert r\rvert r = ∣ r ∣ . We regularize the logarithmic fundamental solution
G ( r ) = log r 2 π . (3) G(r)=\frac{\log r}{2\pi}. \tag*{(3)} G ( r ) = 2 π log r . (3)
The single-layer potential with smooth density f f f on Γ \Gamma Γ is
( S f ) ( y ) = ∫ Γ G ( ∣ x − y ∣ ) f ( x ) d s x . (4) (Sf)(y)=\int_{\Gamma}G(\lvert x-y\rvert)f(x)\,\mathrm{d}s_x. \tag*{(4)} ( S f ) ( y ) = ∫ Γ G (∣ x − y ∣) f ( x ) d s x . (4)
Because the double-layer and Stokes kernels are formed from derivatives of (3), we obtain their regularized versions by differentiating the regularized version of G G G .
Let
φ δ ( r ) = 1 π δ 2 e − r 2 / δ 2 , ∫ R 2 φ δ d A = 1 , (5) \varphi_\delta(r)=\frac{1}{\pi\delta^2}e^{-r^2/\delta^2},\qquad\int_{\mathbb{R}^2}\varphi_\delta\,\mathrm{d}A=1, \tag*{(5)} φ δ ( r ) = π δ 2 1 e − r 2 / δ 2 , ∫ R 2 φ δ d A = 1 , (5)
and set G δ = G ∗ φ δ G_\delta=G*\varphi_\delta G δ = G ∗ φ δ . Since Δ G \Delta G Δ G is the Dirac delta at the origin, we have Δ G δ = φ δ \Delta G_\delta=\varphi_\delta Δ G δ = φ δ ; thus G δ G_\delta G δ is the potential of a radially symmetric Gaussian source of unit mass and width δ \delta δ .
By radial symmetry and the divergence theorem, ∇ G δ \nabla G_\delta ∇ G δ equals ∇ G \nabla G ∇ G times the mass of φ δ \varphi_\delta φ δ inside the disk of radius r r r , namely 2 π ∫ 0 r v φ δ ( v ) d v = 1 − e − r 2 / δ 2 2\pi\int_0^r v\varphi_\delta(v)\,\mathrm{d}v=1-e^{-r^2/\delta^2} 2 π ∫ 0 r v φ δ ( v ) d v = 1 − e − r 2 / δ 2 . Thus, as in [8 eq. (1.6)] ,
∇ G δ = ( 1 − e − r 2 / δ 2 ) ∇ G , ∂ r G δ = 1 2 π r ( 1 − e − r 2 / δ 2 ) . (6) \nabla G_\delta=(1-e^{-r^2/\delta^2})\nabla G,\qquad\partial_rG_\delta=\frac{1}{2\pi r}(1-e^{-r^2/\delta^2}). \tag*{(6)} ∇ G δ = ( 1 − e − r 2 / δ 2 ) ∇ G , ∂ r G δ = 2 π r 1 ( 1 − e − r 2 / δ 2 ) . (6)
Integrating (6) in r r r and choosing the constant of integration so that G δ − G → 0 G_\delta-G\to0 G δ − G → 0 as r → ∞ r\to\infty r → ∞ gives the regularized kernel
G δ ( 1 ) ( r ) = G ( r ) + 1 4 π E 1 ( ξ ) , ξ = ρ 2 , ρ = r δ , E 1 ( ξ ) = ∫ ξ ∞ e − v v d v . (7) G_\delta^{(1)}(r)=G(r)+\frac{1}{4\pi}E_1(\xi),\qquad\xi=\rho^2,\qquad\rho=\frac{r}{\delta},\qquad E_1(\xi)=\int_\xi^\infty\frac{e^{-v}}{v}\,\mathrm{d}v. \tag*{(7)} G δ ( 1 ) ( r ) = G ( r ) + 4 π 1 E 1 ( ξ ) , ξ = ρ 2 , ρ = δ r , E 1 ( ξ ) = ∫ ξ ∞ v e − v d v . (7)
In three dimensions, the analogous calculation gives erf ( ρ ) / r \operatorname{erf}(\rho)/r erf ( ρ ) / r , which differs from 1 / r 1/r 1/ r by a term with Gaussian decay; in two dimensions, it gives log r / ( 2 π ) \log r/(2\pi) log r / ( 2 π ) plus a multiple of E 1 E_1 E 1 . The function E 1 E_1 E 1 is available in standard mathematical libraries; for small r / δ r/\delta r / δ we evaluate G δ ( 1 ) G_\delta^{(1)} G δ ( 1 ) from its power series (Appendix A).
For y y y near Γ \Gamma Γ , we can expand the regularization error in powers of δ \delta δ . In Section 4.1 we show that
S δ ( 1 ) f − S f ∼ 1 4 π [ w 0 δ I 0 ( 1 ) ( λ ) + w 2 δ 3 I 2 ( 1 ) ( λ ) + w 4 δ 5 I 4 ( 1 ) ( λ ) + ⋯ ] , S_\delta^{(1)}f-Sf\sim\frac{1}{4\pi}\left[w_0\delta I_0^{(1)}(\lambda)+w_2\delta^3I_2^{(1)}(\lambda)+w_4\delta^5I_4^{(1)}(\lambda)+\cdots\right], S δ ( 1 ) f − S f ∼ 4 π 1 [ w 0 δ I 0 ( 1 ) ( λ ) + w 2 δ 3 I 2 ( 1 ) ( λ ) + w 4 δ 5 I 4 ( 1 ) ( λ ) + ⋯ ] ,
I n ( 1 ) ( λ ) = ∫ R t n E 1 ( t 2 + λ 2 ) d t , n = 0 , 2 , 4 , … (8) I_n^{(1)}(\lambda)=\int_{\mathbb{R}}t^nE_1(t^2+\lambda^2)\,\mathrm{d}t,\qquad n=0,2,4,\ldots \tag*{(8)} I n ( 1 ) ( λ ) = ∫ R t n E 1 ( t 2 + λ 2 ) d t , n = 0 , 2 , 4 , … (8)
Here η \eta η is a coordinate along Γ \Gamma Γ with ∣ x − y ∣ 2 = η 2 + b 2 \lvert x-y\rvert^2=\eta^2+b^2 ∣ x − y ∣ 2 = η 2 + b 2 , of the same sign as the arclength from x 0 x_0 x 0 , and w n ( x 0 , b ) w_n(x_0,b) w n ( x 0 , b ) are the even Taylor coefficients in η \eta η of the density times d s / d η \mathrm{d}s/\mathrm{d}\eta d s / d η . They depend on the curve, the density, and the target distance b b b , but not on the regularized kernel; the kernel enters only through the moments I n ( 1 ) ( λ ) I_n^{(1)}(\lambda) I n ( 1 ) ( λ ) . The power δ n + 1 \delta^{n+1} δ n + 1 comes from η n = δ n t n \eta^n=\delta^nt^n η n = δ n t n and d η = δ d t \mathrm{d}\eta=\delta\,\mathrm{d}t d η = δ d t . Since E 1 E_1 E 1 is positive and decreasing, the first moment I 0 ( 1 ) ( λ ) I_0^{(1)}(\lambda) I 0 ( 1 ) ( λ ) is positive and decreases with ∣ λ ∣ \lvert\lambda\rvert ∣ λ ∣ . Hence, whenever w 0 ≠ 0 w_0\ne0 w 0 = 0 , the E 1 E_1 E 1 kernel is only first order in δ \delta δ , and this error cannot be reduced by taking more quadrature points.
Higher-order kernels
To get higher order, we modify the E 1 E_1 E 1 term so that the leading moments in (8) are zero. Since successive terms of (8) differ by a factor δ 2 \delta^2 δ 2 , cancelling each moment increases the order by two and requires one free coefficient in the kernel. The kernel (7) is smooth because the logarithmic singularities of G G G and E 1 E_1 E 1 cancel at r = 0 r=0 r = 0 . A term added to E 1 E_1 E 1 must keep this cancellation and decay exponentially, so that the regularized kernel differs from G G G only near the source; a multiple of E 1 E_1 E 1 would spoil the cancellation. The scaled derivatives ξ k E 1 [ k ] \xi^k E_1^{[k]} ξ k E 1 [ k ] , k ≥ 1 k\geq1 k ≥ 1 , where E 1 [ k ] = d k E 1 / d ξ k E_1^{[k]}=\mathrm{d}^kE_1/\mathrm{d}\xi^k E 1 [ k ] = d k E 1 / d ξ k , have both properties: each is a polynomial times e − ξ e^{-\xi} e − ξ (§ 4.2 ). They are easy to differentiate for the double-layer and Stokes kernels, and their moments are known in closed form.
Because the moments depend on λ \lambda λ , the moment conditions must hold for each λ \lambda λ in the band of interest, which cannot be achieved with constant coefficients [8 Sect. 2] . We therefore allow the coefficients to depend on λ \lambda λ and, for order p = 2 m + 1 p=2m+1 p = 2 m + 1 , define
s ( p ) ( ξ ; λ ) = E 1 ( ξ ) + ∑ k = 1 m a k ( p ) ( λ ) ξ k E 1 [ k ] ( ξ ) , (9) s^{(p)}(\xi;\lambda)=E_1(\xi)+\sum_{k=1}^{m}a_k^{(p)}(\lambda)\xi^kE_1^{[k]}(\xi), \tag*{(9)} s ( p ) ( ξ ; λ ) = E 1 ( ξ ) + k = 1 ∑ m a k ( p ) ( λ ) ξ k E 1 [ k ] ( ξ ) , (9)
with a k ( p ) ( λ ) = 0 a_k^{(p)}(\lambda)=0 a k ( p ) ( λ ) = 0 for k > m k>m k > m . The regularized single-layer kernel is
G δ ( p ) ( r ; b ) = G ( r ) + 1 4 π s ( p ) ( ξ ; λ ) , ξ = r 2 δ 2 , (10) G_\delta^{(p)}(r;b)=G(r)+\frac{1}{4\pi}s^{(p)}(\xi;\lambda),\qquad\xi=\frac{r^2}{\delta^2}, \tag*{(10)} G δ ( p ) ( r ; b ) = G ( r ) + 4 π 1 s ( p ) ( ξ ; λ ) , ξ = δ 2 r 2 , (10)
where p = 1 p=1 p = 1 gives the E 1 E_1 E 1 kernel of Section 2.1, and p = 3 , 5 , 7 p=3,5,7 p = 3 , 5 , 7 give third-, fifth-, and seventh-order kernels. We call s ( p ) s^{(p)} s ( p ) the regularizing function. Although s ( p ) s^{(p)} s ( p ) has a logarithmic singularity at ξ = 0 \xi=0 ξ = 0 , its sum with G G G in (10) is smooth. For a fixed target y y y , λ \lambda λ is treated as a constant when we differentiate the kernel with respect to x x x . Note that the coefficients depend only on λ \lambda λ , and not on the density, the curve, or the mesh; this dependence on λ \lambda λ is what distinguishes our kernel from a convolution with a fixed blob. The regularized single-layer potential is
S δ ( p ) f ( y ) = ∫ Γ G δ ( p ) ( ∣ x − y ∣ ; b ) f ( x ) d s x , (11) S_\delta^{(p)}f(y)=\int_\Gamma G_\delta^{(p)}(|x-y|;b)f(x)\,\mathrm{d}s_x, \tag*{(11)} S δ ( p ) f ( y ) = ∫ Γ G δ ( p ) ( ∣ x − y ∣ ; b ) f ( x ) d s x , (11)
with b b b and λ = b / δ \lambda=b/\delta λ = b / δ defined in (2) .
The logarithmic singularity cancels for every p p p . Since E 1 ( ξ ) = − γ − log ξ + ξ − ⋯ E_1(\xi)=-\gamma-\log\xi+\xi-\cdots E 1 ( ξ ) = − γ − log ξ + ξ − ⋯ , where γ \gamma γ is Euler’s constant, we have ξ k E 1 [ k ] ( ξ ) → ( − 1 ) k ( k − 1 ) ! \xi^kE_1^{[k]}(\xi)\to(-1)^k(k-1)! ξ k E 1 [ k ] ( ξ ) → ( − 1 ) k ( k − 1 )! as ξ ↓ 0 \xi\downarrow0 ξ ↓ 0 for k ≥ 1 k\geq1 k ≥ 1 , and the value of the regularized kernel at r = 0 r=0 r = 0 is finite:
G δ ( p ) ( 0 ; 0 ) = 1 4 π [ 2 log δ − γ + ∑ k = 1 m ( − 1 ) k ( k − 1 ) ! a k ( p ) ( 0 ) ] . (12) G_\delta^{(p)}(0;0)=\frac{1}{4\pi}\left[2\log\delta-\gamma+\sum_{k=1}^{m}(-1)^k(k-1)!a_k^{(p)}(0)\right]. \tag*{(12)} G δ ( p ) ( 0 ; 0 ) = 4 π 1 [ 2 log δ − γ + k = 1 ∑ m ( − 1 ) k ( k − 1 )! a k ( p ) ( 0 ) ] . (12)
Since r ≥ ∣ b ∣ r\geq|b| r ≥ ∣ b ∣ , we have r = 0 r=0 r = 0 only when the target lies on the curve (b = λ = 0 b=\lambda=0 b = λ = 0 ) and x = y x=y x = y ; using (12) directly avoids subtracting two large numbers (with a k ( p ) ( 0 ) a_k^{(p)}(0) a k ( p ) ( 0 ) from (16) –(18) , the sum in (12) equals − 2 -2 − 2 , − 8 3 -\frac{8}{3} − 3 8 , and − 46 15 -\frac{46}{15} − 15 46 for p = 3 , 5 , 7 p=3,5,7 p = 3 , 5 , 7 ).
The same expansion for the kernel (10) gives
S δ ( p ) f − S f ∼ 1 4 π ∑ n ≥ 0 n e v e n w n ( x 0 , b ) δ n + 1 I n ( p ) ( λ ) , I n ( p ) ( λ ) = ∫ R t n s ( p ) ( t 2 + λ 2 ; λ ) d t . (13) S_\delta^{(p)}f-Sf\sim\frac{1}{4\pi}\sum_{\substack{n\geq0\\ n\ \mathrm{even}}}w_n(x_0,b)\delta^{n+1}I_n^{(p)}(\lambda),\qquad I_n^{(p)}(\lambda)=\int_{\mathbb{R}}t^n s^{(p)}(t^2+\lambda^2;\lambda)\,\mathrm{d}t. \tag*{(13)} S δ ( p ) f − S f ∼ 4 π 1 n ≥ 0 n even ∑ w n ( x 0 , b ) δ n + 1 I n ( p ) ( λ ) , I n ( p ) ( λ ) = ∫ R t n s ( p ) ( t 2 + λ 2 ; λ ) d t . (13)
Here only the moments change, while the Taylor coefficients w n w_n w n are the same as in (8) . For the first m m m terms of (13) to vanish for arbitrary curves and densities, the first m m m even moments must be zero:
I n ( p ) ( λ ) = 0 , n = 0 , 2 , … , p − 3. (14) I_n^{(p)}(\lambda)=0,\qquad n=0,2,\ldots,p-3. \tag*{(14)} I n ( p ) ( λ ) = 0 , n = 0 , 2 , … , p − 3. (14)
This gives m m m linear equations for the m m m coefficients a k ( p ) ( λ ) a_k^{(p)}(\lambda) a k ( p ) ( λ ) in (9) . When (14) holds, the leading nonzero term in (13) is n = p − 1 n=p-1 n = p − 1 , which is O ( δ p ) O(\delta^p) O ( δ p ) because I p − 1 ( p ) I_{p-1}^{(p)} I p − 1 ( p ) and w p − 1 w_{p-1} w p − 1 are uniformly bounded for ∣ b ∣ ≤ Λ δ |b|\leq\Lambda\delta ∣ b ∣ ≤ Λ δ (with Λ \Lambda Λ fixed) when the curve and density are smooth. Thus the kernel (10) has O ( δ p ) O(\delta^p) O ( δ p ) regularization error.
For p = 3 , 5 , 7 p=3,5,7 p = 3 , 5 , 7 , the linear system (14) can be solved explicitly (§ 4.2 ) in terms of the integrals
H n ( λ ) = ∫ 1 ∞ e − λ 2 ( v − 1 ) v − ( n + 3 ) / 2 d v , H n ( 0 ) = 2 n + 1 , n = 0 , 2 , 4 , … , (15) \mathcal{H}_n(\lambda)=\int_1^\infty e^{-\lambda^2(v-1)}v^{-(n+3)/2}\,\mathrm{d}v,\qquad\mathcal{H}_n(0)=\frac{2}{n+1},\qquad n=0,2,4,\ldots, \tag*{(15)} H n ( λ ) = ∫ 1 ∞ e − λ 2 ( v − 1 ) v − ( n + 3 ) /2 d v , H n ( 0 ) = n + 1 2 , n = 0 , 2 , 4 , … , (15)
which are positive, bounded, even functions of λ \lambda λ that decrease with ∣ λ ∣ |\lambda| ∣ λ ∣ . Appendix § 10 .§ 2 gives H 0 \mathcal{H}_0 H 0 in closed form and a recurrence for the other H n \mathcal{H}_n H n , and describes how to evaluate them and the coefficients accurately for all λ \lambda λ . In terms of H n \mathcal{H}_n H n , we have I n ( 1 ) ( λ ) = Γ ( ( n + 1 ) / 2 ) e − λ 2 H n ( λ ) I_n^{(1)}(\lambda)=\Gamma((n+1)/2)e^{-\lambda^2}\mathcal{H}_n(\lambda) I n ( 1 ) ( λ ) = Γ (( n + 1 ) /2 ) e − λ 2 H n ( λ ) for even n ≥ 0 n\geq0 n ≥ 0 , and the coefficients are
p = 3 : a 1 ( 3 ) = H 0 , (16) \begin{aligned}
p=3:\quad&a_1^{(3)}=\mathcal{H}_0, \tag*{(16)}
\end{aligned} p = 3 : a 1 ( 3 ) = H 0 , (16)
p = 5 : a 2 ( 5 ) = H 0 − H 2 , a 1 ( 5 ) = H 0 + ( λ 2 + 3 2 ) a 2 ( 5 ) , (17) \begin{aligned}
p=5:\quad&a_2^{(5)}=\mathcal{H}_0-\mathcal{H}_2,\qquad a_1^{(5)}=\mathcal{H}_0+\left(\lambda^2+\frac{3}{2}\right)a_2^{(5)}, \tag*{(17)}
\end{aligned} p = 5 : a 2 ( 5 ) = H 0 − H 2 , a 1 ( 5 ) = H 0 + ( λ 2 + 2 3 ) a 2 ( 5 ) , (17)
p = 7 : a 3 ( 7 ) = 1 2 ( H 0 − 2 H 2 + H 4 ) , a 2 ( 7 ) = H 0 − H 2 + ( 2 λ 2 + 5 ) a 3 ( 7 ) , a 1 ( 7 ) = H 0 + ( λ 2 + 3 2 ) a 2 ( 7 ) − ( λ 4 + 3 λ 2 + 15 4 ) a 3 ( 7 ) . (18) \begin{aligned}
p=7:\quad&a_3^{(7)}=\frac{1}{2}(\mathcal{H}_0-2\mathcal{H}_2+\mathcal{H}_4),\\
&a_2^{(7)}=\mathcal{H}_0-\mathcal{H}_2+(2\lambda^2+5)a_3^{(7)},\\
&a_1^{(7)}=\mathcal{H}_0+\left(\lambda^2+\frac{3}{2}\right)a_2^{(7)}-\left(\lambda^4+3\lambda^2+\frac{15}{4}\right)a_3^{(7)}. \tag*{(18)}
\end{aligned} p = 7 : a 3 ( 7 ) = 2 1 ( H 0 − 2 H 2 + H 4 ) , a 2 ( 7 ) = H 0 − H 2 + ( 2 λ 2 + 5 ) a 3 ( 7 ) , a 1 ( 7 ) = H 0 + ( λ 2 + 2 3 ) a 2 ( 7 ) − ( λ 4 + 3 λ 2 + 4 15 ) a 3 ( 7 ) . (18)
On the curve (λ = 0 \lambda=0 λ = 0 ), these take the simple values a 1 ( 3 ) = 2 a_1^{(3)}=2 a 1 ( 3 ) = 2 , ( a 1 ( 5 ) , a 2 ( 5 ) ) = ( 4 , 4 3 ) (a_1^{(5)},a_2^{(5)})=(4,\frac{4}{3}) ( a 1 ( 5 ) , a 2 ( 5 ) ) = ( 4 , 3 4 ) , and ( a 1 ( 7 ) , a 2 ( 7 ) , a 3 ( 7 ) ) = ( 6 , 4 , 8 15 ) (a_1^{(7)},a_2^{(7)},a_3^{(7)})=(6,4,\frac{8}{15}) ( a 1 ( 7 ) , a 2 ( 7 ) , a 3 ( 7 ) ) = ( 6 , 4 , 15 8 ) . Just as in [13 ] , the formulas are in triangular form: setting a 3 ( 7 ) = 0 a_3^{(7)}=0 a 3 ( 7 ) = 0 in the p = 7 p=7 p = 7 formulas gives the coefficients for p = 5 p=5 p = 5 , and setting also a 2 ( 5 ) = 0 a_2^{(5)}=0 a 2 ( 5 ) = 0 gives a 1 ( 3 ) a_1^{(3)} a 1 ( 3 ) .
With p = 3 p=3 p = 3 , the kernel cancels the leading error term that Beale and Lai compute and subtract [8 ] . Their correction requires derivatives of the density; the modified kernel does not, so it is simpler to use, and p = 5 p=5 p = 5 or 7 7 7 cancels further terms in the same way.
Radial factors
To obtain the double-layer and Stokes kernels, we write the radial derivative of the regularized kernel as ∂ r G \partial_rG ∂ r G multiplied by a radial factor:
∂ r [ G ( r ) + 1 4 π s ( p ) ( ξ ; λ ) ] = 1 2 π r ( 1 + ξ ∂ ξ s ( p ) ( ξ ; λ ) ) , ξ = r 2 δ 2 , (19) \partial_r\left[G(r)+\frac{1}{4\pi}s^{(p)}(\xi;\lambda)\right]=\frac{1}{2\pi r}\left(1+\xi\partial_\xi s^{(p)}(\xi;\lambda)\right),\qquad\xi=\frac{r^2}{\delta^2}, \tag*{(19)} ∂ r [ G ( r ) + 4 π 1 s ( p ) ( ξ ; λ ) ] = 2 π r 1 ( 1 + ξ ∂ ξ s ( p ) ( ξ ; λ ) ) , ξ = δ 2 r 2 , (19)
so the factor is $1-Q^{(p)}$ with Q ( p ) = − ξ ∂ ξ s ( p ) Q^{(p)}=-\xi\partial_\xi s^{(p)} Q ( p ) = − ξ ∂ ξ s ( p ) . For p = 1 p=1 p = 1 , we have Q ( 1 ) = e − ξ Q^{(1)}=e^{-\xi} Q ( 1 ) = e − ξ by (6) ; adding the scaled derivatives of E 1 E_1 E 1 multiplies this Gaussian by a polynomial of degree m m m . Specifically, − ξ ∂ ξ E 1 = e − ξ -\xi\partial_\xi E_1=e^{-\xi} − ξ ∂ ξ E 1 = e − ξ and − ξ ∂ ξ [ ξ k E 1 [ k ] ] = ( − 1 ) k k ξ k e − ξ -\xi\partial_\xi[\xi^kE_1^{[k]}]=(-1)^k k\xi^k e^{-\xi} − ξ ∂ ξ [ ξ k E 1 [ k ] ] = ( − 1 ) k k ξ k e − ξ for $k\geq 1 ({xref}Section 4.2). Differentiating {xref}(9)` term by term gives
Q ( p ) ( ξ ; λ ) : = − ξ ∂ ξ s ( p ) ( ξ ; λ ) = e − ξ P ( p ) ( ξ ; λ ) , Q^{(p)}(\xi;\lambda):=-\xi\partial_\xi s^{(p)}(\xi;\lambda)=e^{-\xi}P^{(p)}(\xi;\lambda), Q ( p ) ( ξ ; λ ) := − ξ ∂ ξ s ( p ) ( ξ ; λ ) = e − ξ P ( p ) ( ξ ; λ ) ,
P ( p ) ( ξ ; λ ) : = 1 + ∑ k = 1 m ( − 1 ) k a k ( p ) ( λ ) ξ k . (20) P^{(p)}(\xi;\lambda):=1+\sum_{k=1}^{m}(-1)^k a_k^{(p)}(\lambda)\xi^k. \tag*{(20)} P ( p ) ( ξ ; λ ) := 1 + k = 1 ∑ m ( − 1 ) k a k ( p ) ( λ ) ξ k . (20)
We then define the radial factors
d ( p ) ( ξ ; λ ) : = 1 − Q ( p ) ( ξ ; λ ) = 1 + ξ ∂ ξ s ( p ) ( ξ ; λ ) , d^{(p)}(\xi;\lambda):=1-Q^{(p)}(\xi;\lambda)=1+\xi\partial_\xi s^{(p)}(\xi;\lambda), d ( p ) ( ξ ; λ ) := 1 − Q ( p ) ( ξ ; λ ) = 1 + ξ ∂ ξ s ( p ) ( ξ ; λ ) ,
d ~ ( p ) ( ξ ; λ ) : = d ( p ) ( ξ ; λ ) − ξ ∂ ξ d ( p ) ( ξ ; λ ) = 1 − ξ 2 ∂ ξ 2 s ( p ) ( ξ ; λ ) . (21) \widetilde d^{(p)}(\xi;\lambda):=d^{(p)}(\xi;\lambda)-\xi\partial_\xi d^{(p)}(\xi;\lambda)=1-\xi^2\partial_\xi^2s^{(p)}(\xi;\lambda). \tag*{(21)} d ( p ) ( ξ ; λ ) := d ( p ) ( ξ ; λ ) − ξ ∂ ξ d ( p ) ( ξ ; λ ) = 1 − ξ 2 ∂ ξ 2 s ( p ) ( ξ ; λ ) . (21)
Thus the single-layer kernel is regularized by adding s ( p ) / ( 4 π ) s^{(p)}/(4\pi) s ( p ) / ( 4 π ) , while the double-layer kernel is ∇ G \nabla G ∇ G times d ( p ) d^{(p)} d ( p ) . The second factor d ~ ( p ) \tilde{d}^{(p)} d ~ ( p ) is used only in the stresslet kernel. This construction is the two-dimensional counterpart of [13 eq. (13)] , where the three-dimensional smoothing factor erf ρ \operatorname{erf}\rho erf ρ is modified by adding multiples of ρ ( erf ρ ) ′ \rho(\operatorname{erf}\rho)' ρ ( erf ρ ) ′ , ρ 2 ( erf ρ ) ′ ′ \rho^2(\operatorname{erf}\rho)'' ρ 2 ( erf ρ ) ′′ , and higher scaled derivatives.
Since the coefficients (16)–(18) are polynomials in λ 2 \lambda^2 λ 2 and the bounded functions H n H_n H n , they remain bounded for ∣ λ ∣ ≤ Λ |\lambda| \le\Lambda ∣ λ ∣ ≤ Λ for any fixed Λ > 0 \Lambda> 0 Λ > 0 . Consequently, Q ( p ) = e − ξ P ( p ) Q^{(p)} = e^{-\xi}P^{(p)} Q ( p ) = e − ξ P ( p ) decays like a Gaussian uniformly for ∣ λ ∣ ≤ Λ |\lambda| \le\Lambda ∣ λ ∣ ≤ Λ , and G δ ( p ) − G G_\delta^{(p)} - G G δ ( p ) − G is exponentially small when ρ ≫ 1 \rho\gg1 ρ ≫ 1 . The local expansion (13) therefore holds for all the regularized kernels G δ ( p ) G_\delta^{(p)} G δ ( p ) .
The double-layer potential
The double-layer kernel is the normal derivative of G G G with respect to the source point x x x :
K ( x , y ) = ∂ n ( x ) G ( r ) = n ( x ) ⋅ r 2 π r 2 , r = x − y , (22) K(x,y) = \partial_{n(x)}G(r) = \frac{n(x)\cdot r}{2\pi r^2}, \qquad r = x-y, \tag*{(22)} K ( x , y ) = ∂ n ( x ) G ( r ) = 2 π r 2 n ( x ) ⋅ r , r = x − y , (22)
and the double-layer potential with smooth density g g g is
( D g ) ( y ) = ∫ Γ K ( x , y ) g ( x ) d s x . (23) (Dg)(y) = \int_\Gamma K(x,y)g(x)\,\mathrm{d}s_x. \tag*{(23)} ( D g ) ( y ) = ∫ Γ K ( x , y ) g ( x ) d s x . (23)
When the target y y y lies on Γ \Gamma Γ , D g Dg D g stands for the direct value K g Kg K g of the integral, which equals the average of the interior and exterior limits D − g D^-g D − g and D + g D^+g D + g ; we use the same convention for the other potentials that jump across Γ \Gamma Γ . The one-sided limits are given by the jump relations [29 ]
D − g = K g + 1 2 g , D + g = K g − 1 2 g . (24) D^-g = Kg + \frac{1}{2}g,\qquad D^+g = Kg - \frac{1}{2}g. \tag*{(24)} D − g = K g + 2 1 g , D + g = K g − 2 1 g . (24)
Our singularity subtraction uses Gauss’s identity
∫ Γ K ( x , y ) d s x = χ ( y ) , χ ( y ) = { 1 , y ∈ Ω , 0 , y ∉ Ω ‾ , 1 2 , y ∈ Γ . (25) \int_\Gamma K(x,y)\,\mathrm{d}s_x = \chi(y), \qquad\chi(y) =
\begin{cases}
1, & y \in\Omega,\\
0, & y \notin\overline{\Omega},\\
\frac{1}{2}, & y \in\Gamma.
\end{cases} \tag*{(25)} ∫ Γ K ( x , y ) d s x = χ ( y ) , χ ( y ) = ⎩ ⎨ ⎧ 1 , 0 , 2 1 , y ∈ Ω , y ∈ / Ω , y ∈ Γ. (25)
Differentiating (10) with respect to x x x while keeping y y y , b b b , δ \delta δ , and thus λ \lambda λ and a k ( p ) ( λ ) a_k^{(p)}(\lambda) a k ( p ) ( λ ) fixed gives the regularized double-layer kernel
K δ ( p ) ( x , y ; b ) = ∂ n ( x ) G δ ( p ) = n ( x ) ⋅ ( x − y ) 2 π ∣ x − y ∣ 2 d ( p ) ( ∣ x − y ∣ 2 δ 2 ; λ ) . (26) K_\delta^{(p)}(x,y;b) = \partial_{n(x)}G_\delta^{(p)} = \frac{n(x)\cdot(x-y)}{2\pi|x-y|^2}d^{(p)}\left(\frac{|x-y|^2}{\delta^2};\lambda\right). \tag*{(26)} K δ ( p ) ( x , y ; b ) = ∂ n ( x ) G δ ( p ) = 2 π ∣ x − y ∣ 2 n ( x ) ⋅ ( x − y ) d ( p ) ( δ 2 ∣ x − y ∣ 2 ; λ ) . (26)
Because d ( p ) ( 0 ; λ ) = 0 d^{(p)}(0;\lambda)=0 d ( p ) ( 0 ; λ ) = 0 and d ( p ) ( ξ ; λ ) / ξ d^{(p)}(\xi;\lambda)/\xi d ( p ) ( ξ ; λ ) / ξ is analytic at ξ = 0 \xi=0 ξ = 0 with value 1 + a 1 ( p ) 1+a_1^{(p)} 1 + a 1 ( p ) , the kernel (26) is a smooth function of x x x and equals zero when x = y x=y x = y . We define the regularized double-layer potential with subtraction by
D δ ( p ) g ( y ) = ∫ Γ K δ ( p ) ( x , y ; b ) [ g ( x ) − g ( x 0 ) ] d s x + χ ( y ) g ( x 0 ) . (27) D_\delta^{(p)}g(y) = \int_\Gamma K_\delta^{(p)}(x,y;b)\left[g(x)-g(x_0)\right]\,\mathrm{d}s_x + \chi(y)g(x_0). \tag*{(27)} D δ ( p ) g ( y ) = ∫ Γ K δ ( p ) ( x , y ; b ) [ g ( x ) − g ( x 0 ) ] d s x + χ ( y ) g ( x 0 ) . (27)
For y ∈ Γ y \in\Gamma y ∈ Γ , (27) with χ ( y ) = 1 2 \chi(y)=\frac{1}{2} χ ( y ) = 2 1 approximates the direct value K g Kg K g , while setting χ = 1 \chi=1 χ = 1 or 0 0 0 gives the interior or exterior limit in (24) . Because of (25) , (27) is exact when g g g is constant: the integral is zero and χ g \chi g χ g reproduces the exact potential. Subtracting g ( x 0 ) g(x_0) g ( x 0 ) also makes the integrand vanish at the closest point x 0 x_0 x 0 , removing the constant term in the even part of n ( x ) ⋅ ( x − y ) [ g ( x ) − g ( x 0 ) ] n(x)\cdot(x-y)[g(x)-g(x_0)] n ( x ) ⋅ ( x − y ) [ g ( x ) − g ( x 0 )] . The same subtraction is used in three dimensions in [13 Sect. 2] and [47 ] .
The error D δ ( p ) g − D g D_{\delta}^{(p)}g-Dg D δ ( p ) g − D g has a local expansion of the same form as (13) , with the moments
J n ( p ) ( λ ) = ∫ R t n + 2 t 2 + λ 2 Q ( p ) ( t 2 + λ 2 ; λ ) d t , n = 0 , 2 , 4 , … , (28) J_{n}^{(p)}(\lambda)=\int_{\mathbb{R}}\frac{t^{n+2}}{t^{2}+\lambda^{2}}Q^{(p)}(t^{2}+\lambda^{2};\lambda)\,\mathrm{d}t,\qquad n=0,2,4,\ldots, \tag*{(28)} J n ( p ) ( λ ) = ∫ R t 2 + λ 2 t n + 2 Q ( p ) ( t 2 + λ 2 ; λ ) d t , n = 0 , 2 , 4 , … , (28)
in place of I n ( p ) I_{n}^{(p)} I n ( p ) . Section 4.3 derives this expansion and shows that J n ( p ) = n + 1 2 I n ( p ) J_{n}^{(p)}=\frac{n+1}{2}I_{n}^{(p)} J n ( p ) = 2 n + 1 I n ( p ) . Therefore, the same coefficients a k ( p ) ( λ ) a_{k}^{(p)}(\lambda) a k ( p ) ( λ ) give O ( δ p ) O(\delta^{p}) O ( δ p ) regularization error for the double layer as well. Without subtracting g ( x 0 ) g(x_{0}) g ( x 0 ) , the error would not go to zero uniformly as y y y approaches the curve, because K δ ( p ) K_{\delta}^{(p)} K δ ( p ) does not satisfy Gauss’s identity (25) .
Error estimate
For p = 3 , 5 , 7 p=3,5,7 p = 3 , 5 , 7 , using the coefficients (16) –(18) , the regularized potentials have error O ( δ p ) O(\delta^{p}) O ( δ p ) uniformly for targets within a fixed multiple of δ \delta δ of Γ \Gamma Γ . The proof is given in Section 4.6.
Proposition 2.1 (Uniform accuracy near the curve). Let Γ \Gamma Γ and the densities f f f and g g g be smooth, and let Λ > 0 \Lambda>0 Λ > 0 and δ 0 > 0 \delta_{0}>0 δ 0 > 0 satisfy Λ δ 0 ∥ κ ∥ ∞ ≤ 1 2 \Lambda\delta_{0}\lVert\kappa\rVert_{\infty}\le\frac{1}{2} Λ δ 0 ∥ κ ∥ ∞ ≤ 2 1 , with Λ δ 0 \Lambda\delta_{0} Λ δ 0 small enough that every point within distance Λ δ 0 \Lambda\delta_{0} Λ δ 0 of Γ \Gamma Γ has a unique closest point on Γ \Gamma Γ . Let p = 3 , 5 p=3,5 p = 3 , 5 , or 7 7 7 , and let S δ ( p ) f S_{\delta}^{(p)}f S δ ( p ) f and D δ ( p ) g D_{\delta}^{(p)}g D δ ( p ) g be given by (11) and (27) , with the kernels (10) and (26) and the coefficients (16) –(18) at λ = b / δ \lambda=b/\delta λ = b / δ . Then for 0 < δ ≤ δ 0 0<\delta\le\delta_{0} 0 < δ ≤ δ 0 ,
sup x 0 ∈ Γ , ∣ b ∣ ≤ Λ δ ( ∣ S δ ( p ) f − S f ∣ + ∣ D δ ( p ) g − D g ∣ ) ≤ C δ p . (29) \sup_{\substack{x_{0}\in\Gamma,\,|b|\le\Lambda\delta}}\left(\left|S_{\delta}^{(p)}f-Sf\right|+\left|D_{\delta}^{(p)}g-Dg\right|\right)\le C\delta^{p}. \tag*{(29)} x 0 ∈ Γ , ∣ b ∣ ≤ Λ δ sup ( S δ ( p ) f − S f + D δ ( p ) g − D g ) ≤ C δ p . (29) The constant C C C depends on finitely many derivatives of Γ \Gamma Γ and the densities, and on p p p and Λ \Lambda Λ , but not on δ \delta δ , b b b , or x 0 x_{0} x 0 .
For targets farther from the curve (∣ λ ∣ |\lambda| ∣ λ ∣ large), the modification to the kernel is exponentially small in ∣ b ∣ / δ |b|/\delta ∣ b ∣/ δ and can be neglected. Where the curvature is large, the condition Λ δ 0 ∥ κ ∥ ∞ ≤ 1 2 \Lambda\delta_{0}\lVert\kappa\rVert_{\infty}\le\frac{1}{2} Λ δ 0 ∥ κ ∥ ∞ ≤ 2 1 requires δ \delta δ (and thus h h h ) to be small so that Λ \Lambda Λ can be chosen large enough; a small h h h is needed in any case to resolve a region of large curvature.
Remark 2.2. Unlike the single-layer kernel, which has a logarithmic singularity at x = y x=y x = y , the double-layer kernel K ( x , y ) K(x,y) K ( x , y ) of (22) for y ∈ Γ y\in\Gamma y ∈ Γ is a smooth function of x ∈ Γ x\in\Gamma x ∈ Γ , with limit κ ( y ) / ( 4 π ) \kappa(y)/(4\pi) κ ( y ) / ( 4 π ) as x → y x\to y x → y [29 ] . With this diagonal value, the periodic trapezoidal rule for D g ( y ) Dg(y) D g ( y ) converges faster than any power of h h h [29 , 48 ] . Hence no regularization is needed on the curve, though (27) with λ = 0 \lambda=0 λ = 0 remains valid there. This differs from three dimensions, where the double-layer kernel is O ( 1 / r ) O(1/r) O ( 1/ r ) on the surface and is regularized there as well [5 , 47 ] .
Regularized kernels for Stokes flow
We now extend the regularization to two-dimensional Stokes flow, regularizing the Stokeslet and stresslet integrals and expressing the pressure in terms of harmonic layer potentials, using the regularized kernels and coefficients of Section 2.
The Stokeslet and stresslet integrals
For creeping flow with viscosity μ > 0 \mu>0 μ > 0 , velocity u u u and pressure p p p satisfy μ Δ u = ∇ p \mu\Delta u=\nabla p μ Δ u = ∇ p and ∇ ⋅ u = 0 \nabla\cdot u=0 ∇ ⋅ u = 0 , with hydrodynamic stress Σ = − p I + μ ( ∇ u + ∇ u T ) \Sigma=-pI+\mu(\nabla u+\nabla u^{T}) Σ = − p I + μ ( ∇ u + ∇ u T ) [41 ] . From here on, p p p has two meanings, distinguished by use: the pressure is a field, as in ∇ p \nabla p ∇ p , p ± p^{\pm} p ± , and p δ p_{\delta} p δ , and the order is a number, as in the superscript ( p ) (p) ( p ) , δ p \delta^{p} δ p , and p = 3 , 5 , 7 p=3,5,7 p = 3 , 5 , 7 . The single-layer (Stokeslet) kernel S i j S_{ij} S ij and double-layer (stresslet) kernel T i j k T_{ijk} T ij k are
S i j ( r ) = 1 4 π μ ( − δ i j log r + r i r j r 2 ) , T i j k ( r ) = 1 π r i r j r k r 4 , r = x − y , (30) S_{ij}(\boldsymbol{r})=\frac{1}{4\pi\mu}\left(-\delta_{ij}\log r+\frac{r_i r_j}{r^2}\right), \qquad T_{ijk}(\boldsymbol{r})=\frac{1}{\pi}\frac{r_i r_j r_k}{r^4}, \qquad r=x-y, \tag*{(30)} S ij ( r ) = 4 π μ 1 ( − δ ij log r + r 2 r i r j ) , T ij k ( r ) = π 1 r 4 r i r j r k , r = x − y , (30)
with
( S f ) i ( y ) = ∫ Γ S i j ( x − y ) f j ( x ) d s x , ( D q ) i ( y ) = ∫ Γ T i j k ( x − y ) q j ( x ) n k ( x ) d s x , (31) (\mathcal{S}f)_i(y)=\int_{\Gamma}S_{ij}(x-y)f_j(x)\,\mathrm{d}s_x,\qquad(\mathcal{D}q)_i(y)=\int_{\Gamma}T_{ijk}(x-y)q_j(x)n_k(x)\,\mathrm{d}s_x, \tag*{(31)} ( S f ) i ( y ) = ∫ Γ S ij ( x − y ) f j ( x ) d s x , ( D q ) i ( y ) = ∫ Γ T ij k ( x − y ) q j ( x ) n k ( x ) d s x , (31)
where f f f and q q q are prescribed smooth vector densities and repeated indices are summed. We regularize both kernels by expressing them through G G G and its radial derivatives and replacing these by their regularized versions from § 2 . This gives
S i j , δ ( p ) = − δ i j 2 μ G δ ( p ) + 1 4 π μ r i r j δ 2 A ( p ) ( ξ ; λ ) , A ( p ) ( ξ ; λ ) = d ( p ) ( ξ , λ ) ξ , (32) S_{ij,\delta}^{(p)}=-\frac{\delta_{ij}}{2\mu}G_{\delta}^{(p)}+\frac{1}{4\pi\mu}\frac{r_i r_j}{\delta^2}A^{(p)}(\xi;\lambda),\qquad A^{(p)}(\xi;\lambda)=\frac{\mathrm{d}^{(p)}(\xi,\lambda)}{\xi}, \tag*{(32)} S ij , δ ( p ) = − 2 μ δ ij G δ ( p ) + 4 π μ 1 δ 2 r i r j A ( p ) ( ξ ; λ ) , A ( p ) ( ξ ; λ ) = ξ d ( p ) ( ξ , λ ) , (32)
T i j k , δ ( p ) = 1 π r i r j r k δ 4 B ( p ) ( ξ ; λ ) , B ( p ) ( ξ ; λ ) = d ~ ( p ) ( ξ ; λ ) ξ 2 = − ∂ ξ A ( p ) ( ξ ; λ ) , (33) T_{ijk,\delta}^{(p)}=\frac{1}{\pi}\frac{r_i r_j r_k}{\delta^4}B^{(p)}(\xi;\lambda),\qquad B^{(p)}(\xi;\lambda)=\frac{\widetilde{d}^{(p)}(\xi;\lambda)}{\xi^2}=-\partial_{\xi}A^{(p)}(\xi;\lambda), \tag*{(33)} T ij k , δ ( p ) = π 1 δ 4 r i r j r k B ( p ) ( ξ ; λ ) , B ( p ) ( ξ ; λ ) = ξ 2 d ( p ) ( ξ ; λ ) = − ∂ ξ A ( p ) ( ξ ; λ ) , (33)
where all derivatives in ξ \xi ξ are taken at fixed λ \lambda λ . Both kernels have finite limits as r → 0 r\to0 r → 0 . Since A ( p ) → 1 + a 1 ( p ) A^{(p)}\to1+a_1^{(p)} A ( p ) → 1 + a 1 ( p ) and B ( p ) → 1 2 + a 1 ( p ) + a 2 ( p ) B^{(p)}\to\frac{1}{2}+a_1^{(p)}+a_2^{(p)} B ( p ) → 2 1 + a 1 ( p ) + a 2 ( p ) as ξ ↓ 0 \xi\downarrow0 ξ ↓ 0 , we have T i j k , δ ( p ) = 0 T_{ijk,\delta}^{(p)}=0 T ij k , δ ( p ) = 0 and S i j , δ ( p ) = − δ i j G δ ( p ) ( 0 ; 0 ) / ( 2 μ ) S_{ij,\delta}^{(p)}=-\delta_{ij}G_{\delta}^{(p)}(0;0)/(2\mu) S ij , δ ( p ) = − δ ij G δ ( p ) ( 0 ; 0 ) / ( 2 μ ) at r = 0 r=0 r = 0 , with G δ ( p ) ( 0 ; 0 ) G_{\delta}^{(p)}(0;0) G δ ( p ) ( 0 ; 0 ) as in (12) . We evaluate the kernels through A ( p ) A^{(p)} A ( p ) and B ( p ) B^{(p)} B ( p ) , so that r − 2 r^{-2} r − 2 and r − 4 r^{-4} r − 4 are never formed; see Appendix § 10 .
The Stokeslet has an additional moment that the conditions (14) do not cancel; it is removed by a local correction (§ 4.4 ). The corrected Stokes single layer is
( S δ , c o r r ( p ) f ) i = ∫ Γ S i j , δ ( p ) f j d s + ∣ b ∣ N ( p ) ( λ ) 4 π μ α ( n 0 ) i ( n 0 ⋅ f 0 ) , (34) (\mathcal{S}_{\delta,\mathrm{corr}}^{(p)}f)_i=\int_{\Gamma}S_{ij,\delta}^{(p)}f_j\,\mathrm{d}s+\frac{|b|\mathcal{N}^{(p)}(\lambda)}{4\pi\mu\alpha}(n_0)_i(n_0\cdot f_0), \tag*{(34)} ( S δ , corr ( p ) f ) i = ∫ Γ S ij , δ ( p ) f j d s + 4 π μα ∣ b ∣ N ( p ) ( λ ) ( n 0 ) i ( n 0 ⋅ f 0 ) , (34)
where x 0 x_0 x 0 , n 0 n_0 n 0 , and b b b are as in (2) , f 0 = f ( x 0 ) f_0=f(x_0) f 0 = f ( x 0 ) , α = 1 + b κ 0 \alpha=\sqrt{1+b\kappa_0} α = 1 + b κ 0 , κ 0 \kappa_0 κ 0 is the curvature at x 0 x_0 x 0 , and
N ( p ) ( λ ) = π erfc ∣ λ ∣ + π ∣ λ ∣ e − λ 2 [ − a 1 ( p ) + a 2 ( p ) ( λ 2 + 1 2 ) − a 3 ( p ) ( λ 4 + λ 2 + 3 4 ) ] , (35) \mathcal{N}^{(p)}(\lambda)=\pi\operatorname{erfc}|\lambda|+\sqrt{\pi}|\lambda|e^{-\lambda^2}\left[-a_1^{(p)}+a_2^{(p)}\left(\lambda^2+\frac{1}{2}\right)-a_3^{(p)}\left(\lambda^4+\lambda^2+\frac{3}{4}\right)\right], \tag*{(35)} N ( p ) ( λ ) = π erfc ∣ λ ∣ + π ∣ λ ∣ e − λ 2 [ − a 1 ( p ) + a 2 ( p ) ( λ 2 + 2 1 ) − a 3 ( p ) ( λ 4 + λ 2 + 4 3 ) ] , (35)
with the coefficients a k ( p ) a_k^{(p)} a k ( p ) of (16) –(18) and a k ( p ) = 0 a_k^{(p)}=0 a k ( p ) = 0 for k > m k>m k > m . The function N ( p ) \mathcal{N}^{(p)} N ( p ) is bounded and equals π \pi π at λ = 0 \lambda=0 λ = 0 . In three dimensions [13 eq. (41)] , the same moment is treated differently: ( n 0 ⋅ f 0 ) n (n_0\cdot f_0)n ( n 0 ⋅ f 0 ) n is subtracted from the density, which does not change the Stokeslet integral since ∫ Γ S i j n j d s = 0 \int_{\Gamma}S_{ij}n_j\,\mathrm{d}s=0 ∫ Γ S ij n j d s = 0 . Here the density is not modified; because N ( p ) \mathcal{N}^{(p)} N ( p ) has the closed form (35) , the error term is computed and added directly.
With the sign convention of (30) , the analogue of Gauss’s identity (25) is
∫ Γ T i j k ( x − y ) n k ( x ) d s x = χ ( y ) δ i j . \int_{\Gamma}T_{ijk}(x-y)n_k(x)\,\mathrm{d}s_x=\chi(y)\delta_{ij}. ∫ Γ T ij k ( x − y ) n k ( x ) d s x = χ ( y ) δ ij .
We use this identity for singularity subtraction. The moments of the stresslet expansion are multiples of the J n ( p ) J_n^{(p)} J n ( p ) of (28) , except the lowest one, which is 1 2 \frac{1}{2} 2 1 times the Stokeslet moment left uncancelled (§ 4.5 ). Adding the corresponding local correction gives the corrected double layer
( D δ , c o r r ( p ) q ) i = ∫ Γ T i j k , δ ( p ) ( q j − q 0 , j ) n k d s + χ q 0 , i + ∣ b ∣ N ( p ) ( λ ) 2 π V i , (36) (\mathcal{D}_{\delta,\mathrm{corr}}^{(p)}q)_i=\int_{\Gamma}T_{ijk,\delta}^{(p)}(q_j-q_{0,j})n_k\,\mathrm{d}s+\chi q_{0,i}+\frac{|b|\mathcal{N}^{(p)}(\lambda)}{2\pi}V_i, \tag*{(36)} ( D δ , corr ( p ) q ) i = ∫ Γ T ij k , δ ( p ) ( q j − q 0 , j ) n k d s + χ q 0 , i + 2 π ∣ b ∣ N ( p ) ( λ ) V i , (36)
where q 0 = q ( x 0 ) q_0=q(x_0) q 0 = q ( x 0 ) and
V = α − 3 { n 0 ( τ 0 ⋅ q s ) + τ 0 ( n 0 ⋅ q s ) } − b 2 α 3 n 0 ( n 0 ⋅ q s s ) + b 2 κ s 2 α 5 n 0 ( n 0 ⋅ q s ) . (37) \mathcal{V}=\alpha^{-3}\{n_0(\tau_0\cdot q_s)+\tau_0(n_0\cdot q_s)\}-\frac{b}{2\alpha^3}n_0(n_0\cdot q_{ss})+\frac{b^2\kappa_s}{2\alpha^5}n_0(n_0\cdot q_s). \tag*{(37)} V = α − 3 { n 0 ( τ 0 ⋅ q s ) + τ 0 ( n 0 ⋅ q s )} − 2 α 3 b n 0 ( n 0 ⋅ q ss ) + 2 α 5 b 2 κ s n 0 ( n 0 ⋅ q s ) . (37)
Here all quantities are evaluated at x 0 x_0 x 0 , κ s \kappa_s κ s is the arclength derivative of the curvature, and q s q_s q s , q s s q_{ss} q ss denote componentwise first and second arclength derivatives of the density.
Both corrections vanish at b = 0 b=0 b = 0 , so no correction is needed for targets on the curve. In the computations, a k ( p ) a_k^{(p)} a k ( p ) is obtained from (16)–(18), N ( p ) \mathcal{N}^{(p)} N ( p ) from (35) , and the vector V \mathcal{V} V directly from (37) . Density values and derivatives at x 0 x_0 x 0 are obtained from the trigonometric interpolants of the nodal values; arclength derivatives are computed at the nodes by Fourier differentiation, as ∣ x ′ ( t ) ∣ − 1 ∂ t |x'(t)|^{-1}\partial_t ∣ x ′ ( t ) ∣ − 1 ∂ t , and q s s q_{ss} q ss by differentiating the computed nodal values of q s q_s q s .
Remark 3.1. The regularized integrals approximate the layer potentials for evaluation, rather than representing exact Stokes flows with smoothed forces as in the method of regularized Stokeslets [16 ] . For fixed δ > 0 \delta>0 δ > 0 , the regularized velocity kernel is not divergence-free in general (holding λ \lambda λ fixed gives ∂ i S i j , δ ( p ) = r j ∂ ξ d ( p ) ( ξ ; λ ) / ( 2 π μ δ 2 ) \partial_iS_{ij,\delta}^{(p)}=r_j\partial_\xi d^{(p)}(\xi;\lambda)/(2\pi\mu\delta^2) ∂ i S ij , δ ( p ) = r j ∂ ξ d ( p ) ( ξ ; λ ) / ( 2 π μ δ 2 ) ), and the local corrections do not restore incompressibility at finite δ \delta δ ; the divergence-free condition is recovered to O ( δ p ) O(\delta^p) O ( δ p ) as δ → 0 \delta\to0 δ → 0 . The pressure is computed separately from the harmonic potentials S S S and D D D .
The pressure
We express the pressure for the Stokes layer potentials in terms of the harmonic potentials S S S and D D D of (4) and (23) , keeping the sign conventions of (3) and (30) . The pressure of the Stokes single layer is
p S ( y ) = − 1 2 π ∫ Γ r ⋅ f ( x ) r 2 d s x = − ∫ Γ ∇ x G ( r ) ⋅ f ( x ) d s x , r = x − y . (38) p^S(y)=-\frac{1}{2\pi}\int_\Gamma\frac{r\cdot f(x)}{r^2}\,ds_x=-\int_\Gamma\nabla_xG(r)\cdot f(x)\,ds_x,\qquad r=x-y. \tag*{(38)} p S ( y ) = − 2 π 1 ∫ Γ r 2 r ⋅ f ( x ) d s x = − ∫ Γ ∇ x G ( r ) ⋅ f ( x ) d s x , r = x − y . (38)
Decompose f = ( f ⋅ n ) n + ( f ⋅ τ ) τ f=(f\cdot n)n+(f\cdot\tau)\tau f = ( f ⋅ n ) n + ( f ⋅ τ ) τ ; the same split is used for surfaces in [13 Sect. 2] , which treats the tangential part by a subtraction instead. The normal part of (38) is − D [ f ⋅ n ] -D[f\cdot n] − D [ f ⋅ n ] . For the tangential part, ∂ τ G = ∂ s G \partial_\tau G=\partial_sG ∂ τ G = ∂ s G and periodic integration by parts give − ∫ Γ ∂ s G ( f ⋅ τ ) d s = S [ ( f ⋅ τ ) s ] -\int_\Gamma\partial_sG(f\cdot\tau)\,ds=S[(f\cdot\tau)_s] − ∫ Γ ∂ s G ( f ⋅ τ ) d s = S [( f ⋅ τ ) s ] . Hence
p S = S [ ( f ⋅ τ ) s ] − D [ f ⋅ n ] . (39) p^S=S[(f\cdot\tau)_s]-D[f\cdot n]. \tag*{(39)} p S = S [( f ⋅ τ ) s ] − D [ f ⋅ n ] . (39)
To compute p S p^S p S , we replace S S S and D D D by their regularized versions.
For y ∉ Γ y\notin\Gamma y ∈ / Γ , the pressure of the Stokes double layer is
p D ( y ) = 2 μ ∫ Γ q ( x ) ⋅ [ ∇ x 2 G ( r ) ] n ( x ) d s x = μ π ∫ Γ [ q ⋅ n r 2 − 2 ( q ⋅ r ) ( n ⋅ r ) r 4 ] d s x . (40) p^D(y)=2\mu\int_\Gamma q(x)\cdot[\nabla_x^2G(r)]n(x)\,ds_x=\frac{\mu}{\pi}\int_\Gamma\left[\frac{q\cdot n}{r^2}-2\frac{(q\cdot r)(n\cdot r)}{r^4}\right]\,ds_x. \tag*{(40)} p D ( y ) = 2 μ ∫ Γ q ( x ) ⋅ [ ∇ x 2 G ( r )] n ( x ) d s x = π μ ∫ Γ [ r 2 q ⋅ n − 2 r 4 ( q ⋅ r ) ( n ⋅ r ) ] d s x . (40)
where ∇ x 2 G \nabla_x^2G ∇ x 2 G denotes the Hessian matrix of G G G . This integral is hypersingular for y ∈ Γ y\in\Gamma y ∈ Γ . We therefore reduce it to S S S and D D D , whose limits on Γ \Gamma Γ follow from the continuity of S S S and from (24) . Writing q q q and its derivatives in the ( τ , n ) (\tau,n) ( τ , n ) frame and using τ s = − κ n \tau_s=-\kappa n τ s = − κn , n s = κ τ n_s=\kappa\tau n s = κ τ , and Δ G = 0 \Delta G=0 Δ G = 0 away from the source, integration by parts along Γ \Gamma Γ gives
∫ Γ q ⋅ ( ∇ x 2 G ) n d s = ∫ Γ ( q s ⋅ n ) ∂ τ G d s − ∫ Γ ( q s ⋅ τ ) ∂ n G d s . (41) \int_\Gamma q\cdot(\nabla_x^2G)n\,ds=\int_\Gamma(q_s\cdot n)\,\partial_\tau G\,ds-\int_\Gamma(q_s\cdot\tau)\,\partial_nG\,ds. \tag*{(41)} ∫ Γ q ⋅ ( ∇ x 2 G ) n d s = ∫ Γ ( q s ⋅ n ) ∂ τ G d s − ∫ Γ ( q s ⋅ τ ) ∂ n G d s . (41)
Integrating the first term by parts once more yields
p D = − 2 μ S [ ( q s ⋅ n ) s ] − 2 μ D [ q s ⋅ τ ] , ( q s ⋅ n ) s = q s s ⋅ n + κ q s ⋅ τ . (42) p^D=-2\mu S[(q_s\cdot n)_s]-2\mu D[q_s\cdot\tau],\qquad(q_s\cdot n)_s=q_{ss}\cdot n+\kappa q_s\cdot\tau. \tag*{(42)} p D = − 2 μ S [( q s ⋅ n ) s ] − 2 μ D [ q s ⋅ τ ] , ( q s ⋅ n ) s = q ss ⋅ n + κ q s ⋅ τ . (42)
For the sum of the Stokeslet and stresslet integrals, define
ϕ = ( f ⋅ τ ) s − 2 μ ( q s ⋅ n ) s , ψ = f ⋅ n + 2 μ q s ⋅ τ . (43) \phi= (f \cdot\tau)_s - 2\mu(q_s \cdot n)_s,\qquad\psi= f \cdot n + 2\mu q_s \cdot\tau. \tag*{(43)} ϕ = ( f ⋅ τ ) s − 2 μ ( q s ⋅ n ) s , ψ = f ⋅ n + 2 μ q s ⋅ τ . (43)
By (39) and (42) , the pressure is
p = S ϕ − D ψ . (44) p = S\phi- D\psi. \tag*{(44)} p = S ϕ − D ψ . (44)
The pressure can therefore be computed with the regularized kernels and coefficients of § 2 . The regularized pressure is
p δ = S δ ( p ) ϕ − D δ ( p ) ψ , (45) p_\delta= S_\delta^{(p)}\phi- D_\delta^{(p)}\psi, \tag*{(45)} p δ = S δ ( p ) ϕ − D δ ( p ) ψ , (45)
where D δ ( p ) ψ D_\delta^{(p)}\psi D δ ( p ) ψ is the subtracted double layer (27) with density ψ \psi ψ . For a periodic parametrization x ( t ) x(t) x ( t ) , we compute q s q_s q s , ( f ⋅ τ ) s (f \cdot\tau)_s ( f ⋅ τ ) s , and ( q s ⋅ n ) s (q_s \cdot n)_s ( q s ⋅ n ) s by Fourier differentiation, using ∂ s = ∣ x ′ ( t ) ∣ − 1 ∂ t \partial_s = \lvert x'(t)\rvert^{-1}\partial_t ∂ s = ∣ x ′ ( t ) ∣ − 1 ∂ t . Thus the hypersingular pressure kernel is replaced by tangential derivatives of smooth periodic densities.
Error estimates
The corrected Stokeslet and stresslet integrals have the same accuracy as the potentials S S S and D D D in Proposition 2.1 ; the proofs are given in § 4.6 .
Proposition 3.2 (Uniform Stokes accuracy). Let Γ \Gamma Γ , Λ \Lambda Λ , and δ 0 \delta_0 δ 0 be as in Proposition 2.1 , let p = 3 p = 3 p = 3 , 5 5 5 , or 7 7 7 , and let the vector densities f f f and q q q be smooth. Let S δ , c o r r ( p ) f S_{\delta,\mathrm{corr}}^{(p)}f S δ , corr ( p ) f and D δ , c o r r ( p ) q D_{\delta,\mathrm{corr}}^{(p)}q D δ , corr ( p ) q be given by (34) and (36) , with the kernels (32) and (33) , N ( p ) \mathcal{N}^{(p)} N ( p ) and V \mathcal{V} V from (35) and (37) , and the coefficients (16) –(18) at λ = b / δ \lambda= b/\delta λ = b / δ , and let S f Sf S f and D q Dq D q be the potentials (31) . Then for 0 < δ ≤ δ 0 0 < \delta\le\delta_0 0 < δ ≤ δ 0 ,
sup x 0 ∈ Γ , ∣ b ∣ ≤ Λ δ ( ∣ S δ , c o r r ( p ) f − S f ∣ + ∣ D δ , c o r r ( p ) q − D q ∣ ) ≤ C δ p . (46) \sup_{\substack{x_0 \in\Gamma,\ \lvert b\rvert\le\Lambda\delta}} \left(\left\lvert S_{\delta,\mathrm{corr}}^{(p)}f - Sf\right\rvert+ \left\lvert D_{\delta,\mathrm{corr}}^{(p)}q - Dq\right\rvert\right) \le C\delta^p. \tag*{(46)} x 0 ∈ Γ , ∣ b ∣ ≤ Λ δ sup ( S δ , corr ( p ) f − S f + D δ , corr ( p ) q − D q ) ≤ C δ p . (46) The constant C C C depends on finitely many derivatives of Γ \Gamma Γ and the densities, and on p p p and Λ \Lambda Λ , but not on δ \delta δ , b b b , or x 0 x_0 x 0 .
Since the pressure is computed from S S S and D D D , the same holds for the pressure.
Proposition 3.3 (Uniform pressure accuracy). Let Γ \Gamma Γ , Λ \Lambda Λ , δ 0 \delta_0 δ 0 , and the order be as in Proposition 2.1 , and let f f f and q q q be smooth, so that the pressure densities ϕ \phi ϕ and ψ \psi ψ of (43) are smooth. Let p δ p_\delta p δ be the regularized pressure (45) , with S δ ( p ) S_\delta^{(p)} S δ ( p ) and D δ ( p ) D_\delta^{(p)} D δ ( p ) given by (11) and (27) . Then for 0 < δ ≤ δ 0 0 < \delta\le\delta_0 0 < δ ≤ δ 0 ,
sup x 0 ∈ Γ , ∣ b ∣ ≤ Λ δ ∣ p δ − p ∣ ≤ C δ p . (47) \sup_{\substack{x_0 \in\Gamma,\ \lvert b\rvert\le\Lambda\delta}} \lvert p_\delta- p\rvert\le C\delta^p. \tag*{(47)} x 0 ∈ Γ , ∣ b ∣ ≤ Λ δ sup ∣ p δ − p ∣ ≤ C δ p . (47)
Remark 3.4. As for the Laplace double layer (Remark 2.2), for y ∈ Γ y \in\Gamma y ∈ Γ the stresslet kernel contracted with the normal, T i j k ( x − y ) n k ( x ) T_{ijk}(x-y)n_k(x) T ij k ( x − y ) n k ( x ) , is a smooth function of x ∈ Γ x \in\Gamma x ∈ Γ , with limit κ ( y ) τ i ( y ) τ j ( y ) / ( 2 π ) \kappa(y)\tau_i(y)\tau_j(y)/(2\pi) κ ( y ) τ i ( y ) τ j ( y ) / ( 2 π ) as x → y x \to y x → y [21 ] . With this diagonal value, the stresslet integral needs no regularization on the curve, whereas the Stokeslet and the single-layer part of the pressure remain singular on Γ \Gamma Γ .
Error analysis
We now derive the kernel representations and local corrections of § 2 and § 3 , establishing the uniform error bounds of Propositions 2.1 , 3.2 , and 3.3 . The analysis constructs a local asymptotic expansion in a tubular neighborhood of the closest point x 0 x_0 x 0 . Following [8 , 12 ] , each error term factors into a geometric Taylor coefficient and an error moment depending solely on λ = b / δ \lambda= b/\delta λ = b / δ .
The single layer
Let y = x 0 + b n 0 y=x_0+bn_0 y = x 0 + b n 0 be a target as in (2) , so that b b b is the signed normal distance from the curve to y y y , and let λ = b / δ \lambda=b/\delta λ = b / δ . Under the hypotheses of Proposition 2.1, the closest point x 0 x_0 x 0 is unique and 1 + b κ 0 ≥ 1 2 1+b\kappa_0\ge\frac{1}{2} 1 + b κ 0 ≥ 2 1 , where κ 0 \kappa_0 κ 0 is the curvature at x 0 x_0 x 0 . Let σ \sigma σ be signed arclength from x 0 x_0 x 0 , positive in the direction of τ 0 \tau_0 τ 0 , and let r ( σ ) = x ( σ ) − y = x ( σ ) − x 0 − b n 0 r(\sigma)=x(\sigma)-y=x(\sigma)-x_0-bn_0 r ( σ ) = x ( σ ) − y = x ( σ ) − x 0 − b n 0 . Taylor expansion of the curve about x 0 x_0 x 0 , with τ s = − κ n \tau_s=-\kappa n τ s = − κn and n s = κ τ n_s=\kappa\tau n s = κ τ , gives
r = − b n 0 + σ τ 0 − 1 2 κ 0 σ 2 n 0 − 1 6 σ 3 ( κ 0 2 τ 0 + κ s n 0 ) + O ( σ 4 ) , (48) r=-bn_0+\sigma\tau_0-\frac{1}{2}\kappa_0\sigma^2n_0-\frac{1}{6}\sigma^3(\kappa_0^2\tau_0+\kappa_sn_0)+O(\sigma^4), \tag*{(48)} r = − b n 0 + σ τ 0 − 2 1 κ 0 σ 2 n 0 − 6 1 σ 3 ( κ 0 2 τ 0 + κ s n 0 ) + O ( σ 4 ) , (48)
where κ s \kappa_s κ s is the arclength derivative of the curvature at x 0 x_0 x 0 . Hence r ⋅ n ( σ ) = − b + O ( σ 2 ) r\cdot n(\sigma)=-b+O(\sigma^2) r ⋅ n ( σ ) = − b + O ( σ 2 ) and r 2 − b 2 = α 2 σ 2 + 1 3 b κ s σ 3 + O ( σ 4 ) r^2-b^2=\alpha^2\sigma^2+\frac{1}{3}b\kappa_s\sigma^3+O(\sigma^4) r 2 − b 2 = α 2 σ 2 + 3 1 b κ s σ 3 + O ( σ 4 ) , with α = 1 + b κ 0 \alpha=\sqrt{1+b\kappa_0} α = 1 + b κ 0 . Since α 2 ≥ 1 2 \alpha^2\ge\frac{1}{2} α 2 ≥ 2 1 , the coordinate η = sgn ( σ ) r 2 − b 2 \eta=\operatorname{sgn}(\sigma)\sqrt{r^2-b^2} η = sgn ( σ ) r 2 − b 2 along the curve is a smooth, invertible function of σ \sigma σ near 0 0 0 , uniformly in x 0 x_0 x 0 and b b b , and r 2 = η 2 + b 2 r^2=\eta^2+b^2 r 2 = η 2 + b 2 . Taking the square root and inverting the series gives
η = α σ + b κ s 6 α σ 2 + O ( σ 3 ) , σ = η α − b κ s 6 α 4 η 2 + O ( η 3 ) , (49) \eta=\alpha\sigma+\frac{b\kappa_s}{6\alpha}\sigma^2+O(\sigma^3),\qquad\sigma=\frac{\eta}{\alpha}-\frac{b\kappa_s}{6\alpha^4}\eta^2+O(\eta^3), \tag*{(49)} η = α σ + 6 α b κ s σ 2 + O ( σ 3 ) , σ = α η − 6 α 4 b κ s η 2 + O ( η 3 ) , (49)
so the Jacobian J ( η ) = d σ / d η \mathcal{J}(\eta)=\mathrm{d}\sigma/\mathrm{d}\eta J ( η ) = d σ / d η satisfies J ( 0 ) = 1 / α \mathcal{J}(0)=1/\alpha J ( 0 ) = 1/ α . The coordinate η \eta η is the two-dimensional analogue of the one used in [12 Sect. 3] . In this coordinate, a radial kernel becomes a function of η 2 + b 2 \eta^2+b^2 η 2 + b 2 , and the density, the Jacobian, and the numerators of the Stokes kernels have Taylor expansions in η \eta η . With ξ = ( η 2 + b 2 ) / δ 2 \xi=(\eta^2+b^2)/\delta^2 ξ = ( η 2 + b 2 ) / δ 2 , the Gaussian e − ξ e^{-\xi} e − ξ is concentrated where η 2 + b 2 = O ( δ 2 ) \eta^2+b^2=O(\delta^2) η 2 + b 2 = O ( δ 2 ) , that is, on ∣ η ∣ = O ( δ ) |\eta|=O(\delta) ∣ η ∣ = O ( δ ) when b = O ( δ ) b=O(\delta) b = O ( δ ) , uniformly for ∣ λ ∣ ≤ Λ |\lambda|\le\Lambda ∣ λ ∣ ≤ Λ . The part of the curve outside a fixed neighborhood of x 0 x_0 x 0 therefore contributes only exponentially small terms, and only the behavior of the integrand near η = 0 \eta=0 η = 0 matters.
For the Laplace single layer, the kernel difference is G δ ( 1 ) − G = 1 4 π E 1 ( ξ ) G_\delta^{(1)}-G=\frac{1}{4\pi}E_1(\xi) G δ ( 1 ) − G = 4 π 1 E 1 ( ξ ) . We combine the density and the Jacobian into the smooth function
w ( η ; b ) = f ( x ( σ ( η ) ) ) J ( η ) , (50) w(\eta;b)=f(x(\sigma(\eta)))\mathcal{J}(\eta), \tag*{(50)} w ( η ; b ) = f ( x ( σ ( η ))) J ( η ) , (50)
and insert a smooth even cutoff function that equals one near η = 0 \eta=0 η = 0 and vanishes outside the coordinate patch. Suppressing the cutoff in the notation and extending the integral to R \mathbb{R} R , which changes it by an exponentially small amount, we obtain
S δ ( 1 ) f − S f = 1 4 π ∫ R w ( η ; b ) E 1 ( η 2 + b 2 δ 2 ) d η + exponentially small terms . (51) S_\delta^{(1)}f-Sf=\frac{1}{4\pi}\int_{\mathbb{R}}w(\eta;b)E_1\left(\frac{\eta^2+b^2}{\delta^2}\right)\,\mathrm{d}\eta+\text{exponentially small terms}. \tag*{(51)} S δ ( 1 ) f − S f = 4 π 1 ∫ R w ( η ; b ) E 1 ( δ 2 η 2 + b 2 ) d η + exponentially small terms . (51)
Write the even part of w w w as
w e v ( η ; b ) = 1 2 [ w ( η ; b ) + w ( − η ; b ) ] = ∑ n ≥ 0 n e v e n w n ( x 0 , b ) η n ; (52) w_{\mathrm{ev}}(\eta;b)=\frac{1}{2}[w(\eta;b)+w(-\eta;b)]=\sum_{\substack{n\ge0\\ n\ \mathrm{even}}}w_n(x_0,b)\eta^n; \tag*{(52)} w ev ( η ; b ) = 2 1 [ w ( η ; b ) + w ( − η ; b )] = n ≥ 0 n even ∑ w n ( x 0 , b ) η n ; (52)
finite Taylor expansions with remainder suffice throughout, so analyticity is not required. Only w e v w_{\mathrm{ev}} w ev contributes, because the kernel factor is even in η \eta η for fixed b b b , so the odd part of the integrand integrates to zero. Substituting (52) into (51) and setting η = δ t \eta=\delta t η = δ t gives (8). For the kernel (10), the same steps give (13).
The coefficients
We first derive the form of the correction terms. Differentiating E 1 ( ξ ) = ∫ ξ ∞ e − v v − 1 d v E_1(\xi)=\int_{\xi}^{\infty}e^{-v}v^{-1}\,\mathrm{d}v E 1 ( ξ ) = ∫ ξ ∞ e − v v − 1 d v k k k times and substituting u = ξ v u=\xi v u = ξ v gives
ξ k E 1 [ k ] ( ξ ) = ( − 1 ) k ∫ ξ ∞ u k − 1 e − u d u = ( − 1 ) k ( k − 1 ) ! e − ξ ∑ ℓ = 0 k − 1 ξ ℓ ℓ ! , k ≥ 1 , (53) \xi^k E_1^{[k]}(\xi)=(-1)^k\int_{\xi}^{\infty}u^{k-1}e^{-u}\,\mathrm{d}u=(-1)^k(k-1)!e^{-\xi}\sum_{\ell=0}^{k-1}\frac{\xi^\ell}{\ell!},\qquad k\ge1, \tag*{(53)} ξ k E 1 [ k ] ( ξ ) = ( − 1 ) k ∫ ξ ∞ u k − 1 e − u d u = ( − 1 ) k ( k − 1 )! e − ξ ℓ = 0 ∑ k − 1 ℓ ! ξ ℓ , k ≥ 1 , (53)
where the second equality follows by integrating by parts k − 1 k-1 k − 1 times. Differentiating the integral gives d d ξ [ ξ k E 1 [ k ] ( ξ ) ] = ( − 1 ) k + 1 ξ k − 1 e − ξ \frac{d}{d\xi}\left[\xi^k E_1^{[k]}(\xi)\right]=(-1)^{k+1}\xi^{k-1}e^{-\xi} d ξ d [ ξ k E 1 [ k ] ( ξ ) ] = ( − 1 ) k + 1 ξ k − 1 e − ξ .
We now solve the moment conditions (14) . The polynomial factor of ξ k E 1 [ k ] \xi^k E_1^{[k]} ξ k E 1 [ k ] has degree k − 1 k-1 k − 1 , which increases with k k k , so the conditions can be solved by successive elimination, as we now show. The E 1 E_1 E 1 term in (9) contributes a function H n ( λ ) H_n(\lambda) H n ( λ ) , n = 0 , 2 , 4 , … n=0,2,4,\ldots n = 0 , 2 , 4 , … , to each moment, and each correction term ξ k E 1 [ k ] \xi^k E_1^{[k]} ξ k E 1 [ k ] contributes an elementary Gaussian integral. Together they give a small linear system for the a k ( p ) a_k^{(p)} a k ( p ) . For even n ≥ 0 n\geq0 n ≥ 0 , inserting the integral definition of E 1 E_1 E 1 and interchanging the order of integration gives
∫ R t n E 1 ( t 2 + λ 2 ) d t = ∫ R t n ∫ 1 ∞ e − ( t 2 + λ 2 ) v v d v d t = Γ ( n + 1 2 ) ∫ 1 ∞ v − ( n + 3 ) / 2 e − λ 2 v d v = Γ ( n + 1 2 ) e − λ 2 H n ( λ ) . (54) \begin{aligned}
\int_{\mathbb{R}} t^n E_1(t^2+\lambda^2)\,\mathrm{d}t
&=\int_{\mathbb{R}}t^n\int_1^\infty\frac{e^{-(t^2+\lambda^2)v}}{v}\,\mathrm{d}v\,\mathrm{d}t\\
&=\Gamma\left(\frac{n+1}{2}\right)\int_1^\infty v^{-(n+3)/2}e^{-\lambda^2v}\,\mathrm{d}v\\
&=\Gamma\left(\frac{n+1}{2}\right)e^{-\lambda^2}H_n(\lambda). \tag*{(54)}
\end{aligned} ∫ R t n E 1 ( t 2 + λ 2 ) d t = ∫ R t n ∫ 1 ∞ v e − ( t 2 + λ 2 ) v d v d t = Γ ( 2 n + 1 ) ∫ 1 ∞ v − ( n + 3 ) /2 e − λ 2 v d v = Γ ( 2 n + 1 ) e − λ 2 H n ( λ ) . (54)
Here H n H_n H n is defined in (15) ; the factor e − λ 2 e^{-\lambda^2} e − λ 2 is taken out because it is common to all terms of the moment conditions.
By (53) , the k k k -th correction term in (9) is a polynomial in ξ = t 2 + λ 2 \xi=t^2+\lambda^2 ξ = t 2 + λ 2 times e − ( t 2 + λ 2 ) e^{-(t^2+\lambda^2)} e − ( t 2 + λ 2 ) . Expanding in powers of t 2 t^2 t 2 and applying ∫ R t 2 i e − t 2 d t = Γ ( i + 1 2 ) \int_{\mathbb{R}}t^{2i}e^{-t^2}\,\mathrm{d}t=\Gamma\left(i+\frac{1}{2}\right) ∫ R t 2 i e − t 2 d t = Γ ( i + 2 1 ) to each term gives, for even n ≥ 0 n\geq0 n ≥ 0 ,
∫ R t n ξ k E 1 [ k ] ( ξ ) d t = ( − 1 ) k ( k − 1 ) ! e − λ 2 ∑ ℓ = 0 k − 1 1 ℓ ! ∑ i = 0 ℓ ( ℓ i ) λ 2 ( ℓ − i ) Γ ( i + n + 1 2 ) . (55) \int_{\mathbb{R}}t^n\xi^k E_1^{[k]}(\xi)\,\mathrm{d}t=(-1)^k(k-1)!e^{-\lambda^2}\sum_{\ell=0}^{k-1}\frac{1}{\ell!}\sum_{i=0}^{\ell}\binom{\ell}{i}\lambda^{2(\ell-i)}\Gamma\left(i+\frac{n+1}{2}\right). \tag*{(55)} ∫ R t n ξ k E 1 [ k ] ( ξ ) d t = ( − 1 ) k ( k − 1 )! e − λ 2 ℓ = 0 ∑ k − 1 ℓ ! 1 i = 0 ∑ ℓ ( i ℓ ) λ 2 ( ℓ − i ) Γ ( i + 2 n + 1 ) . (55)
After division by the common factor Γ ( ( n + 1 ) / 2 ) e − λ 2 \Gamma((n+1)/2)e^{-\lambda^2} Γ (( n + 1 ) /2 ) e − λ 2 , the k k k -th correction term contributes a polynomial in λ 2 \lambda^2 λ 2 of degree $k-1 with leading term ( − 1 ) k λ 2 ( k − 1 ) (-1)^k\lambda^{2(k-1)} ( − 1 ) k λ 2 ( k − 1 ) , and the moment conditions become
H n − a 1 ( p ) + ( λ 2 + n + 3 2 ) a 2 ( p ) − [ λ 4 + ( n + 3 ) λ 2 + ( n + 3 ) ( n + 5 ) 4 ] a 3 ( p ) = 0 , (56) H_n-a_1^{(p)}+\left(\lambda^2+\frac{n+3}{2}\right)a_2^{(p)}-\left[\lambda^4+(n+3)\lambda^2+\frac{(n+3)(n+5)}{4}\right]a_3^{(p)}=0, \tag*{(56)} H n − a 1 ( p ) + ( λ 2 + 2 n + 3 ) a 2 ( p ) − [ λ 4 + ( n + 3 ) λ 2 + 4 ( n + 3 ) ( n + 5 ) ] a 3 ( p ) = 0 , (56)
for n = 0 , 2 , … , p − 3 n=0,2,\ldots,p-3 n = 0 , 2 , … , p − 3 . Here a k ( p ) = 0 a_k^{(p)}=0 a k ( p ) = 0 for k > m = ( p − 1 ) / 2 k>m=(p-1)/2 k > m = ( p − 1 ) /2 ; the alternating signs come from the factors ( − 1 ) k (-1)^k ( − 1 ) k in ξ k E 1 [ k ] \xi^k E_1^{[k]} ξ k E 1 [ k ] .
Since the coefficient of a k ( p ) a_k^{(p)} a k ( p ) in (56) is a polynomial of degree k − 1 k-1 k − 1 in n n n , taking differences of the equations for consecutive values of n n n eliminates the unknowns one at a time. For p = 7 p=7 p = 7 , subtracting the equation for n = 0 n=0 n = 0 from that for n = 2 n=2 n = 2 gives a 2 ( 7 ) = H 0 − H 2 + ( 2 λ 2 + 5 ) a 3 ( 7 ) a_2^{(7)}=H_0-H_2+(2\lambda^2+5)a_3^{(7)} a 2 ( 7 ) = H 0 − H 2 + ( 2 λ 2 + 5 ) a 3 ( 7 ) , the second difference gives a 3 ( 7 ) = 1 2 ( H 0 − 2 H 2 + H 4 ) a_3^{(7)}=\frac{1}{2}(H_0-2H_2+H_4) a 3 ( 7 ) = 2 1 ( H 0 − 2 H 2 + H 4 ) , and the equation for n = 0 n=0 n = 0 then gives a 1 ( 7 ) a_1^{(7)} a 1 ( 7 ) . The cases p = 5 p=5 p = 5 and p = 3 p=3 p = 3 follow in the same way, which gives (16) –(18) .
The double layer
For the local error analysis of the subtracted double layer (27) , define the numerator of its integrand, without the factor d ( p ) / ( 2 π r 2 ) d^{(p)}/(2\pi r^2) d ( p ) / ( 2 π r 2 ) of (26) and with the Jacobian included,
w D ( η ; b ) = [ n ( x ) ⋅ ( x − y ) ] [ g ( x ) − g ( x 0 ) ] J ( η ) , w D ( 0 ; b ) = 0 , (57) w_D(\eta;b)=\left[n(x)\cdot(x-y)\right]\left[g(x)-g(x_0)\right]\mathcal{J}(\eta),\qquad w_D(0;b)=0, \tag*{(57)} w D ( η ; b ) = [ n ( x ) ⋅ ( x − y ) ] [ g ( x ) − g ( x 0 ) ] J ( η ) , w D ( 0 ; b ) = 0 , (57)
with the Jacobian J \mathcal{J} J of § 4.1 . The terms χ g ( x 0 ) \chi g(x_0) χ g ( x 0 ) in (27) and in D g = ∫ Γ K ( g − g ( x 0 ) ) d s + χ g ( x 0 ) Dg=\int_\Gamma K\left(g-g(x_0)\right)\,\mathrm{d}s+\chi g(x_0) D g = ∫ Γ K ( g − g ( x 0 ) ) d s + χ g ( x 0 ) cancel, and K δ ( p ) − K = − n ( x ) ⋅ ( x − y ) Q ( p ) ( ξ ; λ ) / ( 2 π r 2 ) K_\delta^{(p)}-K=-n(x)\cdot(x-y)Q^{(p)}(\xi;\lambda)/(2\pi r^2) K δ ( p ) − K = − n ( x ) ⋅ ( x − y ) Q ( p ) ( ξ ; λ ) / ( 2 π r 2 ) by (26) , since d ( p ) = 1 − Q ( p ) d^{(p)}=1-Q^{(p)} d ( p ) = 1 − Q ( p ) , with Q ( p ) Q^{(p)} Q ( p ) as in (20) . The argument of § 4.1 , applied to the even part of w D w_D w D , which begins with η 2 \eta^2 η 2 , then gives
D δ ( p ) g − D g ∼ − 1 2 π ∑ n ≥ 0 n e v e n w D , n + 2 ( x 0 , b ) δ n + 1 J n ( p ) ( λ ) , (58) D_\delta^{(p)}g-Dg\sim-\frac{1}{2\pi}\sum_{\substack{n\geq0\\ n\ {\rm even}}}w_{D,n+2}(x_0,b)\,\delta^{n+1}J_n^{(p)}(\lambda), \tag*{(58)} D δ ( p ) g − D g ∼ − 2 π 1 n ≥ 0 n even ∑ w D , n + 2 ( x 0 , b ) δ n + 1 J n ( p ) ( λ ) , (58)
with the moments J n ( p ) J_n^{(p)} J n ( p ) of (28) , where w D , n + 2 ( x 0 , b ) w_{D,n+2}(x_0,b) w D , n + 2 ( x 0 , b ) is the coefficient of η n + 2 \eta^{n+2} η n + 2 in the Taylor expansion of w D ( η ; b ) w_D(\eta;b) w D ( η ; b ) . The moments J n ( p ) J_n^{(p)} J n ( p ) are multiples of the single-layer moments.
Lemma 4.1 (Relation between the moments). For every even n ≥ 0 n\geq0 n ≥ 0 and every λ \lambda λ ,
J n ( p ) ( λ ) = n + 1 2 I n ( p ) ( λ ) . (59) J_n^{(p)}(\lambda)=\frac{n+1}{2}I_n^{(p)}(\lambda). \tag*{(59)} J n ( p ) ( λ ) = 2 n + 1 I n ( p ) ( λ ) . (59) Proof. Put u = t 2 + λ 2 u=t^2+\lambda^2 u = t 2 + λ 2 . By (20) , Q ( p ) ( u ; λ ) = − u ∂ u s ( p ) ( u ; λ ) Q^{(p)}(u;\lambda)=-u\partial_u s^{(p)}(u;\lambda) Q ( p ) ( u ; λ ) = − u ∂ u s ( p ) ( u ; λ ) . Integrating by parts and using
d d t s ( p ) ( u ; λ ) = 2 t ∂ u s ( p ) ( u ; λ ) , \frac{d}{dt}s^{(p)}(u;\lambda)=2t\partial_u s^{(p)}(u;\lambda), d t d s ( p ) ( u ; λ ) = 2 t ∂ u s ( p ) ( u ; λ ) , J n ( p ) = − ∫ R t n + 2 ∂ u s ( p ) ( u ; λ ) d t = − 1 2 ∫ R t n + 1 d d t s ( p ) ( u ; λ ) d t = n + 1 2 ∫ R t n s ( p ) ( u ; λ ) d t . J_n^{(p)}=-\int_{\mathbb{R}}t^{n+2}\partial_u s^{(p)}(u;\lambda)\,dt=-\frac{1}{2}\int_{\mathbb{R}}t^{n+1}\frac{d}{dt}s^{(p)}(u;\lambda)\,dt=\frac{n+1}{2}\int_{\mathbb{R}}t^n s^{(p)}(u;\lambda)\,dt. J n ( p ) = − ∫ R t n + 2 ∂ u s ( p ) ( u ; λ ) d t = − 2 1 ∫ R t n + 1 d t d s ( p ) ( u ; λ ) d t = 2 n + 1 ∫ R t n s ( p ) ( u ; λ ) d t . The boundary terms vanish at infinity because s ( p ) s^{(p)} s ( p ) decays like a Gaussian times algebraic factors. At λ = 0 \lambda=0 λ = 0 , split the integration at t = 0 t=0 t = 0 ; the additional endpoint contributions vanish since t n + 1 log ( t 2 ) → 0 t^{n+1}\log(t^2)\to0 t n + 1 log ( t 2 ) → 0 .
Hence the coefficients that cancel the first m m m moments I n ( p ) I_n^{(p)} I n ( p ) also cancel the first m m m moments J n ( p ) J_n^{(p)} J n ( p ) , and the double layer has order p p p with the same coefficients. This relation is the two-dimensional analogue of the relation between single- and double-layer moments in [13 eq. (34)] .
Singularity subtraction is needed for uniform accuracy near the curve. Without it, the regularized double layer ∫ Γ K δ ( p ) g d s \int_\Gamma K_\delta^{(p)}g\,ds ∫ Γ K δ ( p ) g d s does not converge uniformly for b = λ δ b=\lambda\delta b = λ δ with fixed λ ≠ 0 \lambda\ne0 λ = 0 , because the regularized kernel does not satisfy Gauss’s identity (25) . By (25) –(26) , the kernel difference integrates to
∫ Γ ( K δ ( p ) − K ) d s x = β ( p ) ( λ ) ( 1 + O ( δ ) ) , β ( p ) ( λ ) = λ 2 π ∫ R Q ( p ) ( t 2 + λ 2 ; λ ) t 2 + λ 2 d t , (60) \int_\Gamma\left(K_\delta^{(p)}-K\right)\,ds_x=\beta^{(p)}(\lambda)\left(1+O(\delta)\right),\qquad\beta^{(p)}(\lambda)=\frac{\lambda}{2\pi}\int_{\mathbb{R}}\frac{Q^{(p)}(t^2+\lambda^2;\lambda)}{t^2+\lambda^2}\,dt, \tag*{(60)} ∫ Γ ( K δ ( p ) − K ) d s x = β ( p ) ( λ ) ( 1 + O ( δ ) ) , β ( p ) ( λ ) = 2 π λ ∫ R t 2 + λ 2 Q ( p ) ( t 2 + λ 2 ; λ ) d t , (60)
so without subtraction the error tends to g ( x 0 ) β ( p ) ( λ ) g(x_0)\beta^{(p)}(\lambda) g ( x 0 ) β ( p ) ( λ ) as δ → 0 \delta\to0 δ → 0 , which is O ( 1 ) O(1) O ( 1 ) . Singularity subtraction removes this error exactly. Although D g = ∫ K ( g − g ˉ ) d s + χ g ˉ Dg=\int K(g-\bar g)\,ds+\chi\bar g D g = ∫ K ( g − g ˉ ) d s + χ g ˉ holds for any constant g ˉ \bar g g ˉ , we subtract the value at the closest point, since then the numerator vanishes at η = 0 \eta=0 η = 0 and the error expansion (58) holds.
The Stokeslet
Substituting (10) into (32) and comparing with (30) gives
S i j , δ ( p ) − S i j = − δ i j 8 π μ s ( p ) ( ξ ; λ ) − 1 4 π μ r i r j r 2 Q ( p ) ( ξ ; λ ) . (61) S_{ij,\delta}^{(p)}-S_{ij}=-\frac{\delta_{ij}}{8\pi\mu}s^{(p)}(\xi;\lambda)-\frac{1}{4\pi\mu}\frac{r_i r_j}{r^2}Q^{(p)}(\xi;\lambda). \tag*{(61)} S ij , δ ( p ) − S ij = − 8 π μ δ ij s ( p ) ( ξ ; λ ) − 4 π μ 1 r 2 r i r j Q ( p ) ( ξ ; λ ) . (61)
The first term is a multiple of the regularizing function, whose moments vanish by (14) . The second term contains r i r j / r 2 r_i r_j/r^2 r i r j / r 2 . By (48) and (49) , the numerator r i r j r_i r_j r i r j has the expansion
r i r j = b 2 ( n 0 ) i ( n 0 ) j − b α η [ ( n 0 ) i ( τ 0 ) j + ( τ 0 ) i ( n 0 ) j ] + η 2 Θ i j ( b ) + O ( η 3 ) , (62) r_i r_j=b^2(n_0)_i(n_0)_j-\frac{b}{\alpha}\eta\left[(n_0)_i(\tau_0)_j+(\tau_0)_i(n_0)_j\right]+\eta^2\Theta_{ij}(b)+O(\eta^3), \tag*{(62)} r i r j = b 2 ( n 0 ) i ( n 0 ) j − α b η [ ( n 0 ) i ( τ 0 ) j + ( τ 0 ) i ( n 0 ) j ] + η 2 Θ ij ( b ) + O ( η 3 ) , (62)
where Θ i j ( b ) \Theta_{ij}(b) Θ ij ( b ) is a smooth tensor coefficient, bounded for b b b in the band around the curve, with Θ i j ( 0 ) = ( τ 0 ) i ( τ 0 ) j \Theta_{ij}(0)=(\tau_0)_i(\tau_0)_j Θ ij ( 0 ) = ( τ 0 ) i ( τ 0 ) j . After division by r 2 = η 2 + b 2 r^2=\eta^2+b^2 r 2 = η 2 + b 2 and the scaling η = δ t \eta=\delta t η = δ t , an even term η n + 2 \eta^{n+2} η n + 2 , n ≥ 0 n\geq0 n ≥ 0 , gives δ n + 1 J n ( p ) \delta^{n+1}J_n^{(p)} δ n + 1 J n ( p ) , with J n ( p ) J_n^{(p)} J n ( p ) as in (28) . These are multiples of the I n ( p ) I_n^{(p)} I n ( p ) by Lemma 4.1 , so they impose no new conditions. The only even term without a factor η 2 \eta^2 η 2 is b 2 ( n 0 ) i ( n 0 ) j b^2(n_0)_i(n_0)_j b 2 ( n 0 ) i ( n 0 ) j , and it gives δ − 1 b 2 ∫ R Q ( p ) ( t 2 + λ 2 ; λ ) ( t 2 + λ 2 ) − 1 d t = ∣ b ∣ N ( p ) ( λ ) \delta^{-1}b^2\int_{\mathbb{R}}Q^{(p)}(t^2+\lambda^2;\lambda)(t^2+\lambda^2)^{-1}\,dt=|b|\mathcal{N}^{(p)}(\lambda) δ − 1 b 2 ∫ R Q ( p ) ( t 2 + λ 2 ; λ ) ( t 2 + λ 2 ) − 1 d t = ∣ b ∣ N ( p ) ( λ ) , where
N ( p ) ( λ ) = ∣ λ ∣ ∫ R Q ( p ) ( t 2 + λ 2 ; λ ) t 2 + λ 2 d t , λ ≠ 0. (63) \mathcal{N}^{(p)}(\lambda)=|\lambda|\int_{\mathbb{R}}\frac{Q^{(p)}(t^2+\lambda^2;\lambda)}{t^2+\lambda^2}\,dt,\qquad\lambda\ne0. \tag*{(63)} N ( p ) ( λ ) = ∣ λ ∣ ∫ R t 2 + λ 2 Q ( p ) ( t 2 + λ 2 ; λ ) d t , λ = 0. (63)
The conditions (14) do not cancel this term, and we remove it by a local correction.
The integral in (63) behaves like π / ∣ λ ∣ \pi/|\lambda| π /∣ λ ∣ as λ → 0 \lambda\to0 λ → 0 , but N ( p ) \mathcal{N}^{(p)} N ( p ) has the closed form (35) , which is bounded and equal to π \pi π at λ = 0 \lambda=0 λ = 0 . To see this, write P ( p ) ( ξ ; λ ) = 1 + ξ Z P^{(p)}(\xi;\lambda)=1+\xi Z P ( p ) ( ξ ; λ ) = 1 + ξ Z , where Z = ∑ k = 1 m ( − 1 ) k a k ( p ) ( λ ) ξ k − 1 Z=\sum_{k=1}^{m}(-1)^k a_k^{(p)}(\lambda)\xi^{k-1} Z = ∑ k = 1 m ( − 1 ) k a k ( p ) ( λ ) ξ k − 1 , and split the integrand into two parts. The first part, e − ξ / ( t 2 + λ 2 ) e^{-\xi}/(t^2+\lambda^2) e − ξ / ( t 2 + λ 2 ) , gives π erfc ∣ λ ∣ \pi\operatorname{erfc}|\lambda| π erfc ∣ λ ∣ , because ∫ R e − t 2 / ( t 2 + λ 2 ) d t = ( π / ∣ λ ∣ ) e λ 2 erfc ∣ λ ∣ \int_{\mathbb{R}}e^{-t^2}/(t^2+\lambda^2)\,dt=(\pi/|\lambda|)e^{\lambda^2}\operatorname{erfc}|\lambda| ∫ R e − t 2 / ( t 2 + λ 2 ) d t = ( π /∣ λ ∣ ) e λ 2 erfc ∣ λ ∣ . The second part, ξ Z ( ξ ) e − ξ / ( t 2 + λ 2 ) = Z ( ξ ) e − ξ \xi Z(\xi)e^{-\xi}/(t^2+\lambda^2)=Z(\xi)e^{-\xi} ξ Z ( ξ ) e − ξ / ( t 2 + λ 2 ) = Z ( ξ ) e − ξ , is a polynomial times a Gaussian; integrating it term by term as in (55) gives the remaining terms of (35) .
To remove the remaining Stokeslet term, we need the constant term of the even part of the numerator, including the Jacobian J \mathcal{J} J , and set
w i U ( η ; b ) = r i r j ( η ; b ) f j ( x ( σ ( η ) ) ) J ( η ) . (64) w_i^U(\eta;b)=r_i r_j(\eta;b)f_j(x(\sigma(\eta)))\mathcal{J}(\eta). \tag*{(64)} w i U ( η ; b ) = r i r j ( η ; b ) f j ( x ( σ ( η ))) J ( η ) . (64)
The second term of (61) then contributes − 1 4 π μ ∫ R w i U ( η ; b ) Q ( p ) ( ξ ; λ ) / ( η 2 + b 2 ) d η -\frac{1}{4\pi\mu}\int_{\mathbb{R}}w_i^U(\eta;b)Q^{(p)}(\xi;\lambda)/(\eta^2+b^2)\,d\eta − 4 π μ 1 ∫ R w i U ( η ; b ) Q ( p ) ( ξ ; λ ) / ( η 2 + b 2 ) d η to the error. Since r i r j f j = b 2 ( n 0 ) i ( n 0 ⋅ f 0 ) + O ( η ) r_i r_j f_j=b^2(n_0)_i(n_0\cdot f_0)+O(\eta) r i r j f j = b 2 ( n 0 ) i ( n 0 ⋅ f 0 ) + O ( η ) by (62) and J ( 0 ) = α − 1 \mathcal{J}(0)=\alpha^{-1} J ( 0 ) = α − 1 by (49) , the constant term of w i U w_i^U w i U is ( b 2 / α ) ( n 0 ) i ( n 0 ⋅ f 0 ) (b^2/\alpha)(n_0)_i(n_0\cdot f_0) ( b 2 / α ) ( n 0 ) i ( n 0 ⋅ f 0 ) , and by (63) its contribution to the regularization error is − ∣ b ∣ N ( p ) ( λ ) 4 π μ α ( n 0 ) i ( n 0 ⋅ f 0 ) -\frac{|b|\mathcal{N}^{(p)}(\lambda)}{4\pi\mu\alpha}(n_0)_i(n_0\cdot f_0) − 4 π μα ∣ b ∣ N ( p ) ( λ ) ( n 0 ) i ( n 0 ⋅ f 0 ) , which vanishes as b → 0 b\to0 b → 0 . Subtracting it gives the correction in (34) .
The stresslet
From (21) , d ~ ( p ) = 1 − ξ 2 ∂ ξ 2 s ( p ) \widetilde{d}^{(p)}=1-\xi^2\partial_\xi^2s^{(p)} d ( p ) = 1 − ξ 2 ∂ ξ 2 s ( p ) at fixed λ \lambda λ . Comparing (33) with (30) therefore gives the exact identity
T i j k , δ ( p ) − T i j k = − 1 π r i r j r k r 4 R ( p ) ( ξ ; λ ) , R ( p ) ( ξ ; λ ) = 1 − d ~ ( p ) ( ξ ; λ ) = ξ 2 ∂ ξ 2 s ( p ) ( ξ ; λ ) . (65) T_{ijk,\delta}^{(p)}-T_{ijk}=-\frac{1}{\pi}\frac{r_i r_j r_k}{r^4}R^{(p)}(\xi;\lambda),\qquad R^{(p)}(\xi;\lambda)=1-\widetilde{d}^{(p)}(\xi;\lambda)=\xi^2\partial_\xi^2s^{(p)}(\xi;\lambda). \tag*{(65)} T ij k , δ ( p ) − T ij k = − π 1 r 4 r i r j r k R ( p ) ( ξ ; λ ) , R ( p ) ( ξ ; λ ) = 1 − d ( p ) ( ξ ; λ ) = ξ 2 ∂ ξ 2 s ( p ) ( ξ ; λ ) . (65)
We subtract the density value at x 0 x_0 x 0 as in (27) and include the Jacobian in the numerator, which after contraction with n k ( x ) n_k(x) n k ( x ) is
w i H ( η ; b ) = r i r j ( r ⋅ n ( x ) ) ( q j ( x ) − q 0 , j ) J ( η ) . (66) w_i^H(\eta;b)=r_i r_j(r\cdot n(x))(q_j(x)-q_{0,j})\mathcal{J}(\eta). \tag*{(66)} w i H ( η ; b ) = r i r j ( r ⋅ n ( x )) ( q j ( x ) − q 0 , j ) J ( η ) . (66)
The kernel difference (65) then contributes − 1 π ∫ R w i H ( η ; b ) R ( p ) ( ξ ; λ ) / ( η 2 + b 2 ) 2 d η -\frac{1}{\pi}\int_{\mathbb{R}}w_i^H(\eta;b)R^{(p)}(\xi;\lambda)/(\eta^2+b^2)^2\,d\eta − π 1 ∫ R w i H ( η ; b ) R ( p ) ( ξ ; λ ) / ( η 2 + b 2 ) 2 d η , and the stresslet moments are
K n ( p ) ( λ ) = ∫ R t n + 2 ( t 2 + λ 2 ) 2 R ( p ) ( t 2 + λ 2 ; λ ) d t , n = 0 , 2 , 4 , … . (67) K_n^{(p)}(\lambda)=\int_{\mathbb{R}}\frac{t^{n+2}}{(t^2+\lambda^2)^2}R^{(p)}(t^2+\lambda^2;\lambda)\,dt,\qquad n=0,2,4,\ldots. \tag*{(67)} K n ( p ) ( λ ) = ∫ R ( t 2 + λ 2 ) 2 t n + 2 R ( p ) ( t 2 + λ 2 ; λ ) d t , n = 0 , 2 , 4 , … . (67)
Integrating by parts twice reduces them to the scalar moments. With u = t 2 + λ 2 u=t^2+\lambda^2 u = t 2 + λ 2 and ∂ u 2 s ( p ) ( u ; λ ) = 1 2 t d d t ∂ u s ( p ) ( u ; λ ) \partial_u^2s^{(p)}(u;\lambda)=\frac{1}{2t}\frac{d}{dt}\partial_us^{(p)}(u;\lambda) ∂ u 2 s ( p ) ( u ; λ ) = 2 t 1 d t d ∂ u s ( p ) ( u ; λ ) ,
K n ( p ) = ∫ R t n + 2 ∂ u 2 s ( p ) ( u ; λ ) d t = − n + 1 2 ∫ R t n ∂ u s ( p ) ( u ; λ ) d t = ( n + 1 ) ( n − 1 ) 4 ∫ R t n − 2 s ( p ) ( u ; λ ) d t , K_n^{(p)}=\int_{\mathbb{R}}t^{n+2}\partial_u^2s^{(p)}(u;\lambda)\,dt=-\frac{n+1}{2}\int_{\mathbb{R}}t^n\partial_us^{(p)}(u;\lambda)\,dt=\frac{(n+1)(n-1)}{4}\int_{\mathbb{R}}t^{n-2}s^{(p)}(u;\lambda)\,dt, K n ( p ) = ∫ R t n + 2 ∂ u 2 s ( p ) ( u ; λ ) d t = − 2 n + 1 ∫ R t n ∂ u s ( p ) ( u ; λ ) d t = 4 ( n + 1 ) ( n − 1 ) ∫ R t n − 2 s ( p ) ( u ; λ ) d t ,
so that
K n ( p ) = ( n + 1 ) ( n − 1 ) 4 I n − 2 ( p ) ( n = 2 , 4 , 6 , … ) , ∣ λ ∣ K 0 ( p ) = 1 2 N ( p ) ( λ ≠ 0 ) . (68) K_n^{(p)}=\frac{(n+1)(n-1)}{4}I_{n-2}^{(p)}\quad(n=2,4,6,\ldots),\qquad|\lambda|K_0^{(p)}=\frac{1}{2}\mathcal{N}^{(p)}\quad(\lambda\ne0). \tag*{(68)} K n ( p ) = 4 ( n + 1 ) ( n − 1 ) I n − 2 ( p ) ( n = 2 , 4 , 6 , … ) , ∣ λ ∣ K 0 ( p ) = 2 1 N ( p ) ( λ = 0 ) . (68)
In particular, K 2 ( p ) = 3 4 I 0 ( p ) K_{2}^{(p)}=\frac{3}{4}I_{0}^{(p)} K 2 ( p ) = 4 3 I 0 ( p ) and K 4 ( p ) = 15 4 I 2 ( p ) K_{4}^{(p)}=\frac{15}{4}I_{2}^{(p)} K 4 ( p ) = 4 15 I 2 ( p ) . The relation for K 0 ( p ) K_{0}^{(p)} K 0 ( p ) follows from a single integration by parts: K 0 ( p ) = ∫ R t 2 ∂ u 2 s ( p ) ( u ; λ ) d t = − 1 2 ∫ R ∂ u s ( p ) ( u ; λ ) d t = 1 2 ∫ R Q ( p ) ( u ; λ ) u − 1 d t K_{0}^{(p)}=\int_{\mathbb{R}}t^{2}\partial_{u}^{2}s^{(p)}(u;\lambda)\,\mathrm{d}t=-\frac{1}{2}\int_{\mathbb{R}}\partial_{u}s^{(p)}(u;\lambda)\,\mathrm{d}t=\frac{1}{2}\int_{\mathbb{R}}Q^{(p)}(u;\lambda)u^{-1}\,\mathrm{d}t K 0 ( p ) = ∫ R t 2 ∂ u 2 s ( p ) ( u ; λ ) d t = − 2 1 ∫ R ∂ u s ( p ) ( u ; λ ) d t = 2 1 ∫ R Q ( p ) ( u ; λ ) u − 1 d t by (20) , which is N ( p ) / ( 2 ∣ λ ∣ ) \mathcal{N}^{(p)}/(2|\lambda|) N ( p ) / ( 2∣ λ ∣ ) by (63) . Thus the conditions (14) cancel all stresslet moments except K 0 ( p ) K_{0}^{(p)} K 0 ( p ) .
Since q − q 0 q-q_{0} q − q 0 vanishes at η = 0 \eta=0 η = 0 , the even part of w H w^{H} w H begins with a multiple of η 2 \eta^{2} η 2 , and this term pairs with K 0 ( p ) K_{0}^{(p)} K 0 ( p ) . To compute its coefficient, we expand the numerator through second order in σ \sigma σ , before multiplying by J \mathcal{J} J :
− b 3 n 0 ( n 0 ⋅ q s ) σ + { b 2 [ n 0 ( τ 0 ⋅ q s ) + τ 0 ( n 0 ⋅ q s ) ] − 1 2 b 3 n 0 ( n 0 ⋅ q s s ) } σ 2 + O ( σ 3 ) , (69) -b^{3}n_{0}(n_{0}\cdot q_{s})\sigma+\left\{b^{2}\left[n_{0}(\tau_{0}\cdot q_{s})+\tau_{0}(n_{0}\cdot q_{s})\right]-\frac{1}{2}b^{3}n_{0}(n_{0}\cdot q_{ss})\right\}\sigma^{2}+O(\sigma^{3}), \tag*{(69)} − b 3 n 0 ( n 0 ⋅ q s ) σ + { b 2 [ n 0 ( τ 0 ⋅ q s ) + τ 0 ( n 0 ⋅ q s ) ] − 2 1 b 3 n 0 ( n 0 ⋅ q ss ) } σ 2 + O ( σ 3 ) , (69)
using (48) . The first term is odd in σ \sigma σ and the second is even, but after the change of variable the first term also contributes to the even part. Substituting the series for σ \sigma σ in (49) and its derivative into (69) and collecting the coefficient of η 2 \eta^{2} η 2 gives b 2 V b^{2}\mathcal{V} b 2 V , with the vector V \mathcal{V} V of (37) . The last term of V \mathcal{V} V comes from the η 2 \eta^{2} η 2 term of σ \sigma σ acting on the first term of (69) . After η = δ t \eta=\delta t η = δ t , the term b 2 V η 2 b^{2}\mathcal{V}\eta^{2} b 2 V η 2 contributes − π − 1 δ − 1 b 2 K 0 ( p ) V = − ∣ b ∣ N ( p ) ( λ ) V / ( 2 π ) -\pi^{-1}\delta^{-1}b^{2}K_{0}^{(p)}\mathcal{V}=-|b|\mathcal{N}^{(p)}(\lambda)\mathcal{V}/(2\pi) − π − 1 δ − 1 b 2 K 0 ( p ) V = − ∣ b ∣ N ( p ) ( λ ) V / ( 2 π ) to the error, by (68) . Subtracting it gives the correction in (36) , which thus involves the same function N ( p ) \mathcal{N}^{(p)} N ( p ) as the Stokeslet correction (34) .
Proofs of the error estimates
We now prove Propositions 2.1, 3.2, and 3.3.
Proof of Proposition 2.1. Expand w e v w_{\mathrm{ev}} w ev through degree p − 3 p-3 p − 3 and w D , e v w_{D,\mathrm{ev}} w D , ev through degree p − 1 p-1 p − 1 , with remainder. After the integrals are extended to R \mathbb{R} R , which changes them by an exponentially small amount, the terms of the Taylor polynomials integrate to zero by (14) and Lemma 4.1; the coefficients (16) –(18) satisfy (14) by construction (Section 4.2). The remainders are bounded uniformly by C ∣ η ∣ p − 1 C|\eta|^{p-1} C ∣ η ∣ p − 1 for the single layer and by C ∣ η ∣ p + 1 C|\eta|^{p+1} C ∣ η ∣ p + 1 for the double layer, whose kernel contributes the factor ( η 2 + b 2 ) − 1 (\eta^{2}+b^{2})^{-1} ( η 2 + b 2 ) − 1 . After η = δ t \eta=\delta t η = δ t , both carry the prefactor δ p \delta^{p} δ p , multiplying
∫ R ∣ t ∣ p − 1 ∣ s ( p ) ( t 2 + λ 2 ; λ ) ∣ d t and ∫ R ∣ t ∣ p + 1 t 2 + λ 2 ∣ Q ( p ) ( t 2 + λ 2 ; λ ) ∣ d t ≤ ∫ R ∣ t ∣ p − 1 ∣ Q ( p ) ( t 2 + λ 2 ; λ ) ∣ d t , \int_{\mathbb{R}}|t|^{p-1}|s^{(p)}(t^{2}+\lambda^{2};\lambda)|\,\mathrm{d}t
\quad\text{and}\quad
\int_{\mathbb{R}}\frac{|t|^{p+1}}{t^{2}+\lambda^{2}}|Q^{(p)}(t^{2}+\lambda^{2};\lambda)|\,\mathrm{d}t
\le
\int_{\mathbb{R}}|t|^{p-1}|Q^{(p)}(t^{2}+\lambda^{2};\lambda)|\,\mathrm{d}t, ∫ R ∣ t ∣ p − 1 ∣ s ( p ) ( t 2 + λ 2 ; λ ) ∣ d t and ∫ R t 2 + λ 2 ∣ t ∣ p + 1 ∣ Q ( p ) ( t 2 + λ 2 ; λ ) ∣ d t ≤ ∫ R ∣ t ∣ p − 1 ∣ Q ( p ) ( t 2 + λ 2 ; λ ) ∣ d t ,
and these integrals are bounded uniformly for ∣ λ ∣ ≤ Λ |\lambda|\le\Lambda ∣ λ ∣ ≤ Λ : s ( p ) s^{(p)} s ( p ) has at most a logarithmic singularity, and Q ( p ) Q^{(p)} Q ( p ) is a Gaussian times a polynomial with coefficients bounded on [ − Λ , Λ ] [-\Lambda,\Lambda] [ − Λ , Λ ] , as noted in Section 2.3. The part of the curve outside the coordinate patch contributes an exponentially small amount. □ \square □
Proof of Proposition 3.2. We localize the integrals as in the proof of Proposition 2.1. The terms involving s ( p ) s^{(p)} s ( p ) vanish by (14) . The other Stokeslet moments are the J n ( p ) J_{n}^{(p)} J n ( p ) , and the stresslet moments K n ( p ) K_{n}^{(p)} K n ( p ) with n ≥ 2 n\ge2 n ≥ 2 are multiples of the I n − 2 ( p ) I_{n-2}^{(p)} I n − 2 ( p ) by (68) ; these also vanish. The terms with N ( p ) \mathcal{N}^{(p)} N ( p ) and K 0 ( p ) K_{0}^{(p)} K 0 ( p ) are removed by the local corrections in (34) and (36) . The Taylor remainders are bounded by C ∣ η ∣ p − 1 C|\eta|^{p-1} C ∣ η ∣ p − 1 , C ∣ η ∣ p + 1 C|\eta|^{p+1} C ∣ η ∣ p + 1 , and C ∣ η ∣ p + 3 C|\eta|^{p+3} C ∣ η ∣ p + 3 in the scalar, Stokeslet tensor, and stresslet tensor families, respectively, and η = δ t \eta=\delta t η = δ t leaves δ p \delta^{p} δ p times integrals bounded uniformly in λ \lambda λ , since ∣ t ∣ p + 1 / ( t 2 + λ 2 ) ≤ ∣ t ∣ p − 1 |t|^{p+1}/(t^{2}+\lambda^{2})\le|t|^{p-1} ∣ t ∣ p + 1 / ( t 2 + λ 2 ) ≤ ∣ t ∣ p − 1 and ∣ t ∣ p + 3 / ( t 2 + λ 2 ) 2 ≤ ∣ t ∣ p − 1 |t|^{p+3}/(t^{2}+\lambda^{2})^{2}\le|t|^{p-1} ∣ t ∣ p + 3 / ( t 2 + λ 2 ) 2 ≤ ∣ t ∣ p − 1 . Finally, N ( p ) \mathcal{N}^{(p)} N ( p ) and V \mathcal{V} V are bounded, so the corrections are O ( ∣ b ∣ ) O(|b|) O ( ∣ b ∣ ) , vanish at b = 0 b=0 b = 0 , and the estimate also holds on the curve. □ \square □
Proof of Proposition 3.3. By (44) and (45) , p δ − p = ( S δ ( p ) − S ) ϕ − ( D δ ( p ) − D ) ψ p_{\delta}-p=(S_{\delta}^{(p)}-S)\phi-(D_{\delta}^{(p)}-D)\psi p δ − p = ( S δ ( p ) − S ) ϕ − ( D δ ( p ) − D ) ψ . Proposition 2.1 applies to the two terms directly. □ \square □
The discretization error
The regularized integrals are computed with the periodic trapezoidal rule, whose error is small when δ / h \delta/h δ / h is large. Following [12 , 13 ] , we bound this error to see how δ \delta δ should depend on h h h . Let x ( t ) x(t) x ( t ) be a smooth 2 π 2\pi 2 π -periodic parametrization of Γ \Gamma Γ with min t ∣ x ′ ( t ) ∣ > 0 \min_t \lvert x'(t)\rvert>0 min t ∣ x ′ ( t )∣ > 0 , and let the densities be smooth. Let the targets lie in a band around Γ \Gamma Γ in which closest points are unique, with ∣ b ∣ ≤ Λ δ \lvert b\rvert\leq\Lambda\delta ∣ b ∣ ≤ Λ δ and 1 + b κ 0 1+b\kappa_0 1 + b κ 0 bounded below by a positive constant. Let F δ , y ( t ) F_{\delta,y}(t) F δ , y ( t ) be any scalar component of an integrand, and h = 2 π / N Γ h=2\pi/N_\Gamma h = 2 π / N Γ . For each integer ℓ ≥ 2 \ell\geq2 ℓ ≥ 2 ,
∣ h ∑ j = 0 N Γ − 1 F δ , y ( j h ) − ∫ 0 2 π F δ , y ( t ) d t ∣ ≤ C ℓ h ℓ ∥ ∂ t ℓ F δ , y ∥ L 1 ( 0 , 2 π ) , \left\lvert h\sum_{j=0}^{N_\Gamma-1}F_{\delta,y}(jh)-\int_0^{2\pi}F_{\delta,y}(t)\,\mathrm{d}t\right\rvert\leq C_\ell h^\ell\left\|\partial_t^\ell F_{\delta,y}\right\|_{L^1(0,2\pi)}, h j = 0 ∑ N Γ − 1 F δ , y ( j h ) − ∫ 0 2 π F δ , y ( t ) d t ≤ C ℓ h ℓ ∂ t ℓ F δ , y L 1 ( 0 , 2 π ) ,
since the error of the rule is 2 π 2\pi 2 π times the sum of the Fourier coefficients of F δ , y F_{\delta,y} F δ , y at the nonzero multiples of N Γ N_\Gamma N Γ , and integration by parts bounds the coefficient at k k k by ∣ k ∣ − ℓ ∥ ∂ t ℓ F δ , y ∥ L 1 / ( 2 π ) \lvert k\rvert^{-\ell}\left\|\partial_t^\ell F_{\delta,y}\right\|_{L^1}/(2\pi) ∣ k ∣ − ℓ ∂ t ℓ F δ , y L 1 / ( 2 π ) .
A crude bound on the derivatives is
∥ ∂ t ℓ F δ , y ∥ L 1 ≤ C ℓ ( 1 + ∣ log δ ∣ ) δ − ℓ − 1 , 0 < δ ≤ 1 , \left\|\partial_t^\ell F_{\delta,y}\right\|_{L^1}\leq C_\ell(1+\lvert\log\delta\rvert)\delta^{-\ell-1},\qquad0<\delta\leq1, ∂ t ℓ F δ , y L 1 ≤ C ℓ ( 1 + ∣ log δ ∣) δ − ℓ − 1 , 0 < δ ≤ 1 ,
uniformly in x 0 x_0 x 0 and ∣ λ ∣ ≤ Λ \lvert\lambda\rvert\leq\Lambda ∣ λ ∣ ≤ Λ . With the coefficients held fixed in t t t , the single-layer and Stokeslet kernels are smooth functions of ( x − y ) / δ (x-y)/\delta ( x − y ) / δ plus a multiple of log δ \log\delta log δ , so they are O ( 1 + ∣ log δ ∣ ) O(1+\lvert\log\delta\rvert) O ( 1 + ∣ log δ ∣) on Γ \Gamma Γ , and each t t t -derivative brings a factor δ − 1 \delta^{-1} δ − 1 . The double-layer and stresslet kernels are δ − 1 \delta^{-1} δ − 1 times smooth bounded functions of ( x − y ) / δ (x-y)/\delta ( x − y ) / δ , since d ( p ) = O ( ξ ) d^{(p)}=O(\xi) d ( p ) = O ( ξ ) and d ~ ( p ) = O ( ξ 2 ) \tilde d^{(p)}=O(\xi^2) d ~ ( p ) = O ( ξ 2 ) at ξ = 0 \xi=0 ξ = 0 and 1 − d ( p ) 1-d^{(p)} 1 − d ( p ) and 1 − d ~ ( p ) 1-\tilde d^{(p)} 1 − d ~ ( p ) decay exponentially. The chain and product rules give the bound. The pressure uses the same kernels, with the densities ϕ \phi ϕ and ψ \psi ψ .
Let V V V be any of the potentials and V δ , h V_{\delta,h} V δ , h its trapezoidal approximation, including the local corrections and the term χ g 0 \chi g_0 χ g 0 , which involve no quadrature. With exact data and arithmetic, Propositions 2.1, 3.2, and 3.3 and the two bounds above give
∣ V δ , h − V ∣ ≤ C δ p + C ℓ h ℓ δ − ℓ − 1 ( 1 + ∣ log δ ∣ ) . \lvert V_{\delta,h}-V\rvert\leq C\delta^p+C_\ell h^\ell\delta^{-\ell-1}(1+\lvert\log\delta\rvert). ∣ V δ , h − V ∣ ≤ C δ p + C ℓ h ℓ δ − ℓ − 1 ( 1 + ∣ log δ ∣) .
For δ = c h q \delta=ch^q δ = c h q with 0 < q < 1 0<q<1 0 < q < 1 , the second term is O ( h ℓ ( 1 − q ) − q ∣ log h ∣ ) O(h^{\ell(1-q)-q}\lvert\log h\rvert) O ( h ℓ ( 1 − q ) − q ∣ log h ∣) , which is o ( h p q ) o(h^{pq}) o ( h pq ) if
ℓ ( 1 − q ) > q ( p + 1 ) . (70) \ell(1-q)>q(p+1). \tag*{(70)} ℓ ( 1 − q ) > q ( p + 1 ) . (70)
Since ℓ \ell ℓ is arbitrary for smooth data, the error is O ( h p q ) O(h^{pq}) O ( h pq ) for every q < 1 q<1 q < 1 . When δ \delta δ is proportional to h h h , the second term is O ( h − 1 ∣ log h ∣ ) O(h^{-1}\lvert\log h\rvert) O ( h − 1 ∣ log h ∣) for every ℓ \ell ℓ , and the bound does not imply convergence.
Numerical examples
We test the regularized kernels on the unit circle, a kite, and a starfish (the kite and starfish are shown in Figure 9 ). In this section, target points y = x ( t 0 ) + b n ( t 0 ) y=x(t_0)+bn(t_0) y = x ( t 0 ) + bn ( t 0 ) are specified directly by the closest-point parameter t 0 t_0 t 0 and signed normal distance b b b . On the Cartesian grids of § 8 , t 0 t_0 t 0 is found by Newton’s method applied to ( x ( t ) − y ) ⋅ x ′ ( t ) = 0 (x(t)-y)\cdot x'(t)=0 ( x ( t ) − y ) ⋅ x ′ ( t ) = 0 , starting from the closest of 2 12 2^{12} 2 12 sample points on Γ \Gamma Γ , with b = ( y − x 0 ) ⋅ n 0 b=(y-x_0)\cdot n_0 b = ( y − x 0 ) ⋅ n 0 (cf. [8 Sect. 4] ). The curvature κ \kappa κ and its derivative κ s \kappa_s κ s are evaluated from the parametrization of Γ \Gamma Γ .
Figure 9. log 10 \log_{10} log 10 of the error on the Cartesian grid at M = 2 10 M = 2^{10} M = 2 10 : (a) Laplace solution for the kite; (b) pressure and (c) velocity for the starfish.
Unless stated otherwise, the reported error is the maximum over the 97 parameters t 0 = 2 π ( j + 0.37 ) / 97 t_0=2\pi(j+0.37)/97 t 0 = 2 π ( j + 0.37 ) /97 (j = 0 , … , 96 j=0,\ldots,96 j = 0 , … , 96 ) and b / δ = − 1 / 2 , 0 , 1 / 2 b/\delta=-1/2,0,1/2 b / δ = − 1/2 , 0 , 1/2 . Targets on the curve (b = 0 b=0 b = 0 ) use the regularized kernels with λ = 0 \lambda=0 λ = 0 so that a single formula covers all targets, although on Γ \Gamma Γ the double-layer and stresslet kernels are smooth and need no regularization (Remarks 2.2 and 3.4). Table 1 summarizes the convergence rates for the main tests. All computations were done in double precision in Python with NumPy and SciPy [49 ] on a MacBook Air with an Apple M3 chip and 8 GB of memory.
Table 1: Fitted convergence rates in h h h (top) and in Δ s \Delta s Δ s (bottom). The fits use N Γ = 2 11 N_{\Gamma}=2^{11} N Γ = 2 11 –2 14 2^{14} 2 14 for the circle with p = 3 , 5 p=3,5 p = 3 , 5 and 2 9 2^9 2 9 –2 12 2^{12} 2 12 with p = 7 p=7 p = 7 ; 2 12 2^{12} 2 12 –2 14 2^{14} 2 14 for the kite and the starfish with p = 3 , 5 p=3,5 p = 3 , 5 and 2 13 2^{13} 2 13 –2 15 2^{15} 2 15 with p = 7 p=7 p = 7 ; and 2 12 2^{12} 2 12 –2 14 2^{14} 2 14 in the bottom block, except 2 11 2^{11} 2 11 –2 13 2^{13} 2 13 for the starfish pressure.
Table 1. Fitted convergence rates in h h h (top) and in Δ s \Delta s Δ s (bottom). The fits use N Γ = 2 11 N_{\Gamma}=2^{11} N Γ = 2 11 –2 14 2^{14} 2 14 for the circle with p = 3 , 5 p=3,5 p = 3 , 5 and 2 9 2^9 2 9 –2 12 2^{12} 2 12 with p = 7 p=7 p = 7 ; 2 12 2^{12} 2 12 –2 14 2^{14} 2 14 for the kite and the starfish with p = 3 , 5 p=3,5 p = 3 , 5 and 2 13 2^{13} 2 13 –2 15 2^{15} 2 15 with p = 7 p=7 p = 7 ; and 2 12 2^{12} 2 12 –2 14 2^{14} 2 14 in the bottom block, except 2 11 2^{11} 2 11 –2 13 2^{13} 2 13 for the starfish pressure.
Two choices of δ \delta δ
We use two choices of δ \delta δ . With the parameter spacing h = 2 π / N Γ h = 2\pi/N_{\Gamma} h = 2 π / N Γ , the first is
δ = c h q , c = 1 , ( p , q ) = ( 3 , 2 / 3 ) , ( 5 , 4 / 5 ) , ( 7 , 5 / 7 ) , (71) \delta= ch^q,\qquad c = 1,\qquad(p,q) = (3,2/3),(5,4/5),(7,5/7), \tag*{(71)} δ = c h q , c = 1 , ( p , q ) = ( 3 , 2/3 ) , ( 5 , 4/5 ) , ( 7 , 5/7 ) , (71)
so the regularization error is O ( h p q ) O(h^{pq}) O ( h pq ) and the predicted orders are p q = 2 , 4 , 5 pq = 2,4,5 pq = 2 , 4 , 5 . Since q < 1 q < 1 q < 1 , δ / h → ∞ \delta/h \to\infty δ / h → ∞ , and (70) holds with ℓ = 9 , 25 , 21 \ell= 9,25,21 ℓ = 9 , 25 , 21 ; for smooth curves and densities the discretization error is therefore asymptotically smaller than the regularization error. Throughout, we estimate the discretization error by the change when N Γ N_{\Gamma} N Γ is doubled with δ \delta δ and the targets fixed.
The second choice, as in [13 ] , takes δ \delta δ proportional to the mesh spacing: δ = c h \delta= ch δ = c h for the circle, and δ = c Δ s \delta= c\Delta s δ = c Δ s for the kite and the starfish, whose nodes are not equally spaced along the curve. Here
Δ s : = max j ∣ x ( t j + 1 ) − x ( t j ) ∣ , t j = j h , (72) \Delta s := \max_j \lvert x(t_{j+1}) - x(t_j)\rvert,\qquad t_j = jh, \tag*{(72)} Δ s := j max ∣ x ( t j + 1 ) − x ( t j )∣ , t j = j h , (72)
is the largest distance between adjacent nodes, and rates are fitted against Δ s \Delta s Δ s . The regularization error is O ( h p ) O(h^p) O ( h p ) , but the estimate of § 4.7 does not imply convergence, so the discretization error need not become negligible. We therefore fit rates only over ranges where the estimated discretization error is well below the total error. For this choice we also maximize the error over closest points at nodes and halfway between nodes, and for the circle also a quarter of the way.
The Laplace equation
Single- and double-layer potentials for a circle
We first test the single and double layers separately, with Γ \Gamma Γ the unit circle and the density
f ( t ) = 1 + 0.4 cos 3 t + 0.3 sin 5 t , x ( t ) = ( cos t , sin t ) . (73) f(t) = 1 + 0.4\cos3t + 0.3\sin5t,\qquad x(t) = (\cos t,\sin t). \tag*{(73)} f ( t ) = 1 + 0.4 cos 3 t + 0.3 sin 5 t , x ( t ) = ( cos t , sin t ) . (73)
For y = R ( cos θ , sin θ ) y = R(\cos\theta,\sin\theta) y = R ( cos θ , sin θ ) and G = + log r / ( 2 π ) G = +\log r/(2\pi) G = + log r / ( 2 π ) ,
S f ( y ) = log max ( R , 1 ) − 0.4 6 min ( R 3 , R − 3 ) cos 3 θ − 0.3 10 min ( R 5 , R − 5 ) sin 5 θ , (74) Sf(y) = \log\max(R,1) - \frac{0.4}{6}\min(R^3,R^{-3})\cos3\theta- \frac{0.3}{10}\min(R^5,R^{-5})\sin5\theta, \tag*{(74)} S f ( y ) = log max ( R , 1 ) − 6 0.4 min ( R 3 , R − 3 ) cos 3 θ − 10 0.3 min ( R 5 , R − 5 ) sin 5 θ , (74)
which is continuous at R = 1 R = 1 R = 1 . The double layer is
D f ( y ) = { 1 + 0.2 R 3 cos 3 θ + 0.15 R 5 sin 5 θ , R < 1 , − 0.2 R − 3 cos 3 θ − 0.15 R − 5 sin 5 θ , R > 1 , K f = 1 2 . (75) Df(y) =
\begin{cases}
1 + 0.2R^3\cos3\theta+ 0.15R^5\sin5\theta, & R < 1,\\
-0.2R^{-3}\cos3\theta- 0.15R^{-5}\sin5\theta, & R > 1,
\end{cases}
\qquad Kf = \frac{1}{2}. \tag*{(75)} D f ( y ) = { 1 + 0.2 R 3 cos 3 θ + 0.15 R 5 sin 5 θ , − 0.2 R − 3 cos 3 θ − 0.15 R − 5 sin 5 θ , R < 1 , R > 1 , K f = 2 1 . (75)
These formulas follow mode by mode from the Fourier expansion of log ∣ x − y ∣ \log\lvert x-y\rvert log ∣ x − y ∣ and satisfy D − f − D + f = f D^-f-D^+f=f D − f − D + f = f .
For p = 3 , 5 , 7 p = 3,5,7 p = 3 , 5 , 7 , both layers converge at the predicted orders 2 , 4 , 5 2,4,5 2 , 4 , 5 (Table 1 , Figure 1 (a),(Figure 1 b)). With δ = c h \delta= ch δ = c h , c = 2 , 3 , 4 c = 2,3,4 c = 2 , 3 , 4 , both seventh-order layers follow the reference slope seven until they reach roundoff (Figure 1 (c),(Figure 1 d)).
Figure 1. Single layer (a),(c) and double layer (b),(d) for the unit circle: (a),(b) δ = h q \delta=h^q δ = h q ; (c),(d) p = 7 p=7 p = 7 , δ = c h \delta=ch δ = c h .
A Laplace solution for a kite
Next we test the sum S f + D g Sf + Dg S f + D g of single and double layers, with densities on the kite-shaped curve Γ L \Gamma_L Γ L ,
x L ( t ) = s K ( cos t + 0.65 cos 2 t − 0.65 , 1.5 sin t ) , s K = 179 300 . (76) x_L(t) = s_K(\cos t + 0.65\cos2t - 0.65,\ 1.5\sin t),\qquad s_K = \frac{179}{300}. \tag*{(76)} x L ( t ) = s K ( cos t + 0.65 cos 2 t − 0.65 , 1.5 sin t ) , s K = 300 179 . (76)
With this scaling, max t ∥ x L ( t ) ∥ ∞ = 0.895 \max_t \lVert x_L(t)\rVert_\infty= 0.895 max t ∥ x L ( t ) ∥ ∞ = 0.895 . Let z = x 1 + i x 2 z=x_1+ix_2 z = x 1 + i x 2 and set v ∓ = Re F ∓ ( z ) v^\mp=\operatorname{Re} F^\mp(z) v ∓ = Re F ∓ ( z ) , where
F − ( z ) = 0.12 + 0.20 z + ( 0.07 − 0.04 i ) z 2 + 0.10 z − ζ 1 − 0.08 i z − ζ 2 , (77) F^-(z)=0.12+0.20z+(0.07-0.04i)z^2+\frac{0.10}{z-\zeta_1}-\frac{0.08i}{z-\zeta_2}, \tag*{(77)} F − ( z ) = 0.12 + 0.20 z + ( 0.07 − 0.04 i ) z 2 + z − ζ 1 0.10 − z − ζ 2 0.08 i , (77)
F + ( z ) = 0.16 + 0.03 i z − z 1 − 0.07 − 0.04 i z − z 2 , (78) F^+(z)=\frac{0.16+0.03i}{z-z_1}-\frac{0.07-0.04i}{z-z_2}, \tag*{(78)} F + ( z ) = z − z 1 0.16 + 0.03 i − z − z 2 0.07 − 0.04 i , (78)
with
ζ 1 = 2.40 + 1.10 i , ζ 2 = − 2.50 + 0.80 i , \zeta_1=2.40+1.10i,\qquad\zeta_2=-2.50+0.80i, ζ 1 = 2.40 + 1.10 i , ζ 2 = − 2.50 + 0.80 i ,
z 1 = 0.08 + 0.10 i , z 2 = − 0.22 − 0.08 i . z_1=0.08+0.10i,\qquad z_2=-0.22-0.08i. z 1 = 0.08 + 0.10 i , z 2 = − 0.22 − 0.08 i .
The poles of F − F^- F − lie outside the kite and those of F + F^+ F + lie inside, so v − v^- v − is harmonic inside and v + v^+ v + is harmonic outside. Since ∇ Re F = ( Re F ′ , − Im F ′ ) \nabla\operatorname{Re}F=(\operatorname{Re}F',-\operatorname{Im}F') ∇ Re F = ( Re F ′ , − Im F ′ ) , the normal derivatives are known exactly. With
f = ∂ n v + − ∂ n v − , g = v − − v + , (79) f=\partial_n v^+-\partial_n v^-,\qquad g=v^--v^+, \tag*{(79)} f = ∂ n v + − ∂ n v − , g = v − − v + , (79)
the jump relations give S f + D g = v − Sf+Dg=v^- S f + D g = v − inside, v + v^+ v + outside, and the average ( v − + v + ) / 2 (v^-+v^+)/2 ( v − + v + ) /2 on Γ L \Gamma_L Γ L .
With δ \delta δ as in (71) , the fitted rates are close to the predicted orders, and with the seventh-order kernel and δ = c Δ s \delta=c\Delta s δ = c Δ s , c = 2 , 3 , 4 c=2,3,4 c = 2 , 3 , 4 , they are close to seven (Table 1 , Figure 2 (a),(Figure 2 b)). Figure 2 (c) shows the seventh-order errors separately for ∣ λ ∣ = 0 , 0.5 , 1 , 2 |\lambda|=0,0.5,1,2 ∣ λ ∣ = 0 , 0.5 , 1 , 2 ; for λ ≠ 0 \lambda\ne0 λ = 0 we report the larger of the errors on the two sides. On the finer meshes, the errors decrease as ∣ λ ∣ |\lambda| ∣ λ ∣ increases; at ∣ λ ∣ = 2 |\lambda|=2 ∣ λ ∣ = 2 the error on the finest mesh is at roundoff.
Figure 2. Laplace solution for the kite: (a) δ = h q \delta= h^q δ = h q ; (b) p = 7 p = 7 p = 7 , δ = c Δ s \delta= c\Delta s δ = c Δ s ; (c) p = 7 p = 7 p = 7 , δ = h 5 / 7 \delta= h^{5/7} δ = h 5/7 , for several ∣ λ ∣ |\lambda| ∣ λ ∣ .
Stokes flow
A Stokes solution for a starfish
For the Stokes test, we compute the velocity from the sum of the corrected single and double layers, and the pressure from (45) , for the starfish Γ S \Gamma_S Γ S ,
x S ( t ) = s S ( 1 + 0.3 cos 5 t ) ( cos t , sin t ) , s S = 179 260 . (80) x_S(t)=s_S(1+0.3\cos5t)(\cos t,\sin t),\qquad s_S=\frac{179}{260}. \tag*{(80)} x S ( t ) = s S ( 1 + 0.3 cos 5 t ) ( cos t , sin t ) , s S = 260 179 . (80)
Again max t ∥ x S ( t ) ∥ ∞ = 0.895 \max_t\lVert x_S(t)\rVert_\infty=0.895 max t ∥ x S ( t ) ∥ ∞ = 0.895 . For μ = 1 \mu=1 μ = 1 , define the interior stream function
Ψ − = Re { z ˉ F ( z ) + W ( z ) } , u − = ( Ψ x 2 − , − Ψ x 1 − ) , p − = − 4 Im F ′ ( z ) , (81) \Psi^-=\operatorname{Re}\{\bar{z}F(z)+W(z)\},\qquad u^-=(\Psi^-_{x_2},-\Psi^-_{x_1}),\qquad p^-=-4\operatorname{Im}F'(z), \tag*{(81)} Ψ − = Re { z ˉ F ( z ) + W ( z )} , u − = ( Ψ x 2 − , − Ψ x 1 − ) , p − = − 4 Im F ′ ( z ) , (81)
with u + = 0 u^+=0 u + = 0 and p + = 0 p^+=0 p + = 0 . The analytic functions are
F ( z ) = 0.04 + ( 0.055 + 0.025 i ) z + ( 0.018 − 0.012 i ) z 2 + 0.035 z − ζ 1 − 0.025 i z − ζ 2 , (82) F(z)=0.04+(0.055+0.025i)z+(0.018-0.012i)z^2+\frac{0.035}{z-\zeta_1}-\frac{0.025i}{z-\zeta_2}, \tag*{(82)} F ( z ) = 0.04 + ( 0.055 + 0.025 i ) z + ( 0.018 − 0.012 i ) z 2 + z − ζ 1 0.035 − z − ζ 2 0.025 i , (82)
W ( z ) = − 0.03 i + ( 0.045 − 0.020 i ) z + ( 0.014 + 0.010 i ) z 2 + 0.030 i z − ζ 3 − 0.022 z − ζ 4 . (83) W(z)=-0.03i+(0.045-0.020i)z+(0.014+0.010i)z^2+\frac{0.030i}{z-\zeta_3}-\frac{0.022}{z-\zeta_4}. \tag*{(83)} W ( z ) = − 0.03 i + ( 0.045 − 0.020 i ) z + ( 0.014 + 0.010 i ) z 2 + z − ζ 3 0.030 i − z − ζ 4 0.022 . (83)
The poles are
ζ 1 = 2.50 + 1.00 i , ζ 2 = − 2.40 + 1.20 i , ζ 3 = 2.30 − 1.40 i , ζ 4 = − 2.60 − 0.90 i . \zeta_1=2.50+1.00i,\qquad\zeta_2=-2.40+1.20i,\qquad\zeta_3=2.30-1.40i,\qquad\zeta_4=-2.60-0.90i. ζ 1 = 2.50 + 1.00 i , ζ 2 = − 2.40 + 1.20 i , ζ 3 = 2.30 − 1.40 i , ζ 4 = − 2.60 − 0.90 i .
All poles lie outside the starfish, so F F F and W W W are analytic in the interior. With the stress Σ \Sigma Σ of § 3.1 , computed exactly from F F F and W W W , we prescribe on Γ S \Gamma_S Γ S
q = u − ∣ Γ S , f = Σ − n . (84) q=u^-\rvert_{\Gamma_S},\qquad f=\Sigma^-n. \tag*{(84)} q = u − ∣ Γ S , f = Σ − n . (84)
By the jump relations, the resulting velocity and pressure potentials equal u − u^{-} u − and p − p^{-} p − inside Γ S \Gamma_S Γ S , vanish outside, and equal u − / 2 u^{-}/2 u − /2 and p − / 2 p^{-}/2 p − /2 on Γ S \Gamma_S Γ S . The pressure densities ϕ \phi ϕ and ψ \psi ψ are computed by Fourier differentiation, as for general data.
We measure the maximum Euclidean norm of the velocity error and the maximum pressure error. For δ \delta δ as in (71), the velocity and pressure converge at close to the predicted orders (Table 1 , Figure 3 (a),(Figure 3 b)).
Figure 3. Stokes solution for the starfish: velocity (a),(c) and pressure (b),(d), with δ = h q \delta=h^q δ = h q in (a),(b) and p = 7 p=7 p = 7 , δ = c Δ s \delta=c\Delta s δ = c Δ s in (c),(d).
With the seventh-order kernel and δ = c Δ s \delta=c\Delta s δ = c Δ s , c = 2 , 3 , 4 c=2,3,4 c = 2 , 3 , 4 , the velocity error decreases at a rate close to seven (Table 1 , Figure 3 (c)). The pressure error decreases at a similar rate over N Γ = 2 11 N_{\Gamma}=2^{11} N Γ = 2 11 –2 13 2^{13} 2 13 (Table 1 , Figure 3 (d)); over this range the velocity rates are 6.75, 6.49, and 6.17. At N Γ = 2 14 N_{\Gamma}=2^{14} N Γ = 2 14 the pressure error levels off: it is limited by roundoff and by the Fourier differentiation of the pressure densities (§ 7.3 ), and its estimated discretization error is no longer well below the total error.
The Stokeslet and stresslet integrals and the local corrections
To verify the accuracy of each layer without possible error cancellation between them, we test the Stokeslet and stresslet integrals separately on the starfish. Exact integral values are unavailable, so we compare with numerical reference values, prescribing smooth densities independently as
q ( t ) = ( sin ( 2 t ) + 0.25 cos ( 5 t ) , 0.6 cos ( 3 t ) − 0.2 sin t ) , q(t) = \left(\sin(2t) + 0.25\cos(5t),\ 0.6\cos(3t) - 0.2\sin t\right), q ( t ) = ( sin ( 2 t ) + 0.25 cos ( 5 t ) , 0.6 cos ( 3 t ) − 0.2 sin t ) ,
f ( t ) = ( 0.7 cos ( 2 t ) + 0.15 sin ( 4 t ) , 0.5 sin ( 3 t ) + 0.2 cos t ) . f(t) = \left(0.7\cos(2t) + 0.15\sin(4t),\ 0.5\sin(3t) + 0.2\cos t\right). f ( t ) = ( 0.7 cos ( 2 t ) + 0.15 sin ( 4 t ) , 0.5 sin ( 3 t ) + 0.2 cos t ) .
The targets have the 49 closest-point parameters t 0 = 2 π ( j + 0.37 ) / 49 t_0 = 2\pi(j + 0.37)/49 t 0 = 2 π ( j + 0.37 ) /49 , j = 0 , … , 48 j = 0,\ldots,48 j = 0 , … , 48 , and b = ± δ / 2 b = \pm\delta/2 b = ± δ /2 . We halve δ \delta δ from 0.04 to 0.00125, with N Γ = max { 2 13 , 2 ⌈ log 2 ( 32 / δ ) ⌉ } N_{\Gamma} = \max\{2^{13}, 2^{\lceil\log_2(32/\delta)\rceil}\} N Γ = max { 2 13 , 2 ⌈ l o g 2 ( 32/ δ )⌉ } . Reference values are computed from the unregularized kernels by Gauss–Legendre quadrature on panels graded by bisection, with singularity subtraction for the stresslet; the 48-point and 64-point rules agree to within 10 − 14 10^{-14} 1 0 − 14 .
Without the local corrections, the errors of both integrals are only about first order in δ \delta δ (Figure 4 , dashed curves). With them, the rates between δ = 0.0025 \delta= 0.0025 δ = 0.0025 and 0.00125 are close to p p p : between 2.92 and 3.04 for p = 3 p = 3 p = 3 , 4.85 and 5.05 for p = 5 p = 5 p = 5 , and 6.72 and 7.06 for p = 7 p = 7 p = 7 , for both integrals and both sides of the curve. For the seventh-order Stokeslet integral, whose error at δ = 0.00125 \delta= 0.00125 δ = 0.00125 is at the level of the reference error, the rate is taken between δ = 0.005 \delta= 0.005 δ = 0.005 and 0.0025.
Figure 4. Stokeslet and stresslet integrals for the starfish at b = − δ / 2 b = -\delta/2 b = − δ /2 , with (solid) and without (dashed) the local corrections.
Differentiation of the pressure densities
For the exact starfish solution of § 7.1 , with the boundary data (84) , the pressure densities on Γ S \Gamma_S Γ S are also given by
ϕ = − ∂ n p − , ψ = − p − . \phi= -\partial_n p^{-},\qquad\psi= -p^{-}. ϕ = − ∂ n p − , ψ = − p − .
Indeed, incompressibility gives ψ = − p − \psi= -p^{-} ψ = − p − , while the Stokes equations give ϕ = μ ∂ s ω − = − ∂ n p − \phi= \mu\partial_s\omega^{-} = -\partial_n p^{-} ϕ = μ ∂ s ω − = − ∂ n p − , with ω − = ∂ x 1 u 2 − − ∂ x 2 u 1 − \omega^{-} = \partial_{x_1}u_2^{-} - \partial_{x_2}u_1^{-} ω − = ∂ x 1 u 2 − − ∂ x 2 u 1 − and the outward normal convention used here. At N Γ = 2 10 N_{\Gamma} = 2^{10} N Γ = 2 10 , the densities computed by Fourier differentiation satisfy the identities for ψ \psi ψ and ϕ \phi ϕ to within 10 − 13 10^{-13} 1 0 − 13 and 10 − 10 10^{-10} 1 0 − 10 , respectively; the residuals grow with N Γ N_{\Gamma} N Γ as roundoff in the differentiation is amplified. To see how this affects the pressure, we compute it once with the Fourier densities and once with the exact ones, with everything else the same. We use the 33 closest-point parameters 2 π j / 33 2\pi j/33 2 π j /33 , each also shifted by half a mesh interval, b / δ = − 1 / 2 , 0 , 1 / 2 b/\delta=-1/2,0,1/2 b / δ = − 1/2 , 0 , 1/2 , the seventh-order kernel, and δ = c Δ s \delta=c\Delta s δ = c Δ s with c = 2 , 4 c=2,4 c = 2 , 4 . (These targets differ from those of § 7.1 .) At N Γ = 2 14 N_{\Gamma}=2^{14} N Γ = 2 14 and δ = 2 Δ s \delta=2\Delta s δ = 2Δ s , the exact densities reduce the maximum pressure error by a factor of about 15, from about 10 − 13 10^{-13} 1 0 − 13 ; on coarser meshes or with c = 4 c=4 c = 4 , the regularization error dominates and the two computations give nearly the same error. Thus the smallest pressure errors are limited by the numerical differentiation of the densities.
Targets approaching the curve
The estimates of § 2.5 and § 3.2.1 are uniform for ∣ b ∣ ≤ Λ δ |b|\leq\Lambda\delta ∣ b ∣ ≤ Λ δ ; here we check that the computed values tend to their limits on the curve as b → 0 b\to0 b → 0 from either side. We use the 33 closest-point parameters t 0 = 2 π ( j + 0.37 ) / 33 t_0=2\pi(j+0.37)/33 t 0 = 2 π ( j + 0.37 ) /33 , j = 0 , … , 32 j=0,\ldots,32 j = 0 , … , 32 , δ = 0.01 \delta=0.01 δ = 0.01 , and N Γ = 2 13 N_{\Gamma}=2^{13} N Γ = 2 13 . With the seventh-order kernels and ∣ b ∣ / δ = 10 − 10 |b|/\delta=10^{-10} ∣ b ∣/ δ = 1 0 − 10 , the Stokeslet and stresslet integrals of § 7.2 differ from their values at b = 0 b=0 b = 0 by less than 10 − 11 10^{-11} 1 0 − 11 ; for the stresslet, the value at b = 0 b=0 b = 0 is computed with χ = 1 \chi=1 χ = 1 or 0, according to the side of the target. For the exact solutions on the kite and the starfish, with p = 7 p=7 p = 7 and ∣ b ∣ / δ = 0.5 |b|/\delta=0.5 ∣ b ∣/ δ = 0.5 , 10 − 2 10^{-2} 1 0 − 2 , 10 − 4 10^{-4} 1 0 − 4 , 10 − 6 10^{-6} 1 0 − 6 , 10 − 8 10^{-8} 1 0 − 8 , and 10 − 10 10^{-10} 1 0 − 10 on both sides, the errors approach those at b = 0 b=0 b = 0 as ∣ b ∣ |b| ∣ b ∣ decreases (Figure 5 ).
Figure 5. Maximum errors as targets approach the curve: Laplace solution for the kite, and Stokes velocity and pressure for the starfish.
Varying δ \delta δ with fixed nodes and targets
To study the choice of δ \delta δ separately from the target locations, we fix N Γ = 2 11 N_{\Gamma}=2^{11} N Γ = 2 11 and b = − 0.003 b=-0.003 b = − 0.003 for the Stokeslet and stresslet integrals of § 7.2 and vary δ / Δ s \delta/\Delta s δ /Δ s from 0.35 to 8 (Figure 6 ). For small δ \delta δ the estimated discretization error is comparable with the total error; for large δ \delta δ it is negligible, but the regularization error is larger. For these targets the smallest fifth- and seventh-order errors occur near δ / Δ s = 1.5 \delta/\Delta s=1.5 δ /Δ s = 1.5 , but there the estimated discretization error of the seventh-order Stokeslet integral is comparable with the total error. At δ / Δ s = 2 \delta/\Delta s=2 δ /Δ s = 2 , the estimated discretization errors of both integrals are at roundoff level. In practice, one can decrease δ \delta δ while monitoring the estimated discretization error; once it becomes significant, decreasing δ \delta δ further requires more nodes. The product δ ∥ κ ∥ ∞ \delta\lVert\kappa\rVert_{\infty} δ ∥ κ ∥ ∞ must also be small, as Proposition 2.1 assumes. The best ratio δ / Δ s \delta/\Delta s δ /Δ s depends on the geometry, the density, the order, and the target distance.
Figure 6. Stokeslet and stresslet integrals for the starfish against δ / Δ s \delta/\Delta s δ /Δ s : errors (solid) and estimated discretization errors (dashed).
Values on a Cartesian grid
In many applications the solution is needed on a whole grid rather than at a few targets, for example when the flow is coupled with an equation for a heat or solute concentration in the surrounding fluid, or when the layer potentials supply the interface data for a grid-based method. Evaluating the integrals at every grid point would be costly. Following Mayo [31 , 32 ] and [8 , 13 ] , we evaluate the regularized potentials directly only at grid points near Γ \Gamma Γ and on the outer boundary, and obtain the values elsewhere with a fast Poisson solver based on the discrete sine transform (DST).
Finite-difference discretization
Let h c h_c h c be the grid spacing of a square computational box and M M M the number of interior grid points in each direction. We discretize the Laplacian by the nine-point compact operator [25 ]
( L h c v ) i j = 1 6 h c 2 [ 4 ( v i + 1 , j + v i − 1 , j + v i , j + 1 + v i , j − 1 ) + v i + 1 , j + 1 + v i + 1 , j − 1 + v i − 1 , j + 1 + v i − 1 , j − 1 − 20 v i j ] . (85) (L_{h_c}v)_{ij} = \frac{1}{6h_c^2}\left[4(v_{i+1,j}+v_{i-1,j}+v_{i,j+1}+v_{i,j-1})+v_{i+1,j+1}+v_{i+1,j-1}+v_{i-1,j+1}+v_{i-1,j-1}-20v_{ij}\right]. \tag*{(85)} ( L h c v ) ij = 6 h c 2 1 [ 4 ( v i + 1 , j + v i − 1 , j + v i , j + 1 + v i , j − 1 ) + v i + 1 , j + 1 + v i + 1 , j − 1 + v i − 1 , j + 1 + v i − 1 , j − 1 − 20 v ij ] . (85)
For smooth v v v , Taylor expansion gives
L h c v = Δ v + h c 2 12 Δ 2 v + h c 4 360 ( ∂ x 1 6 v + 5 ∂ x 1 4 ∂ x 2 2 v + 5 ∂ x 1 2 ∂ x 2 4 v + ∂ x 2 6 v ) + O ( h c 6 ) . (86) L_{h_c}v = \Delta v+\frac{h_c^2}{12}\Delta^2v+\frac{h_c^4}{360}(\partial_{x_1}^6v+5\partial_{x_1}^4\partial_{x_2}^2v+5\partial_{x_1}^2\partial_{x_2}^4v+\partial_{x_2}^6v)+O(h_c^6). \tag*{(86)} L h c v = Δ v + 12 h c 2 Δ 2 v + 360 h c 4 ( ∂ x 1 6 v + 5 ∂ x 1 4 ∂ x 2 2 v + 5 ∂ x 1 2 ∂ x 2 4 v + ∂ x 2 6 v ) + O ( h c 6 ) . (86)
The operator is second order accurate for a general smooth function, and its truncation error is O ( h c 6 ) O(h_c^6) O ( h c 6 ) for a harmonic function. For the Stokes velocity, Δ u = μ − 1 ∇ p \Delta u=\mu^{-1}\nabla p Δ u = μ − 1 ∇ p and Δ p = 0 \Delta p=0 Δ p = 0 , hence Δ 2 u = 0 \Delta^2u=0 Δ 2 u = 0 , and the truncation error is O ( h c 4 ) O(h_c^4) O ( h c 4 ) .
We approximate the pressure gradient by the centered fourth-order formulas
( D x 1 ( 4 ) v ) i j = v i − 2 , j − 8 v i − 1 , j + 8 v i + 1 , j − v i + 2 , j 12 h c , (D_{x_1}^{(4)}v)_{ij}=\frac{v_{i-2,j}-8v_{i-1,j}+8v_{i+1,j}-v_{i+2,j}}{12h_c}, ( D x 1 ( 4 ) v ) ij = 12 h c v i − 2 , j − 8 v i − 1 , j + 8 v i + 1 , j − v i + 2 , j ,
( D x 2 ( 4 ) v ) i j = v i , j − 2 − 8 v i , j − 1 + 8 v i , j + 1 − v i , j + 2 12 h c . (87) (D_{x_2}^{(4)}v)_{ij}=\frac{v_{i,j-2}-8v_{i,j-1}+8v_{i,j+1}-v_{i,j+2}}{12h_c}. \tag*{(87)} ( D x 2 ( 4 ) v ) ij = 12 h c v i , j − 2 − 8 v i , j − 1 + 8 v i , j + 1 − v i , j + 2 . (87)
We write ∇ h c ( 4 ) = ( D x 1 ( 4 ) , D x 2 ( 4 ) ) \nabla_{h_c}^{(4)}=(D_{x_1}^{(4)},D_{x_2}^{(4)}) ∇ h c ( 4 ) = ( D x 1 ( 4 ) , D x 2 ( 4 ) ) , with fourth-order one-sided formulas next to the outer boundary.
Irregular points and directly evaluated values
Let W Δ ( x i j ) \mathcal{W}_{\Delta}(x_{ij}) W Δ ( x ij ) consist of x i j x_{ij} x ij , its four axial neighbors, and its four diagonal neighbors, as in (85), and let W ∇ ( x i j ) \mathcal{W}_{\nabla}(x_{ij}) W ∇ ( x ij ) consist of x i j x_{ij} x ij and the points at axial offsets ± h c , ± 2 h c \pm h_c,\pm2h_c ± h c , ± 2 h c , as in (87). A stencil crosses Γ \Gamma Γ if its points do not all lie on the same side of Γ \Gamma Γ ; whether a grid point is inside or outside Ω \Omega Ω is found by a point-in-polygon test. The sets of irregular points [9 , 30 ] are
C Δ = { x i j : W Δ ( x i j ) crosses Γ } , C ∇ = { x i j : W ∇ ( x i j ) crosses Γ } . (88) \mathcal{C}_{\Delta}=\{x_{ij}:\mathcal{W}_{\Delta}(x_{ij})\text{ crosses }\Gamma\},\qquad\mathcal{C}_{\nabla}=\{x_{ij}:\mathcal{W}_{\nabla}(x_{ij})\text{ crosses }\Gamma\}. \tag*{(88)} C Δ = { x ij : W Δ ( x ij ) crosses Γ } , C ∇ = { x ij : W ∇ ( x ij ) crosses Γ } . (88)
Grid points outside C Δ \mathcal{C}_{\Delta} C Δ are regular for the Laplace and pressure problems. For the Stokes velocity, the irregular points are
C u = C Δ ∪ C ∇ , (89) \mathcal{C}_u=\mathcal{C}_{\Delta}\cup\mathcal{C}_{\nabla}, \tag*{(89)} C u = C Δ ∪ C ∇ , (89)
since the pressure-gradient stencil can cross Γ \Gamma Γ when the nine-point stencil does not (Figure 7 ).
Figure 7. Irregular points for the starfish at M = 2 7 M=2^7 M = 2 7 : the set C Δ \mathcal{C}_{\Delta} C Δ and the additional points of C ∇ ∖ C Δ \mathcal{C}_{\nabla}\setminus\mathcal{C}_{\Delta} C ∇ ∖ C Δ , where only the pressure-gradient stencil crosses Γ \Gamma Γ .
The layer potentials are evaluated directly at the irregular points and at all points of their nine-point stencils, that is, on
E L = ⋃ x ∈ C Δ W Δ ( x ) , E p = E u = ⋃ x ∈ C u W Δ ( x ) . (90) \mathcal{E}_L=\bigcup_{x\in\mathcal{C}_{\Delta}}\mathcal{W}_{\Delta}(x),\qquad\mathcal{E}_p=\mathcal{E}_u=\bigcup_{x\in\mathcal{C}_u}\mathcal{W}_{\Delta}(x). \tag*{(90)} E L = x ∈ C Δ ⋃ W Δ ( x ) , E p = E u = x ∈ C u ⋃ W Δ ( x ) . (90)
The pressure equation needs directly evaluated values only on E L \mathcal{E}_L E L , but we evaluate the pressure directly, and keep these values, on the larger set E p = E u \mathcal{E}_p=\mathcal{E}_u E p = E u , where the velocity is also evaluated directly.
Laplace problem
For a piecewise harmonic function v = S f + D g v = Sf + Dg v = S f + D g , we evaluate v I = S δ ( p ) f + D δ ( p ) g v^{I} = S_{\delta}^{(p)}f + D_{\delta}^{(p)}g v I = S δ ( p ) f + D δ ( p ) g on E L \mathcal{E}_{L} E L , taking the interior or exterior value according to the side of each grid point. We set
F h L ( x i j ) = { L h c v I ( x i j ) , x i j ∈ C Δ , 0 , x i j ∉ C Δ , (91) F_{h}^{L}(x_{ij}) =
\begin{cases}
L_{h_c}v^{I}(x_{ij}), & x_{ij} \in C_{\Delta},\\
0, & x_{ij} \notin C_{\Delta},
\end{cases} \tag*{(91)} F h L ( x ij ) = { L h c v I ( x ij ) , 0 , x ij ∈ C Δ , x ij ∈ / C Δ , (91)
and solve L h c v h = F h L L_{h_c}v_h = F_h^L L h c v h = F h L with Dirichlet data on the outer boundary computed from the layer potentials with the original kernels, which are not singular there. The reported solution uses the directly evaluated values on E L \mathcal{E}_{L} E L and the grid solution elsewhere:
v h a l l ( x ) = { v I ( x ) , x ∈ E L , v h ( x ) , x ∉ E L . (92) v_h^{\mathrm{all}}(x) =
\begin{cases}
v^{I}(x), & x \in\mathcal{E}_{L},\\
v_h(x), & x \notin\mathcal{E}_{L}.
\end{cases} \tag*{(92)} v h all ( x ) = { v I ( x ) , v h ( x ) , x ∈ E L , x ∈ / E L . (92)
To relate the error of the grid solution to that of the directly evaluated values, let e I = v I − v e^{I} = v^{I} - v e I = v I − v and e h = v h − v e_h = v_h - v e h = v h − v , and let τ h = L h c v \tau_h = L_{h_c}v τ h = L h c v be the truncation error at the regular points, where Δ v = 0 \Delta v = 0 Δ v = 0 . With exact boundary data,
L h c e h = 1 C Δ L h c e I − 1 C Δ c τ h . (93) L_{h_c}e_h = 1_{C_{\Delta}}L_{h_c}e^{I} - 1_{C_{\Delta}^{c}}\tau_h. \tag*{(93)} L h c e h = 1 C Δ L h c e I − 1 C Δ c τ h . (93)
Thus L h c L_{h_c} L h c is applied to the error of the directly evaluated values before the solve, and a bound on this error in the maximum norm alone allows growth by a factor h c − 2 h_c^{-2} h c − 2 in the right-hand side. This term, however, is supported on C Δ C_{\Delta} C Δ , within O ( h c ) O(h_c) O ( h c ) of Γ \Gamma Γ . For the standard discrete Laplacian, Beale and Layton [9 Lemmas 2.2 and 2.3] show that such a grid function is the discrete divergence of a grid function smaller by a factor O ( h c ) O(h_c) O ( h c ) , and that the discrete solution is bounded by the maximum norm of this function plus that of the truncation error, uniformly in h c h_c h c ; its first differences are bounded by the same quantity times ∣ log h c ∣ |\log h_c| ∣ log h c ∣ . If these estimates also hold for L h c L_{h_c} L h c , as Beale and Tlupova expect for their fourth-order Laplacian [13 Sect. 4.3] , then
max ∣ e h ∣ ≤ C ( h c − 1 max E L ∣ e I ∣ + max C Δ c ∣ τ h ∣ ) . (94) \max|e_h| \le C\left(h_c^{-1}\max_{\mathcal{E}_{L}}|e^{I}|+\max_{C_{\Delta}^{c}}|\tau_h|\right). \tag*{(94)} max ∣ e h ∣ ≤ C ( h c − 1 E L max ∣ e I ∣ + C Δ c max ∣ τ h ∣ ) . (94)
and an error O ( h c 5 ) O(h_c^{5}) O ( h c 5 ) in the directly evaluated values gives an error O ( h c 4 ) O(h_c^{4}) O ( h c 4 ) in the grid solution. For Stokes, the pressure error is also differentiated before the velocity solve. The fourth-order gradient is a combination of first differences, so the same argument gives a velocity error O ( h c 4 ∣ log h c ∣ ) O(h_c^{4}|\log h_c|) O ( h c 4 ∣ log h c ∣ ) .
Stokes problem: pressure first, then velocity
We first evaluate p I = S δ ( p ) ϕ − D δ ( p ) ψ p^{I} = S_{\delta}^{(p)}\phi- D_{\delta}^{(p)}\psi p I = S δ ( p ) ϕ − D δ ( p ) ψ on E p \mathcal{E}_{p} E p and solve
L h c p h = F h p , F h p ( x i j ) = { L h c p I ( x i j ) , x i j ∈ C Δ , 0 , x i j ∉ C Δ . (95) L_{h_c}p_h = F_h^p,\qquad
F_h^p(x_{ij}) =
\begin{cases}
L_{h_c}p^{I}(x_{ij}), & x_{ij} \in C_{\Delta},\\
0, & x_{ij} \notin C_{\Delta}.
\end{cases} \tag*{(95)} L h c p h = F h p , F h p ( x ij ) = { L h c p I ( x ij ) , 0 , x ij ∈ C Δ , x ij ∈ / C Δ . (95)
As in (92) , the reported pressure p h a l l p_h^{\mathrm{all}} p h all is p I p^{I} p I on E p \mathcal{E}_{p} E p and p h p_h p h elsewhere. At the regular points for the velocity, we differentiate p h p_h p h , not p h a l l p_h^{\mathrm{all}} p h all . We then evaluate the corrected Stokes velocity u I u^{I} u I on E u \mathcal{E}_{u} E u and set
F h u ( x i j ) = { L h c u I ( x i j ) , x i j ∈ C u , μ − 1 ∇ h c ( 4 ) p h ( x i j ) , x i j ∉ C u . (96) F_h^u(x_{ij}) =
\begin{cases}
L_{h_c}u^{I}(x_{ij}), & x_{ij} \in C_u,\\
\mu^{-1}\nabla_{h_c}^{(4)}p_h(x_{ij}), & x_{ij} \notin C_u.
\end{cases} \tag*{(96)} F h u ( x ij ) = { L h c u I ( x ij ) , μ − 1 ∇ h c ( 4 ) p h ( x ij ) , x ij ∈ C u , x ij ∈ / C u . (96)
We solve L h c ( u h ) k = ( F h u ) k L_{h_c}(u_h)_k=(F^u_h)_k L h c ( u h ) k = ( F h u ) k for the two velocity components, with Dirichlet data on the outer boundary. The reported velocity u h a l l u_h^{\mathrm{all}} u h all is u ′ u' u ′ on E u \mathcal{E}_u E u and u h u_h u h elsewhere.
The direct evaluations cost O ( h c − 2 ) O(h_c^{-2}) O ( h c − 2 ) , since there are N Γ = O ( h c − 1 ) N_\Gamma=O(h_c^{-1}) N Γ = O ( h c − 1 ) nodes on the curve and O ( h c − 1 ) O(h_c^{-1}) O ( h c − 1 ) grid points where the layer potentials are evaluated directly, and each DST solve costs O ( h c − 2 log ( 1 / h c ) ) O(h_c^{-2}\log(1/h_c)) O ( h c − 2 log ( 1/ h c )) , whereas direct evaluation at all O ( h c − 2 ) O(h_c^{-2}) O ( h c − 2 ) grid points would cost O ( h c − 3 ) O(h_c^{-3}) O ( h c − 3 ) . The direct evaluations could be accelerated by the fast multipole method [22 ] or a treecode [50 ] ; we do not use them here.
Results on a Cartesian grid
We apply this method to the Laplace solution for the kite of § 6.2 and to the Stokes solution for the starfish of § 7.1 , on the box ( − 1.5 , 1.5 ) 2 (-1.5,1.5)^2 ( − 1.5 , 1.5 ) 2 , so that h c = 3 / ( M + 1 ) h_c=3/(M+1) h c = 3/ ( M + 1 ) . We use the seventh-order kernel with δ = h 5 / 7 \delta=h^{5/7} δ = h 5/7 , as in (71) . For each curve and each M = 2 6 , 2 7 , … , 2 10 M=2^6,2^7,\ldots,2^{10} M = 2 6 , 2 7 , … , 2 10 , N Γ N_\Gamma N Γ is the smallest even integer at least 2 π max t ∣ x ′ ( t ) ∣ / h c 2\pi\max_t|x'(t)|/h_c 2 π max t ∣ x ′ ( t ) ∣/ h c with no prime factor larger than 11, for the FFT. The ratio Δ s / h c \Delta s/h_c Δ s / h c , with Δ s \Delta s Δ s as in (72) , then lies between 0.96 and 1.00.
Since h ≍ Δ s ≈ h c h\asymp\Delta s\approx h_c h ≍ Δ s ≈ h c , the regularization error at the directly evaluated points is O ( δ 7 ) = O ( h c 5 ) O(\delta^7)=O(h_c^5) O ( δ 7 ) = O ( h c 5 ) . The truncation errors (§ 8.1 ) are O ( h c 6 ) O(h_c^6) O ( h c 6 ) for the harmonic Laplace and pressure fields and O ( h c 4 ) O(h_c^4) O ( h c 4 ) for the velocity. If (94) holds, the grid solution therefore has error O ( h c 4 ) O(h_c^4) O ( h c 4 ) , or O ( h c 4 ∣ log h c ∣ ) O(h_c^4|\log h_c|) O ( h c 4 ∣ log h c ∣ ) for the velocity. In addition to the maximum error, we report the discrete L 2 L^2 L 2 norm
∥ e ∥ 2 , h = ( h c 2 ∑ i , j = 0 M + 1 ∣ e i j ∣ 2 ) 1 / 2 . (97) \lVert e\rVert_{2,h}=\left(h_c^2\sum_{i,j=0}^{M+1}|e_{ij}|^2\right)^{1/2}. \tag*{(97)} ∥ e ∥ 2 , h = ( h c 2 i , j = 0 ∑ M + 1 ∣ e ij ∣ 2 ) 1/2 . (97)
Here the sum includes the grid points on the outer boundary and, for the velocity, ∣ e i j ∣ |e_{ij}| ∣ e ij ∣ is the Euclidean norm of the error.
Figure 8 (a),(Figure 8 b) shows the maximum and discrete L 2 L^2 L 2 errors. On every grid, all three maximum errors occur at points where the potentials are evaluated directly, so the maximum norm mainly measures the error of these values near the interface, while the discrete L 2 L^2 L 2 norm also reflects the grid solution.
Figure 8. Maximum (a) and discrete L 2 L^2 L 2 (b) errors on the Cartesian grid for the Laplace solution for the kite and the Stokes solution for the starfish.
Between M = 2 9 M=2^9 M = 2 9 and 2 10 2^{10} 2 10 , the rates range from 3.91 for the Laplace solution to 3.29 for the velocity in the maximum norm, and from 4.30 to 3.81 in the discrete L 2 L^2 L 2 norm. The velocity error at M = 2 7 M=2^7 M = 2 7 is anomalously large; we return to this in § 8.7 . On the finest grid, the largest errors occur near the points of highest curvature, the two tips of the kite and the inner corners between the arms of the starfish (Figure 9 ).
Exact values at the directly evaluated points
To isolate the error of the grid method, we repeat the computation with exact values at the directly evaluated points and on the outer boundary, leaving everything else unchanged; for Stokes, the pressure is still computed first and differentiated with ∇ h c ( 4 ) \nabla_{h_c}^{(4)} ∇ h c ( 4 ) . At M = 2 9 M=2^9 M = 2 9 the maximum errors are below 10 − 11 10^{-11} 1 0 − 11 , many orders of magnitude smaller than those of the full method. With the location of the maximum errors noted in § 8.5 , this indicates that the regularized evaluations near the interface determine the maximum error.
To measure the error of the grid method above roundoff, we repeat this test with exact solutions that vary on a shorter length scale. Let
A ( x 1 , x 2 ) = e k ( x 1 − 0.895 ) cos ( k x 2 ) , B ( x 1 , x 2 ) = e k ( x 1 − 0.895 ) sin ( k x 2 ) , k = 8. \mathcal{A}(x_1,x_2)=e^{k(x_1-0.895)}\cos(kx_2),\qquad\mathcal{B}(x_1,x_2)=e^{k(x_1-0.895)}\sin(kx_2),\qquad k=8. A ( x 1 , x 2 ) = e k ( x 1 − 0.895 ) cos ( k x 2 ) , B ( x 1 , x 2 ) = e k ( x 1 − 0.895 ) sin ( k x 2 ) , k = 8.
Since max t ∥ x ( t ) ∥ ∞ = 0.895 \max_t \|x(t)\|_\infty= 0.895 max t ∥ x ( t ) ∥ ∞ = 0.895 for both curves, the shift by 0.895 gives ∣ A ∣ , ∣ B ∣ ≤ 1 |\mathcal{A}|, |\mathcal{B}| \le1 ∣ A ∣ , ∣ B ∣ ≤ 1 inside them. For Laplace, we take v − = A v^- = \mathcal{A} v − = A inside the kite and v + = 0 v^+ = 0 v + = 0 outside. For Stokes, with μ = 1 \mu= 1 μ = 1 , the stream function Ψ = x 2 A / k \Psi= x_2\mathcal{A}/k Ψ = x 2 A / k inside the starfish gives
u ˉ = ( ∂ x 2 Ψ , − ∂ x 1 Ψ ) = ( A / k − x 2 B , − x 2 A ) , p ˉ = − 2 A , \bar{u} = (\partial_{x_2}\Psi,-\partial_{x_1}\Psi) = (\mathcal{A}/k-x_2\mathcal{B},-x_2\mathcal{A}), \qquad\bar{p}=-2\mathcal{A}, u ˉ = ( ∂ x 2 Ψ , − ∂ x 1 Ψ ) = ( A / k − x 2 B , − x 2 A ) , p ˉ = − 2 A ,
and we set the exterior velocity and pressure to zero. These fields satisfy ∇ ⋅ u ˉ = 0 \nabla\cdot\bar{u} = 0 ∇ ⋅ u ˉ = 0 , Δ u ˉ = ∇ p ˉ \Delta\bar{u} = \nabla\bar{p} Δ u ˉ = ∇ p ˉ , and Δ p ˉ = 0 \Delta\bar{p} = 0 Δ p ˉ = 0 .
The errors decrease at rates close to six for the harmonic Laplace and pressure fields over M = 2 7 M = 2^7 M = 2 7 –2 9 2^9 2 9 , the finer grids being affected by roundoff, and close to four for the velocity between M = 2 9 M = 2^9 M = 2 9 and 2 10 2^{10} 2 10 , with both the computed and the exact pressure gradient (Figure 10 ); the exact gradient reduces the velocity error by a factor of about five. These rates agree with the truncation errors of (86) for harmonic and biharmonic functions.
Figure 10. Cartesian grid method with exact values at the directly evaluated points: (a) Laplace solution and pressure; (b) velocity with computed and exact pressure gradients.
The choice of δ \delta δ and the curvature
We repeat the computations with δ = 2 h c \delta= 2h_c δ = 2 h c , proportional to the mesh size, on the same grids and with the same nodes on the curves. On the finer grids, δ = 2 h c \delta= 2h_c δ = 2 h c is much more accurate than $\delta = h^{5/7}
Figure 11 . At M = 2 6 M = 2^{6} M = 2 6 the two values of δ \delta δ differ by less than 7% and give similar errors. At M = 2 10 M = 2^{10} M = 2 10 , δ = 2 h c \delta= 2h_c δ = 2 h c reduces the errors by about two orders of magnitude. At about 100 targets on this grid, including the points with the largest errors, the estimated discretization errors are below 10 − 13 10^{-13} 1 0 − 13 . Because the estimate of § 4.7 does not imply convergence when δ \delta δ is proportional to the mesh size, this comparison does not indicate the asymptotic rate.
Figure 11. Maximum errors on the Cartesian grid for δ = h 5 / 7 \delta= h^{5/7} δ = h 5/7 (solid) and δ = 2 h c \delta= 2h_c δ = 2 h c (dashed).
On coarse grids, the velocity error is also affected by the geometric factors in the local correction. Table 2 gives, for the starfish, min y ∈ E u ( 1 + b κ 0 ) \min_{y\in\mathcal{E}_u}(1+b\kappa_0) min y ∈ E u ( 1 + b κ 0 ) , which does not depend on δ \delta δ , and δ ∥ κ ∥ ∞ \delta\|\kappa\|_\infty δ ∥ κ ∥ ∞ , with ∥ κ ∥ ∞ ≈ 20.2 \|\kappa\|_\infty\approx20.2 ∥ κ ∥ ∞ ≈ 20.2 . At M = 2 7 M = 2^{7} M = 2 7 , the largest velocity error occurs at a point with α 2 = 1 + b κ 0 ≈ 0.21 \alpha^2 = 1+b\kappa_0 \approx0.21 α 2 = 1 + b κ 0 ≈ 0.21 . There the factors α − 3 \alpha^{-3} α − 3 and α − 5 \alpha^{-5} α − 5 in (37) make the stresslet correction large, while the discretization error and the error in the closest point are negligible. With δ = h 5 / 7 \delta= h^{5/7} δ = h 5/7 on this grid, δ ∥ κ ∥ ∞ ≈ 1.14 \delta\|\kappa\|_\infty\approx1.14 δ ∥ κ ∥ ∞ ≈ 1.14 , so δ \delta δ exceeds the smallest radius of curvature, contrary to the hypothesis Λ δ ∥ κ ∥ ∞ ≤ 1 2 \Lambda\delta\|\kappa\|_\infty\le\frac{1}{2} Λ δ ∥ κ ∥ ∞ ≤ 2 1 of Proposition 3.2 . On the finest grid δ ∥ κ ∥ ∞ ≈ 0.27 \delta\|\kappa\|_\infty\approx0.27 δ ∥ κ ∥ ∞ ≈ 0.27 and α \alpha α is close to one (Table 2 ).
Table 2. Geometric quantities for the Stokes computation for the starfish.
Conclusions
We have constructed regularized versions of the kernels for Laplace and Stokes layer potentials in two dimensions. They allow control of the discretization error: we prove that the regularization error is O ( δ p ) O(\delta^p) O ( δ p ) in the smoothing radius δ \delta δ , p = 3 , 5 , 7 p=3,5,7 p = 3 , 5 , 7 , uniformly for targets on and near the curve, so δ \delta δ can be taken several times the node spacing h h h to keep the trapezoidal discretization error small. The method is easy to implement and use: for every target, the trapezoidal rule is applied on one standard set of nodes on the curve, reusing the same samples of the curve and density.
The kernel coefficients are explicit functions of the scaled distance b / δ b/\delta b / δ to the curve, and the Laplace potentials require no derivatives of the density. The Stokes kernels use the same coefficients together with explicit local corrections that remove one remaining error moment, involve the curvature, and vanish on the curve; without them, the Stokeslet and stresslet errors are only first order in δ \delta δ . The stresslet correction and the pressure formula use tangential derivatives of the densities.
The error bounds hold when δ \delta δ is at most a fixed fraction of the smallest radius of curvature. For δ = c h q \delta=ch^q δ = c h q with q < 1 q<1 q < 1 , the discretization error decays faster than δ p \delta^p δ p , and the observed convergence rates are close to p q pq pq . With p = 7 p=7 p = 7 and δ \delta δ equal to two to four times the largest distance between adjacent nodes, we observe rates of 6.4 to 7 while the discretization error remains small; doubling the number of nodes at fixed δ \delta δ estimates the discretization error.
On a Cartesian grid, a fast Poisson solver supplies values at all grid points from the ones near the curve whose values are computed directly, where the largest errors occur. On finer grids, taking δ \delta δ equal to twice the grid spacing is much more accurate than δ = h 5 / 7 \delta=h^{5/7} δ = h 5/7 once δ \delta δ resolves the curvature. Future work includes choosing δ \delta δ locally along the curve from the node spacing or curvature, solving boundary integral equations, and treating open arcs and corners.
Evaluation of the kernels and coefficients
The formulas below avoid loss of accuracy from cancellation near ξ = 0 \xi=0 ξ = 0 and for large ∣ λ ∣ \lvert\lambda\rvert ∣ λ ∣ .
The logarithmic terms
When the source and target are close, we evaluate 2 log r + E 1 ( ξ ) 2\log r+E_1(\xi) 2 log r + E 1 ( ξ ) as a single expression, since each term diverges as r → 0 r\to0 r → 0 while their sum is analytic. Substituting the expansion [38 Chap. 6]
E 1 ( ξ ) = − γ − log ξ + ∑ k = 1 ∞ ( − 1 ) k + 1 ξ k k k ! , (98) E_1(\xi)=-\gamma-\log\xi+\sum_{k=1}^{\infty}\frac{(-1)^{k+1}\xi^k}{k\,k!}, \tag*{(98)} E 1 ( ξ ) = − γ − log ξ + k = 1 ∑ ∞ k k ! ( − 1 ) k + 1 ξ k , (98)
we obtain $2\log r+E_1(\xi)=2\log\delta-\gamma+\xi-\xi^2/4+\xi^3/18-\cdots$, which contains no singular terms. For ξ < 1 / 2 \xi<1/2 ξ < 1/2 , we evaluate the sum from the first 18 terms of this series; otherwise, we evaluate $E_1 with the function exp1 of SciPy [49 ] ; the same function is available as expint in MATLAB and as gsl_sf_expint_E1 in the GNU Scientific Library.
Coefficients
The functions H n \mathcal{H}_n H n satisfy a two-term recurrence. With erfcx ( x ) = e x 2 erfc ( x ) \operatorname{erfcx}(x) = e^{x^2}\operatorname{erfc}(x) erfcx ( x ) = e x 2 erfc ( x ) ,
H 0 = 2 [ 1 − π ∣ λ ∣ erfcx ∣ λ ∣ ] , H n = 2 n + 1 ( 1 − λ 2 H n − 2 ) , n = 2 , 4 , 6 , … (99) \mathcal{H}_0 = 2\left[1-\sqrt{\pi}\lvert\lambda\rvert\operatorname{erfcx}\lvert\lambda\rvert\right], \qquad\mathcal{H}_n = \frac{2}{n+1}\left(1-\lambda^2\mathcal{H}_{n-2}\right), \qquad n=2,4,6,\ldots \tag*{(99)} H 0 = 2 [ 1 − π ∣ λ ∣ erfcx ∣ λ ∣ ] , H n = n + 1 2 ( 1 − λ 2 H n − 2 ) , n = 2 , 4 , 6 , … (99)
Thus all H n \mathcal{H}_n H n are determined by H 0 \mathcal{H}_0 H 0 .
The recurrence is accurate for moderate ∣ λ ∣ \lvert\lambda\rvert ∣ λ ∣ , but loses digits through cancellation for large ∣ λ ∣ \lvert\lambda\rvert ∣ λ ∣ . The differences H 0 − H 2 \mathcal{H}_0-\mathcal{H}_2 H 0 − H 2 and H 0 − 2 H 2 + H 4 \mathcal{H}_0-2\mathcal{H}_2+\mathcal{H}_4 H 0 − 2 H 2 + H 4 in (17) and (18) decay faster than the individual H n \mathcal{H}_n H n and are even more affected. We therefore compute these differences directly from integrals with positive integrands, obtained from (15) :
U j ( λ ) = λ − 2 j − 2 ∫ 0 ∞ e − v v j ( 1 + v λ 2 ) − j − 3 / 2 d v , j = 0 , 1 , 2 , λ ≠ 0 , (100) U_j(\lambda)=\lambda^{-2j-2}\int_0^\infty e^{-v}v^j\left(1+\frac{v}{\lambda^2}\right)^{-j-3/2}\,dv,\qquad j=0,1,2,\qquad\lambda\ne0, \tag*{(100)} U j ( λ ) = λ − 2 j − 2 ∫ 0 ∞ e − v v j ( 1 + λ 2 v ) − j − 3/2 d v , j = 0 , 1 , 2 , λ = 0 , (100)
for which U 0 = H 0 U_0=\mathcal{H}_0 U 0 = H 0 , U 1 = H 0 − H 2 U_1=\mathcal{H}_0-\mathcal{H}_2 U 1 = H 0 − H 2 , and U 2 = H 0 − 2 H 2 + H 4 U_2=\mathcal{H}_0-2\mathcal{H}_2+\mathcal{H}_4 U 2 = H 0 − 2 H 2 + H 4 . The integrand is bounded, and the small factor λ − 2 j − 2 \lambda^{-2j-2} λ − 2 j − 2 is applied after the integration. We use these formulas for ∣ λ ∣ > 1 \lvert\lambda\rvert>1 ∣ λ ∣ > 1 and the recurrence (99) for ∣ λ ∣ ≤ 1 \lvert\lambda\rvert\le1 ∣ λ ∣ ≤ 1 , with erfcx in H 0 \mathcal{H}_0 H 0 so that the product of e λ 2 e^{\lambda^2} e λ 2 and erfc ∣ λ ∣ \operatorname{erfc}\lvert\lambda\rvert erfc ∣ λ ∣ is never formed. For ∣ λ ∣ > 1 \lvert\lambda\rvert>1 ∣ λ ∣ > 1 , the integrals in (100) are computed with the adaptive QUADPACK routine for infinite intervals [40 ] , called through scipy.integrate.quad [49 ] , with absolute and relative tolerances 2 × 10 − 14 2\times10^{-14} 2 × 1 0 − 14 and at most 50 subintervals.
Radial factors in the Stokes kernels
To avoid cancellation near ξ = 0 \xi=0 ξ = 0 , we rewrite the radial factors A = A ( p ) A=A^{(p)} A = A ( p ) and B = B ( p ) = − A ′ B=B^{(p)}=-A' B = B ( p ) = − A ′ of (32) and (33) , for fixed p p p and λ \lambda λ , with primes denoting ∂ ξ \partial_\xi ∂ ξ . Write P ( p ) ( ξ ; λ ) = 1 + ξ Z P^{(p)}(\xi;\lambda)=1+\xi Z P ( p ) ( ξ ; λ ) = 1 + ξ Z , so that Z Z Z is the polynomial ∑ k = 1 m ( − 1 ) k a k ( p ) ( λ ) ξ k − 1 \sum_{k=1}^{m}(-1)^k a_k^{(p)}(\lambda)\xi^{k-1} ∑ k = 1 m ( − 1 ) k a k ( p ) ( λ ) ξ k − 1 of degree m − 1 m-1 m − 1 . Then
Φ ( ξ ) = 1 − e − ξ ξ , A = Φ − e − ξ Z , \Phi(\xi)=\frac{1-e^{-\xi}}{\xi},\qquad A=\Phi-e^{-\xi}Z, Φ ( ξ ) = ξ 1 − e − ξ , A = Φ − e − ξ Z ,
∂ ξ d ( p ) = e − ξ ( 1 + ξ Z − Z − ξ Z ′ ) , B = − A ′ = − Φ ′ + e − ξ ( Z ′ − Z ) , (101) \partial_\xi d^{(p)}=e^{-\xi}(1+\xi Z-Z-\xi Z'),\qquad B=-A'=-\Phi'+e^{-\xi}(Z'-Z), \tag*{(101)} ∂ ξ d ( p ) = e − ξ ( 1 + ξ Z − Z − ξ Z ′ ) , B = − A ′ = − Φ ′ + e − ξ ( Z ′ − Z ) , (101)
so A A A , B B B , and ∂ ξ d ( p ) \partial_\xi d^{(p)} ∂ ξ d ( p ) are computed from Φ \Phi Φ and the polynomial Z Z Z , without dividing a small difference by ξ \xi ξ . We compute Φ \Phi Φ as − expm1 ( − ξ ) / ξ -\operatorname{expm1}(-\xi)/\xi − expm1 ( − ξ ) / ξ , with Φ ( 0 ) = 1 \Phi(0)=1 Φ ( 0 ) = 1 , where expm1 ( z ) = e z − 1 \operatorname{expm1}(z)=e^z-1 expm1 ( z ) = e z − 1 is accurate for small z z z . For ξ < 1 / 2 \xi<1/2 ξ < 1/2 , including ξ = 0 \xi=0 ξ = 0 , we compute Φ ′ \Phi' Φ ′ from the first 20 terms of the series Φ ′ ( ξ ) = ∑ j ≥ 1 ( − 1 ) j j ξ j − 1 / ( j + 1 ) ! \Phi'(\xi)=\sum_{j\ge1}(-1)^j j\xi^{j-1}/(j+1)! Φ ′ ( ξ ) = ∑ j ≥ 1 ( − 1 ) j j ξ j − 1 / ( j + 1 )! , which avoids the cancellation between ξ e − ξ \xi e^{-\xi} ξ e − ξ and 1 − e − ξ 1-e^{-\xi} 1 − e − ξ ; otherwise we use Φ ′ ( ξ ) = [ ξ e − ξ + expm1 ( − ξ ) ] / ξ 2 \Phi'(\xi)=[\xi e^{-\xi}+\operatorname{expm1}(-\xi)]/\xi^2 Φ ′ ( ξ ) = [ ξ e − ξ + expm1 ( − ξ )] / ξ 2 . The tensors are then assembled as r i r j A / δ 2 r_i r_j A/\delta^2 r i r j A / δ 2 and r i r j ( r ⋅ n ) ( − A ′ ) / δ 4 r_i r_j(\mathbf{r}\cdot\mathbf{n})(-A')/\delta^4 r i r j ( r ⋅ n ) ( − A ′ ) / δ 4 , so neither r − 2 r^{-2} r − 2 nor r − 4 r^{-4} r − 4 is ever formed. These formulas give B ( 0 ) = 1 2 + a 1 ( p ) + a 2 ( p ) B(0)=\frac{1}{2}+a_1^{(p)}+a_2^{(p)} B ( 0 ) = 2 1 + a 1 ( p ) + a 2 ( p ) , in agreement with the limit in Section 3.1.
Acknowledgments
The author thanks J. Thomas Beale for explaining the problem and for many helpful suggestions on the first draft of this article, and Yoichiro Mori for hosting the visit of J. Thomas Beale to the University of Pennsylvania during which this work began.
Declarations
Competing interests The author declares no competing interests.
Use of AI tools The author used Claude Code and Gemini to assist with code generation and writing/editing the manuscript. After using these tools, the author reviewed and verified all mathematical derivations, numerical code, and text, and takes full responsibility for the content of the article.
Code and data availability The code and the data for the numerical examples are available in the GitHub repository https://github.com/han-zhou-math/high-order-reg-2d.git.