Introduction
The convergence behaviour of classical and optimized Schwarz Waveform Relaxation (SWR) methods for parabolic partial differential equations (PDEs) differs markedly from that of Schwarz methods for elliptic problems [5 ] , owing to the non-normal nature of the time derivative operator. It is known that classical SWR methods (i.e., with Dirichlet transmission conditions) converge superlinearly on bounded time intervals, but such convergence deteriorates as the time interval length T T T goes to infinity; in the limit, convergence is only linear [11 , 13 , 12 ] . As the iteration progresses, the convergence curve exhibits different phases: there is an initial rapid phase, then a slowdown, and finally an acceleration to superlinear convergence [10 ] . Moreover, when Robin transmission conditions are used, the best Robin parameter p p p that minimizes error reduction depends on the number of iterations n n n one intends to take, rather than a fixed constant [4 ] .
The classical approach to finding best parameters via Fourier analysis, see [3 ] , is unable to capture this dependence on n n n . Indeed, such an analysis does not even yield superlinear convergence estimates (but see [10 ] for a Kernel Fourier analysis that uses the heat kernel to restore superlinearity with the estimates). As a result, superlinear estimates typically employ kernel techniques [11 , 12 ] or Laplace transforms [8 , 14 , 9 ] . The latter is useful when the iteration function in Laplace space admits a closed form inverse Laplace transform, or can be expressed as a convergent series of such inverse transforms. This is feasible for Dirichlet and Neumann transmission conditions, but not for Robin transmission conditions. As a result, the problem of finding n n n -dependent best Robin parameters remains so far an open problem, even for the simple case of
*Department of Mathematics and Statistics, Université Laval, Quebec, Canada alejandro.alfonso-rodriguez.1@ulaval.ca.
†Section de mathématiques, Université de Genève, Geneva, Switzerland martin.gander@unige.ch.
‡Department of Mathematics and Statistics, Université Laval, Quebec, Canada felix.kwok@mat.ulaval.ca. the one-dimensional heat equation in two subdomains. Note that in the discrete-time setting, it is possible to find the best Robin parameter by numerically minimizing the norm of the appropriate iteration matrix, whose size is proportional to the number of time steps, see [2 , 1 ] .
The goal of this paper is twofold. The first goal is to introduce a theoretical tool for estimating the error of domain decomposition methods with simple Laplace transforms, even when the inverse transforms do not have closed form formulas. The tool relies on a simple relationship between the Laplace and Fourier transforms, together with the Plancherel theorem. Although the tool is easy to understand, its application is not so widespread in the domain decomposition literature; we therefore illustrate its utility by giving a new convergence analysis of the classical SWR method for two subdomains, for which we are able to recover the known superlinear estimates using only a few lines of calculations. The second goal of the paper is to use the same techniques to fully analyze the convergence of optimized SWR (OSWR), first as a function of the Robin parameter p p p , and then by choosing the parameter p p p to minimize this contraction estimate for a given number of iterations n n n . Note that this is the first time we are able to truly optimize p p p in the superlinear regime, even though the algorithm exists and has been used for three decades. In each case, we use numerical examples to illustrate the sharpness of our estimates.
Laplace transform and exponential weighting
Let Ω = R \Omega= \mathbb{R} Ω = R . To explain this new technique, it suffices to focus on the one-dimensional heat equation1
∂ t u − ∂ y y u = f , ( y , t ) ∈ Ω × ( 0 , T ) (1) \partial_t u-\partial_{yy}u=f,\qquad(y,t)\in\Omega\times(0,T) \tag*{(1)} ∂ t u − ∂ y y u = f , ( y , t ) ∈ Ω × ( 0 , T ) (1)
with source term f ∈ L 2 ( 0 , T ; L 2 ( Ω ) ) f\in L^2(0,T;L^2(\Omega)) f ∈ L 2 ( 0 , T ; L 2 ( Ω )) and initial conditions u ( ⋅ , 0 ) = u 0 ∈ H 1 ( Ω ) u(\cdot,0)=u_0\in H^1(\Omega) u ( ⋅ , 0 ) = u 0 ∈ H 1 ( Ω ) . Note the higher regularity assumed in the initial conditions, which is needed for the well-posedness of the SWR methods later. The problem (1) is well posed [6 , 7 ] with a unique solution u ∈ H 2 , 1 ( Ω × ( 0 , T ) ) : = L 2 ( 0 , T ; H 2 ( Ω ) ) ∩ H 1 ( 0 , T ; L 2 ( Ω ) ) u\in H^{2,1}(\Omega\times(0,T)):=L^2(0,T;H^2(\Omega))\cap H^1(0,T;L^2(\Omega)) u ∈ H 2 , 1 ( Ω × ( 0 , T )) := L 2 ( 0 , T ; H 2 ( Ω )) ∩ H 1 ( 0 , T ; L 2 ( Ω )) .
Next, we consider a decomposition of Ω \Omega Ω into two spatial subdomains Ω 1 = ( − ∞ , δ ) \Omega_1=(-\infty,\delta) Ω 1 = ( − ∞ , δ ) , Ω 2 = ( 0 , ∞ ) \Omega_2=(0,\infty) Ω 2 = ( 0 , ∞ ) with an overlap of width δ > 0 \delta>0 δ > 0 . The alternating (O)SWR method solves, for n = 1 , 2 , 3 , … n=1,2,3,\ldots n = 1 , 2 , 3 , … ,
∂ t u 1 n − ∂ y y u 1 n = f in Ω 1 × ( 0 , T ) , ∂ t u 2 n − ∂ y y u 2 n = f in Ω 2 × ( 0 , T ) , u 1 n ( ⋅ , 0 ) = u 0 in Ω 1 , u 2 n ( ⋅ , 0 ) = u 0 in Ω 2 , B 1 u 1 n ( δ , ⋅ ) = B 1 u 2 n − 1 ( δ , ⋅ ) on ( 0 , T ) , B 2 u 2 n ( 0 , ⋅ ) = B 2 u 1 n ( 0 , ⋅ ) on ( 0 , T ) . \begin{aligned}
\partial_t u_1^n-\partial_{yy}u_1^n&=f &&\text{in }\Omega_1\times(0,T),&
\partial_t u_2^n-\partial_{yy}u_2^n&=f &&\text{in }\Omega_2\times(0,T),\\
u_1^n(\cdot,0)&=u_0 &&\text{in }\Omega_1,&
u_2^n(\cdot,0)&=u_0 &&\text{in }\Omega_2,\\
B_1u_1^n(\delta,\cdot)&=B_1u_2^{n-1}(\delta,\cdot) &&\text{on }(0,T),&
B_2u_2^n(0,\cdot)&=B_2u_1^n(0,\cdot) &&\text{on }(0,T).
\end{aligned} ∂ t u 1 n − ∂ y y u 1 n u 1 n ( ⋅ , 0 ) B 1 u 1 n ( δ , ⋅ ) = f = u 0 = B 1 u 2 n − 1 ( δ , ⋅ ) in Ω 1 × ( 0 , T ) , in Ω 1 , on ( 0 , T ) , ∂ t u 2 n − ∂ y y u 2 n u 2 n ( ⋅ , 0 ) B 2 u 2 n ( 0 , ⋅ ) = f = u 0 = B 2 u 1 n ( 0 , ⋅ ) in Ω 2 × ( 0 , T ) , in Ω 2 , on ( 0 , T ) .
where the transmission conditions can be
classical (Dirichlet) conditions: B i = i d B_i=\mathrm{id} B i = id , i.e., the identity operator;
optimized (Robin) conditions: B i = ∂ n i + p B_i=\partial_{n_i}+p B i = ∂ n i + p , where p > 0 p>0 p > 0 is a constant and ∂ n i \partial_{n_i} ∂ n i is the normal derivative, with n i n_i n i being the normal vector pointing outward from Ω i \Omega_i Ω i . In other words, we have B 1 = ∂ y + p B_1=\partial_y+p B 1 = ∂ y + p and B 2 = − ∂ y + p B_2=-\partial_y+p B 2 = − ∂ y + p .
For classical SWR, the iteration is well posed if f ∈ L 2 ( 0 , T ; L 2 ( Ω ) ) f\in L^2(0,T;L^2(\Omega)) f ∈ L 2 ( 0 , T ; L 2 ( Ω )) , u 0 ∣ Ω i ∈ H 1 ( Ω i ) u_0|_{\Omega_i}\in H^1(\Omega_i) u 0 ∣ Ω i ∈ H 1 ( Ω i ) , and if the initial guess u 1 0 ( δ , ⋅ ) u_1^0(\delta,\cdot) u 1 0 ( δ , ⋅ ) is in H 3 / 4 ( 0 , T ) H^{3/4}(0,T) H 3/4 ( 0 , T ) and compatible with u 0 u_0 u 0 at t = 0 t=0 t = 0 . The iterates u i n u_i^n u i n then converge to u ∣ Ω i u|_{\Omega_i} u ∣ Ω i in H 2 , 1 ( Ω i × ( 0 , T ) ) H^{2,1}(\Omega_i\times(0,T)) H 2 , 1 ( Ω i × ( 0 , T )) as n → ∞ n\to\infty n → ∞ . For OSWR with p > 0 p>0 p > 0 , the algorithm is well-posed if f ∈ L 2 ( 0 , T ; L 2 ( Ω ) ) f\in L^2(0,T;L^2(\Omega)) f ∈ L 2 ( 0 , T ; L 2 ( Ω )) , u 0 ∣ Ω i ∈ H 1 ( Ω i ) u_0|_{\Omega_i}\in H^1(\Omega_i) u 0 ∣ Ω i ∈ H 1 ( Ω i ) and if the initial Robin trace B 1 u 1 0 ( δ , ⋅ ) B_1u_1^0(\delta,\cdot) B 1 u 1 0 ( δ , ⋅ ) is in H 1 / 4 ( 0 , T ) H^{1/4}(0,T) H 1/4 ( 0 , T ) ; in that case, the iterates u i n u_i^n u i n all lie in H 2 , 1 ( Ω i × ( 0 , T ) ) H^{2,1}(\Omega_i\times(0,T)) H 2 , 1 ( Ω i × ( 0 , T )) and converge to u ∣ Ω i u|_{\Omega_i} u ∣ Ω i as n → ∞ n\to\infty n → ∞ . See [6 , 7 ] for the proof of these results.
Since the problem (1) is linear, it suffices to analyze the convergence of the local errors E i n = u i n − u ∣ Ω i E_i^n=u_i^n-u|_{\Omega_i} E i n = u i n − u ∣ Ω i to zero. The error equations satisfied by E i n E_i^n E i n have the same form as the original problem, but with zero initial conditions and source term:
∂ t E 1 n − ∂ y 2 E 1 n = 0 in Ω 1 × ( 0 , T ) , ∂ t E 2 n − ∂ y 2 E 2 n = 0 in Ω 2 × ( 0 , T ) , E 1 n ( ⋅ , 0 ) = 0 in Ω 1 , E 2 n ( ⋅ , 0 ) = 0 in Ω 2 , B 1 E 1 n ( δ , ⋅ ) = B 1 E 2 n − 1 ( δ , ⋅ ) on ( 0 , T ) , B 2 E 2 n ( 0 , ⋅ ) = B 2 E 1 n − 1 ( 0 , ⋅ ) on ( 0 , T ) . (2) \begin{aligned}
\partial_t E_1^n-\partial_y^2 E_1^n&=0 &&\text{in }\Omega_1\times(0,T),&
\partial_t E_2^n-\partial_y^2 E_2^n&=0 &&\text{in }\Omega_2\times(0,T),\\
E_1^n(\cdot,0)&=0 &&\text{in }\Omega_1,&
E_2^n(\cdot,0)&=0 &&\text{in }\Omega_2,\\
\mathcal{B}_1E_1^n(\delta,\cdot)&=\mathcal{B}_1E_2^{n-1}(\delta,\cdot) &&\text{on }(0,T),&
\mathcal{B}_2E_2^n(0,\cdot)&=\mathcal{B}_2E_1^{n-1}(0,\cdot) &&\text{on }(0,T). \tag*{(2)}
\end{aligned} ∂ t E 1 n − ∂ y 2 E 1 n E 1 n ( ⋅ , 0 ) B 1 E 1 n ( δ , ⋅ ) = 0 = 0 = B 1 E 2 n − 1 ( δ , ⋅ ) in Ω 1 × ( 0 , T ) , in Ω 1 , on ( 0 , T ) , ∂ t E 2 n − ∂ y 2 E 2 n E 2 n ( ⋅ , 0 ) B 2 E 2 n ( 0 , ⋅ ) = 0 = 0 = B 2 E 1 n − 1 ( 0 , ⋅ ) in Ω 2 × ( 0 , T ) , in Ω 2 , on ( 0 , T ) . (2)
We now introduce the main idea of our paper, namely the Laplace transform as an exponentially weighted Fourier transform. Let g ∈ L 2 ( 0 , T ) g\in L^2(0,T) g ∈ L 2 ( 0 , T ) , and g ~ ∈ L 2 ( R ) \widetilde{g}\in L^2(\mathbb{R}) g ∈ L 2 ( R ) be the extension of g g g by zero outside the interval ( 0 , T ) (0,T) ( 0 , T ) , i.e.,
g ~ ( t ) = { g ( t ) , 0 < t < T , 0 , else . \widetilde{g}(t)=
\begin{cases}
g(t), & 0<t<T,\\
0, & \text{else}.
\end{cases} g ( t ) = { g ( t ) , 0 , 0 < t < T , else .
Recall that the Laplace transform of g g g , which is a function of the complex variable s = r + i ω s=r+i\omega s = r + iω with r ≥ 0 r\geq0 r ≥ 0 , is given by
L { g } ( s ) = ∫ 0 ∞ g ( t ) e − s t d t = ∫ − ∞ ∞ e − i ω t ( e − r t g ~ ( t ) ) d t = F { e − r t g ~ ( t ) } = : g ^ ( r , ω ) , \mathcal{L}\{g\}(s)=\int_0^\infty g(t)e^{-st}\,dt=\int_{-\infty}^{\infty}e^{-i\omega t}(e^{-rt}\widetilde{g}(t))\,dt=\mathcal{F}\{e^{-rt}\widetilde{g}(t)\}=:\widehat{g}(r,\omega), L { g } ( s ) = ∫ 0 ∞ g ( t ) e − s t d t = ∫ − ∞ ∞ e − iω t ( e − r t g ( t )) d t = F { e − r t g ( t )} =: g ( r , ω ) ,
where F \mathcal{F} F denotes the Fourier transform. We can therefore invoke the Plancherel theorem to deduce that
1 2 π ∫ − ∞ ∞ ∣ g ^ ( r , ω ) ∣ 2 d ω = ∫ − ∞ ∞ ∣ e − r t g ~ ( t ) ∣ 2 d t = ∫ 0 T e − 2 r t g 2 ( t ) d t , (3) \frac{1}{2\pi}\int_{-\infty}^{\infty}|\widehat{g}(r,\omega)|^2\,d\omega=\int_{-\infty}^{\infty}|e^{-rt}\widetilde{g}(t)|^2\,dt=\int_0^T e^{-2rt}g^2(t)\,dt, \tag*{(3)} 2 π 1 ∫ − ∞ ∞ ∣ g ( r , ω ) ∣ 2 d ω = ∫ − ∞ ∞ ∣ e − r t g ( t ) ∣ 2 d t = ∫ 0 T e − 2 r t g 2 ( t ) d t , (3)
which is valid for all r ≥ 0 r\geq0 r ≥ 0 .
We now explain how this idea can be used to analyze (O)SWR methods. To illustrate the technique, let us concentrate for now on the classical transmission conditions B 1 = B 2 = id \mathcal{B}_1=\mathcal{B}_2=\operatorname{id} B 1 = B 2 = id ; the case of optimized transmission conditions will be considered in detail in the next sections. We start by considering the error equations (2) and study the behaviour of E i n ( y , t ) E_i^n(y,t) E i n ( y , t ) for all time t > 0 t>0 t > 0 (instead of truncating the solution at t = T t=T t = T ). Writing E ^ 1 n ( y , s ) = L { E 1 n ( y , ⋅ ) } ( s ) \widehat{E}_1^n(y,s)=\mathcal{L}\{E_1^n(y,\cdot)\}(s) E 1 n ( y , s ) = L { E 1 n ( y , ⋅ )} ( s ) , the Laplace transform of the error equation for each y ∈ Ω 1 y\in\Omega_1 y ∈ Ω 1 reads
s E ^ 1 n − ∂ y y E 1 n = 0 , y ∈ ( − ∞ , δ ) , s\widehat{E}_1^n-\partial_{yy}E_1^n=0,\qquad y\in(-\infty,\delta), s E 1 n − ∂ y y E 1 n = 0 , y ∈ ( − ∞ , δ ) ,
with lim y → − ∞ E ^ 1 ( y , s ) = 0 \lim_{y\to-\infty}\widehat{E}_1(y,s)=0 lim y → − ∞ E 1 ( y , s ) = 0 and E ^ 1 n ( δ , s ) = E ^ 2 n − 1 ( 0 , s ) \widehat{E}_1^n(\delta,s)=\widehat{E}_2^{n-1}(0,s) E 1 n ( δ , s ) = E 2 n − 1 ( 0 , s ) for all ℜ ( s ) ≥ 0 \Re(s)\geq0 ℜ ( s ) ≥ 0 . For a fixed s s s , this ODE in y y y can be solved explicitly to yield
E ^ 1 n ( y , s ) = E ^ 2 n − 1 ( δ , s ) e ( y − δ ) s , y < δ , n ≥ 1 , (4) \widehat{E}_1^n(y,s)=\widehat{E}_2^{n-1}(\delta,s)e^{(y-\delta)\sqrt{s}},\qquad y<\delta,\quad n\geq1, \tag*{(4)} E 1 n ( y , s ) = E 2 n − 1 ( δ , s ) e ( y − δ ) s , y < δ , n ≥ 1 , (4)
where s = ξ + i η \sqrt{s}=\xi+i\eta s = ξ + i η is the unique square root of s s s with positive real part. Similarly, taking the Laplace transform in time of the PDE in Ω 2 \Omega_2 Ω 2 and solving the resulting ODE gives
E ^ 2 n ( y , s ) = E ^ 1 n ( 0 , s ) e − y s , y > 0 , n ≥ 1. (5) \widehat{E}_2^n(y,s)=\widehat{E}_1^n(0,s)e^{-y\sqrt{s}},\qquad y>0,\quad n\geq1. \tag*{(5)} E 2 n ( y , s ) = E 1 n ( 0 , s ) e − y s , y > 0 , n ≥ 1. (5)
Combining (4) and (5) yields the recurrence
E ^ 2 n ( δ , s ) = E ^ 2 n − 1 ( δ , s ) e − 2 δ s = ⋯ = E ^ 2 0 ( δ , s ) e − 2 n δ s . (6) \widehat{E}_2^n(\delta,s)=\widehat{E}_2^{n-1}(\delta,s)e^{-2\delta\sqrt{s}}=\cdots=\widehat{E}_2^0(\delta,s)e^{-2n\delta\sqrt{s}}. \tag*{(6)} E 2 n ( δ , s ) = E 2 n − 1 ( δ , s ) e − 2 δ s = ⋯ = E 2 0 ( δ , s ) e − 2 n δ s . (6)
At this point, if we know the inverse Laplace transform ρ n ( t ) \rho_n(t) ρ n ( t ) of e − 2 n δ s e^{-2n\delta\sqrt{s}} e − 2 n δ s , we can take inverse Laplace transforms of both sides to conclude that
E 2 n ( δ , t ) = ∫ 0 t ρ n ( t − τ ) E 2 0 ( δ , τ ) d τ , E_2^n(\delta,t)=\int_0^t \rho_n(t-\tau)E_2^0(\delta,\tau)\,d\tau, E 2 n ( δ , t ) = ∫ 0 t ρ n ( t − τ ) E 2 0 ( δ , τ ) d τ ,
and can use the properties of ρ n \rho_n ρ n to estimate the size of ∥ E 2 n ( δ , t ) ∥ L 2 ( 0 , T ) \|E_2^n(\delta,t)\|_{L^2(0,T)} ∥ E 2 n ( δ , t ) ∥ L 2 ( 0 , T ) . Here, we take a different approach: we instead assume, without loss of generality, that the initial error on the interface E 2 0 ( δ , ⋅ ) E_2^0(\delta,\cdot) E 2 0 ( δ , ⋅ ) has support in the interval ( 0 , T ) (0,T) ( 0 , T ) .2 Then for any r ≥ 0 r\geq0 r ≥ 0 , we write
∥ E 2 n ( δ , ⋅ ) ∥ L 2 ( 0 , T ) 2 = ∫ 0 T ( E 2 n ( δ , t ) ) 2 d t ≤ e 2 r T ∫ 0 T e − 2 r t ( E 2 n ( δ , t ) ) 2 d t ≤ e 2 r T ∫ 0 ∞ e − 2 r t ( E 2 n ( δ , t ) ) 2 d t = e 2 r T 2 π ∫ − ∞ ∞ ∣ E ^ 2 n ( δ , r + i ω ) ∣ 2 d ω , \begin{aligned}
\|E_2^n(\delta,\cdot)\|_{L^2(0,T)}^2=\int_0^T(E_2^n(\delta,t))^2\,dt\leq e^{2rT}\int_0^T e^{-2rt}(E_2^n(\delta,t))^2\,dt \\
&\leq e^{2rT}\int_0^\infty e^{-2rt}(E_2^n(\delta,t))^2\,dt=\frac{e^{2rT}}{2\pi}\int_{-\infty}^{\infty}|\widehat{E}_2^n(\delta,r+i\omega)|^2\,d\omega,
\end{aligned} ∥ E 2 n ( δ , ⋅ ) ∥ L 2 ( 0 , T ) 2 = ∫ 0 T ( E 2 n ( δ , t ) ) 2 d t ≤ e 2 r T ∫ 0 T e − 2 r t ( E 2 n ( δ , t ) ) 2 d t ≤ e 2 r T ∫ 0 ∞ e − 2 r t ( E 2 n ( δ , t ) ) 2 d t = 2 π e 2 r T ∫ − ∞ ∞ ∣ E 2 n ( δ , r + iω ) ∣ 2 d ω ,
where we used the Plancherel theorem (3) twice. With the recurrence (6) , we deduce that
∥ E 2 n ( δ , ⋅ ) ∥ L 2 ( 0 , T ) 2 ≤ e 2 r T 2 π ∫ − ∞ ∞ ∣ E ^ 2 0 ( δ , r + i ω ) ∣ 2 ∣ e − 4 n δ r + i ω ∣ d ω ≤ ( ρ ( r , n ) ) 2 2 π ∫ − ∞ ∞ ∣ E ^ 2 0 ( δ , r + i ω ) ∣ 2 d ω = ( ρ ( r , n ) ) 2 ∥ E 2 0 ( δ , ⋅ ) ∥ L 2 ( 0 , T ) 2 , \begin{aligned}
\|E_2^n(\delta,\cdot)\|_{L^2(0,T)}^2\leq\frac{e^{2rT}}{2\pi}\int_{-\infty}^{\infty}|\widehat{E}_2^0(\delta,r+i\omega)|^2|e^{-4n\delta\sqrt{r+i\omega}}|\,d\omega\\
&\leq\frac{(\rho(r,n))^2}{2\pi}\int_{-\infty}^{\infty}|\widehat{E}_2^0(\delta,r+i\omega)|^2\,d\omega=(\rho(r,n))^2\|E_2^0(\delta,\cdot)\|_{L^2(0,T)}^2,
\end{aligned} ∥ E 2 n ( δ , ⋅ ) ∥ L 2 ( 0 , T ) 2 ≤ 2 π e 2 r T ∫ − ∞ ∞ ∣ E 2 0 ( δ , r + iω ) ∣ 2 ∣ e − 4 n δ r + iω ∣ d ω ≤ 2 π ( ρ ( r , n ) ) 2 ∫ − ∞ ∞ ∣ E 2 0 ( δ , r + iω ) ∣ 2 d ω = ( ρ ( r , n ) ) 2 ∥ E 2 0 ( δ , ⋅ ) ∥ L 2 ( 0 , T ) 2 ,
where we introduced
ρ ( r , n ) = e r T sup ω ∈ R ∣ e − 2 n δ r + i ω ∣ , (7) \rho(r,n)=e^{rT}\sup_{\omega\in\mathbb{R}}|e^{-2n\delta\sqrt{r+i\omega}}|, \tag*{(7)} ρ ( r , n ) = e r T ω ∈ R sup ∣ e − 2 n δ r + iω ∣ , (7)
an upper bound for the n n n -step contraction factor for the L 2 L^2 L 2 -norm (without the square). Note that this upper bound is valid for any r ≥ 0 r\geq0 r ≥ 0 ; thus, we can obtain the best estimate by choosing r r r judiciously. Indeed, if we let r + i ω = x + i b \sqrt{r+i\omega}=x+ib r + iω = x + ib and calculate
r + i ω = ( x + i b ) 2 = ( x 2 − b 2 ) + 2 i b x ⟹ r = x 2 − b 2 , (8) r+i\omega=(x+ib)^2=(x^2-b^2)+2ibx\quad\Longrightarrow\quad r=x^2-b^2, \tag*{(8)} r + iω = ( x + ib ) 2 = ( x 2 − b 2 ) + 2 ib x ⟹ r = x 2 − b 2 , (8)
then together with ∣ r + i ω ∣ = ∣ x + i b ∣ 2 = x 2 + b 2 |r+i\omega|=|x+ib|^2=x^2+b^2 ∣ r + iω ∣ = ∣ x + ib ∣ 2 = x 2 + b 2 , we deduce that x 2 = 1 2 ( r + r 2 + ω 2 ) x^2=\frac{1}{2}(r+\sqrt{r^2+\omega^2}) x 2 = 2 1 ( r + r 2 + ω 2 ) . We therefore obtain
ρ ( r , n ) = e r T sup ω ∈ R e − 2 n δ x ( r , ω ) = e r T − 2 n δ r . \rho(r,n)=e^{rT}\sup_{\omega\in\mathbb{R}}e^{-2n\delta x(r,\omega)}=e^{rT-2n\delta\sqrt{r}}. ρ ( r , n ) = e r T ω ∈ R sup e − 2 n δ x ( r , ω ) = e r T − 2 n δ r .
The best choice of r r r is the one that minimizes ln ( ρ ( r , n ) ) = r T − 2 n δ r \ln(\rho(r,n))=rT-2n\delta\sqrt{r} ln ( ρ ( r , n )) = r T − 2 n δ r , i.e., when r = r ∗ = ( n δ / T ) 2 r=r^*=(n\delta/T)^2 r = r ∗ = ( n δ / T ) 2 . For this value, we have ρ ( r ∗ , n ) = e − n 2 δ 2 / T \rho(r^*,n)=e^{-n^2\delta^2/T} ρ ( r ∗ , n ) = e − n 2 δ 2 / T . Thus, we deduce the estimate
∥ E 2 n ( δ , ⋅ ) ∥ L 2 ( 0 , T ) ≤ e − n 2 δ 2 / T ∥ E 2 0 ( δ , ⋅ ) ∥ L 2 ( 0 , T ) . (9) \|E_2^n(\delta,\cdot)\|_{L^2(0,T)}\leq e^{-n^2\delta^2/T}\|E_2^0(\delta,\cdot)\|_{L^2(0,T)}. \tag*{(9)} ∥ E 2 n ( δ , ⋅ ) ∥ L 2 ( 0 , T ) ≤ e − n 2 δ 2 / T ∥ E 2 0 ( δ , ⋅ ) ∥ L 2 ( 0 , T ) . (9)
Note that r r r is not a parameter of the algorithm, but an analysis parameter that we optimize to get the sharpest bounds possible for this technique.
Recall the well-known superlinear convergence bound in [12 Cor. 5.2] , which can be adapted to the alternating SWR method to yield
∥ E j n ∥ L ∞ ( Ω j × ( 0 , T ) ) ≤ erfc ( n δ T ) max ℓ = 1 , 2 ∥ E ℓ 0 ∥ L ∞ ( Ω j × ( 0 , T ) ) . \|E_j^n\|_{L^\infty(\Omega_j\times(0,T))}\leq\operatorname{erfc}\left(\frac{n\delta}{\sqrt{T}}\right)\max_{\ell=1,2}\|E_\ell^0\|_{L^\infty(\Omega_j\times(0,T))}. ∥ E j n ∥ L ∞ ( Ω j × ( 0 , T )) ≤ erfc ( T n δ ) ℓ = 1 , 2 max ∥ E ℓ 0 ∥ L ∞ ( Ω j × ( 0 , T )) .
Since erfc ( x ) = e − x 2 ( 1 x π + O ( x − 3 ) ) \operatorname{erfc}(x)=e^{-x^2}\left(\frac{1}{x\sqrt{\pi}}+O(x^{-3})\right) erfc ( x ) = e − x 2 ( x π 1 + O ( x − 3 ) ) , we see that (9) captures the dominant order of the superlinear convergence without the need of exact formulas for the heat kernel nor for the inverse Laplace transform! The true power of this new analysis tool will be shown in the next section, where it allows us to obtain new results for OSWR that remained out of reach for three decades.
Convergence analysis of OSWR
We now analyze in detail the convergence of OSWR, which is significantly more involved than for SWR with classical transmission conditions. We will proceed as follows:
Derive the n n n -step convergence factor ρ ( r , p , n ) \rho(r,p,n) ρ ( r , p , n ) , analogous to the expression (7) for classical SWR, but with an extra dependence on the Robin parameter p p p . The formula for ρ ( r , p , n ) \rho(r,p,n) ρ ( r , p , n ) will contain a supremum over all frequencies ω \omega ω , just like in (7) .
Calculate the supremum inside ρ ( r , p , n ) \rho(r,p,n) ρ ( r , p , n ) to obtain upper bounds for each value of r ≥ 0 r \ge0 r ≥ 0 .
Study the behaviour of r ∗ = r ∗ ( p ) r^*=r^*(p) r ∗ = r ∗ ( p ) that minimizes ρ ( r , p , n ) \rho(r,p,n) ρ ( r , p , n ) for given values of p p p and n n n .
For a given value of n n n , find the optimal p p p that minimizes the n n n -step contraction factor ρ ( r ∗ ( p ) , p , n ) \rho(r^*(p),p,n) ρ ( r ∗ ( p ) , p , n ) .
The first three points will be addressed in subsections 3.1–3.3, and the last point will be addressed in section 4.
Derivation of the contraction factor
When Robin transmission conditions are used, the iteration can be written in Laplace variables as
On Ω 1 = ( − ∞ , δ ) \Omega_1=(-\infty,\delta) Ω 1 = ( − ∞ , δ ) :
s E ^ 1 n − ∂ y y E ^ 1 n = 0 , ( ∂ y E ^ 1 n + p E ^ 1 n ) ∣ y = δ = ( ∂ y E ^ 2 n − 1 + p E ^ 2 n − 1 ) ∣ y = δ . s\widehat{E}_1^n-\partial_{yy}\widehat{E}_1^n=0,\qquad\left.(\partial_y\widehat{E}_1^n+p\widehat{E}_1^n)\right|_{y=\delta}=\left.(\partial_y\widehat{E}_2^{n-1}+p\widehat{E}_2^{n-1})\right|_{y=\delta}. s E 1 n − ∂ y y E 1 n = 0 , ( ∂ y E 1 n + p E 1 n ) y = δ = ( ∂ y E 2 n − 1 + p E 2 n − 1 ) y = δ .
On Ω 2 = ( 0 , ∞ ) \Omega_2=(0,\infty) Ω 2 = ( 0 , ∞ ) :
s E ^ 2 n − ∂ y y E ^ 2 n = 0 , ( − ∂ y E ^ 2 n + p E ^ 2 n ) ∣ y = 0 = ( − ∂ y E ^ 1 n − 1 + p E ^ 1 n − 1 ) ∣ y = 0 . s\widehat{E}_2^n-\partial_{yy}\widehat{E}_2^n=0,\qquad\left.(-\partial_y\widehat{E}_2^n+p\widehat{E}_2^n)\right|_{y=0}=\left.(-\partial_y\widehat{E}_1^{n-1}+p\widehat{E}_1^{n-1})\right|_{y=0}. s E 2 n − ∂ y y E 2 n = 0 , ( − ∂ y E 2 n + p E 2 n ) y = 0 = ( − ∂ y E 1 n − 1 + p E 1 n − 1 ) y = 0 .
Defining g n ( s ) : = ( ∂ y E 2 n + p E 2 n ) ∣ y = δ g^n(s):=\left.(\partial_yE_2^n+pE_2^n)\right|_{y=\delta} g n ( s ) := ( ∂ y E 2 n + p E 2 n ) ∣ y = δ , a similar calculation as above leads to (cf. [4 ] )
g n ( s ) = e − 2 n δ s [ p − s p + s ] 2 n g 0 ( s ) . g^n(s)=e^{-2n\delta\sqrt{s}}\left[\frac{p-\sqrt{s}}{p+\sqrt{s}}\right]^{2n}g^0(s). g n ( s ) = e − 2 n δ s [ p + s p − s ] 2 n g 0 ( s ) .
Thus, the n n n -step contraction factor for each r ≥ 0 r\ge0 r ≥ 0 is given by
ρ ( r , q , n ) = e r T sup ω ∈ R ∣ e − 2 n δ r + i ω [ p − r + i ω p + r + i ω ] 2 n ∣ . \rho(r,q,n)=e^{rT}\sup_{\omega\in\mathbb{R}}\left|e^{-2n\delta\sqrt{r+i\omega}}\left[\frac{p-\sqrt{r+i\omega}}{p+\sqrt{r+i\omega}}\right]^{2n}\right|. ρ ( r , q , n ) = e r T ω ∈ R sup e − 2 n δ r + iω [ p + r + iω p − r + iω ] 2 n .
Writing r + i ω = x + i b \sqrt{r+i\omega}=x+ib r + iω = x + ib as before, we obtain
ρ ( r , q , n ) = e r T sup ω ∈ R e − 2 n δ x ∣ p − x − i b p + x + i b ∣ 2 n = e r T sup ω ∈ R e − 2 n δ x [ ( p − x ) 2 + b 2 ( p + x ) 2 + b 2 ] n . \rho(r,q,n)=e^{rT}\sup_{\omega\in\mathbb{R}}e^{-2n\delta x}\left|\frac{p-x-ib}{p+x+ib}\right|^{2n}=e^{rT}\sup_{\omega\in\mathbb{R}}e^{-2n\delta x}\left[\frac{(p-x)^2+b^2}{(p+x)^2+b^2}\right]^n. ρ ( r , q , n ) = e r T ω ∈ R sup e − 2 n δ x p + x + ib p − x − ib 2 n = e r T ω ∈ R sup e − 2 n δ x [ ( p + x ) 2 + b 2 ( p − x ) 2 + b 2 ] n .
Note that we can replace b 2 b^2 b 2 by x 2 − r x^2-r x 2 − r , thanks to (8) . Moreover, we have x = ( 1 2 ( r + r 2 + ω 2 ) ) 1 / 2 x=\left(\frac{1}{2}\left(r+\sqrt{r^2+\omega^2}\right)\right)^{1/2} x = ( 2 1 ( r + r 2 + ω 2 ) ) 1/2 , so that x ≥ r x\ge\sqrt{r} x ≥ r for all ω ∈ R \omega\in\mathbb{R} ω ∈ R . Thus, we can rewrite ρ ( r , p , n ) \rho(r,p,n) ρ ( r , p , n ) as
ρ ( r , p , n ) = sup x ≥ r e r T − 2 n δ x [ ( p − x ) 2 + x 2 − r ( p + x ) 2 + x 2 − r ] n . \rho(r,p,n)=\sup_{x\ge\sqrt{r}}e^{rT-2n\delta x}\left[\frac{(p-x)^2+x^2-r}{(p+x)^2+x^2-r}\right]^n. ρ ( r , p , n ) = x ≥ r sup e r T − 2 n δ x [ ( p + x ) 2 + x 2 − r ( p − x ) 2 + x 2 − r ] n .
This motivates us to define the functions
h ( x , r , p , n ) = δ 2 T ( r T − 2 n δ x + n log [ ( p − x ) 2 + x 2 − r ( p + x ) 2 + x 2 − r ] ) , h(x,r,p,n)=\frac{\delta^2}{T}\left(rT-2n\delta x+n\log\left[\frac{(p-x)^2+x^2-r}{(p+x)^2+x^2-r}\right]\right), h ( x , r , p , n ) = T δ 2 ( r T − 2 n δ x + n log [ ( p + x ) 2 + x 2 − r ( p − x ) 2 + x 2 − r ] ) ,
M ( r , p , n ) = sup x ≥ r h ( x , r , p , n ) , M(r,p,n)=\sup_{x\ge\sqrt{r}}h(x,r,p,n), M ( r , p , n ) = x ≥ r sup h ( x , r , p , n ) ,
so that ρ ( r , p , n ) = exp ( M ( r , p , n ) ⋅ T / δ 2 ) \rho(r,p,n)=\exp(M(r,p,n)\cdot T/\delta^2) ρ ( r , p , n ) = exp ( M ( r , p , n ) ⋅ T / δ 2 ) . Note that by letting r ~ = δ 2 r \tilde{r}=\delta^2r r ~ = δ 2 r and ν = n δ 2 / T \nu=n\delta^2/T ν = n δ 2 / T , we can write3
h ( x , r , p , n ) = r ~ − 2 ν δ x + ν log [ ( p − x ) 2 + x 2 − r ~ / δ 2 ( p + x ) 2 + x 2 − r ~ / δ 2 ] = r ~ − 2 ν δ x + ν log [ ( δ p − δ x ) 2 + δ 2 x 2 − r ~ ( δ p + δ x ) 2 + δ 2 x 2 − r ~ ] , \begin{aligned}
h(x,r,p,n)=\tilde{r}-2\nu\delta x+\nu\log\left[\frac{(p-x)^2+x^2-\tilde{r}/\delta^2}{(p+x)^2+x^2-\tilde{r}/\delta^2}\right] \\
&=\tilde{r}-2\nu\delta x+\nu\log\left[\frac{(\delta p-\delta x)^2+\delta^2x^2-\tilde{r}}{(\delta p+\delta x)^2+\delta^2x^2-\tilde{r}}\right],
\end{aligned} h ( x , r , p , n ) = r ~ − 2 ν δ x + ν log [ ( p + x ) 2 + x 2 − r ~ / δ 2 ( p − x ) 2 + x 2 − r ~ / δ 2 ] = r ~ − 2 ν δ x + ν log [ ( δ p + δ x ) 2 + δ 2 x 2 − r ~ ( δ p − δ x ) 2 + δ 2 x 2 − r ~ ] ,
so by letting x ~ = δ x \tilde{x}=\delta x x ~ = δ x and p ~ = δ p \tilde{p}=\delta p p ~ = δ p , we get
h ( x ~ , r ~ , p ~ , ν ) = r ~ − 2 ν x ~ + ν log [ ( p ~ − x ~ ) 2 + x ~ 2 − r ~ ( p ~ + x ~ ) 2 + x ~ 2 − r ~ ] (10) h(\tilde{x},\tilde{r},\tilde{p},\nu)=\tilde{r}-2\nu\tilde{x}+\nu\log\left[\frac{(\tilde{p}-\tilde{x})^2+\tilde{x}^2-\tilde{r}}{(\tilde{p}+\tilde{x})^2+\tilde{x}^2-\tilde{r}}\right] \tag*{(10)} h ( x ~ , r ~ , p ~ , ν ) = r ~ − 2 ν x ~ + ν log [ ( p ~ + x ~ ) 2 + x ~ 2 − r ~ ( p ~ − x ~ ) 2 + x ~ 2 − r ~ ] (10)
With this change of variables, the dependence on the time horizon T T T and the overlap size δ \delta δ has disappeared, since their effect is only present in the final calculation of the contraction factor ρ \rho ρ . We will therefore analyze (10) in the remainder of this section. To lighten the notation, we will use the same name for the functions h h h and M M M for both the original and rescaled variables; the name of the arguments will make it clear which version is being used.
Remark: In section 4, we will look for the best parameter p p p to optimize convergence. We will continue to use the rescaled variables, with the understanding that if the analysis yields p ~ ∗ \tilde{p}^{*} p ~ ∗ as the optimal parameter for ν = n δ 2 / T \nu=n\delta^2/T ν = n δ 2 / T , then one should use p ∗ = p ~ ∗ / δ p^{*}=\tilde{p}^{*}/\delta p ∗ = p ~ ∗ / δ in the implementation.
Calculating the supremum
In this subsection, we assume r ~ \tilde{r} r ~ , p ~ \tilde{p} p ~ and ν \nu ν to be fixed positive parameters and study the supremum
M ( r ~ , p ~ , ν ) : = sup x ~ ≥ r ~ h ( x ~ , r ~ , p ~ , ν ) . (11) M(\tilde{r},\tilde{p},\nu):=\sup_{\tilde{x}\ge\sqrt{\tilde{r}}}h(\tilde{x},\tilde{r},\tilde{p},\nu). \tag*{(11)} M ( r ~ , p ~ , ν ) := x ~ ≥ r ~ sup h ( x ~ , r ~ , p ~ , ν ) . (11)
To calculate this supremum, let us find the value of x ~ ∈ [ r ~ , ∞ ) \tilde{x}\in[\sqrt{\tilde{r}},\infty) x ~ ∈ [ r ~ , ∞ ) that maximizes h ( x ~ , r ~ , p ~ , ν ) h(\tilde{x},\tilde{r},\tilde{p},\nu) h ( x ~ , r ~ , p ~ , ν ) . A direct calculation gives
∂ h ∂ x ~ = 2 κ ( A r ~ − B − ( 2 x ~ 2 − r ~ − p ~ ) 2 ( p ~ 2 − r ~ ) 2 + 4 x ~ 2 ( x ~ 2 − r ~ ) ) (12) \frac{\partial h}{\partial\tilde{x}}=2\kappa\left(\frac{A\tilde{r}-B-(2\tilde{x}^2-\tilde{r}-\tilde{p})^2}{(\tilde{p}^2-\tilde{r})^2+4\tilde{x}^2(\tilde{x}^2-\tilde{r})}\right) \tag*{(12)} ∂ x ~ ∂ h = 2 κ ( ( p ~ 2 − r ~ ) 2 + 4 x ~ 2 ( x ~ 2 − r ~ ) A r ~ − B − ( 2 x ~ 2 − r ~ − p ~ ) 2 ) (12)
where
A = 4 p ~ + 2 p ~ 2 > 0 , B = p ~ 4 + 2 p ~ 3 − p ~ 2 . (13) A=4\tilde{p}+2\tilde{p}^2>0,\qquad B=\tilde{p}^4+2\tilde{p}^3-\tilde{p}^2. \tag*{(13)} A = 4 p ~ + 2 p ~ 2 > 0 , B = p ~ 4 + 2 p ~ 3 − p ~ 2 . (13)
Note that A A A and B B B depending only on p ~ \tilde{p} p ~ . If A r ~ − B ≤ 0 A\tilde{r}-B\leq0 A r ~ − B ≤ 0 , the function x ~ ↦ h ( x , r ~ , p ~ , ν ~ ) \tilde{x}\mapsto h(x,\tilde{r},\tilde{p},\tilde{\nu}) x ~ ↦ h ( x , r ~ , p ~ , ν ~ ) is decreasing everywhere, so the maximum of h h h is achieved at x ~ = r ~ \tilde{x}=\sqrt{\tilde{r}} x ~ = r ~ . On the other hand, if A r ~ − B > 0 A\tilde{r}-B>0 A r ~ − B > 0 , then solving ∂ h ∂ x ~ ( x , r ~ , p ~ , ν ~ ) = 0 \frac{\partial h}{\partial\tilde{x}}(x,\tilde{r},\tilde{p},\tilde{\nu})=0 ∂ x ~ ∂ h ( x , r ~ , p ~ , ν ~ ) = 0 formally leads to the two roots
x ~ − = ( r ~ + p ~ − A r ~ − B 2 ) 1 / 2 , x ~ + = ( r ~ + p ~ + A r ~ − B 2 ) 1 / 2 . (14) \tilde{x}_{-}=\left(\frac{\tilde{r}+\tilde{p}-\sqrt{A\tilde{r}-B}}{2}\right)^{1/2},\qquad\tilde{x}_{+}=\left(\frac{\tilde{r}+\tilde{p}+\sqrt{A\tilde{r}-B}}{2}\right)^{1/2}. \tag*{(14)} x ~ − = ( 2 r ~ + p ~ − A r ~ − B ) 1/2 , x ~ + = ( 2 r ~ + p ~ + A r ~ − B ) 1/2 . (14)
We see that x ~ + \tilde{x}_{+} x ~ + is always real, whereas x ~ − \tilde{x}_{-} x ~ − may be real or purely imaginary. Since ∂ h ∂ x ~ \frac{\partial h}{\partial\tilde{x}} ∂ x ~ ∂ h is positive for x ~ − < x ~ < x ~ + \tilde{x}_{-}<\tilde{x}<\tilde{x}_{+} x ~ − < x ~ < x ~ + (or simply x ~ < x ~ + \tilde{x}<\tilde{x}_{+} x ~ < x ~ + when x ~ − \tilde{x}_{-} x ~ − is purely imaginary) and negative elsewhere, x ~ ↦ h ( x , r ~ , p ~ , ν ~ ) \tilde{x}\mapsto h(x,\tilde{r},\tilde{p},\tilde{\nu}) x ~ ↦ h ( x , r ~ , p ~ , ν ~ ) is decreasing for x ~ > x ~ + \tilde{x}>\tilde{x}_{+} x ~ > x ~ + and has a local maximum at x ~ + \tilde{x}_{+} x ~ + , see Figure 3.1 . (The critical point x ~ − \tilde{x}_{-} x ~ − , when it is real, is a local minimum.) Therefore, on the interval x ~ ∈ [ r ~ , ∞ ) \tilde{x}\in[\sqrt{\tilde{r}},\infty) x ~ ∈ [ r ~ , ∞ ) , h h h can attain its global maximum either at the end-point x ~ = r ~ \tilde{x}=\sqrt{\tilde{r}} x ~ = r ~ , or at the local maximum x ~ = x ~ + \tilde{x}=\tilde{x}_{+} x ~ = x ~ + provided that x ~ + ≥ r ~ \tilde{x}_{+}\geq\sqrt{\tilde{r}} x ~ + ≥ r ~ , i.e., whenever
Figure 3.1. Behaviour of x ~ ↦ h ( x , r ~ , p ~ , ν ~ ) \tilde{x}\mapsto h(x,\tilde{r},\tilde{p},\tilde{\nu}) x ~ ↦ h ( x , r ~ , p ~ , ν ~ ) for two different values of r ~ \tilde{r} r ~ . Left: A r ~ − B > 0 A\tilde{r}-B>0 A r ~ − B > 0 , right: A r ~ − B < 0 A\tilde{r}-B<0 A r ~ − B < 0 . Note that x ~ ± \tilde{x}_{\pm} x ~ ± and H ^ ( r ~ , p ~ , ν ~ ) \widehat{H}(\tilde{r},\tilde{p},\tilde{\nu}) H ( r ~ , p ~ , ν ~ ) are not defined when A r ~ − B A\tilde{r}-B A r ~ − B is negative.
1 2 ( r ~ + p ~ + A r ~ − B ) ≥ r ~ ⟺ A r ~ − B ≥ r ~ − p ~ . (15) \frac{1}{2}\left(\tilde{r}+\tilde{p}+\sqrt{A\tilde{r}-B}\right)\geq\tilde{r}\quad\Longleftrightarrow\quad\sqrt{A\tilde{r}-B}\geq\tilde{r}-\tilde{p}. \tag*{(15)} 2 1 ( r ~ + p ~ + A r ~ − B ) ≥ r ~ ⟺ A r ~ − B ≥ r ~ − p ~ . (15)
Squaring both sides of (15) leads to A r ~ − B ≥ ( r ~ − p ~ ) 2 A\tilde{r}-B\geq(\tilde{r}-\tilde{p})^2 A r ~ − B ≥ ( r ~ − p ~ ) 2 , which holds when r ~ ∈ [ r ~ − , r ~ + ] \tilde{r}\in[\tilde{r}_{-},\tilde{r}_{+}] r ~ ∈ [ r ~ − , r ~ + ] , where A r ~ ± − B = ( r ~ ± − p ~ ) 2 A\tilde{r}_{\pm}-B=(\tilde{r}_{\pm}-\tilde{p})^2 A r ~ ± − B = ( r ~ ± − p ~ ) 2 , i.e.,
r ~ ± : = p ~ + A 2 ± A 2 4 + A p ~ − B . (16) \tilde{r}_{\pm}:=\tilde{p}+\frac{A}{2}\pm\sqrt{\frac{A^2}{4}+A\tilde{p}-B}. \tag*{(16)} r ~ ± := p ~ + 2 A ± 4 A 2 + A p ~ − B . (16)
Note that r ~ ± \tilde{r}_{\pm} r ~ ± are real (since 1 4 A 2 − B > 0 \frac{1}{4}A^2-B>0 4 1 A 2 − B > 0 for all p ~ > 0 \tilde{p}>0 p ~ > 0 ) and satisfy r ~ + ≥ r ~ − ≥ B / A \tilde{r}_{+}\geq\tilde{r}_{-}\geq B/A r ~ + ≥ r ~ − ≥ B / A , since
A r ~ − − B = ( r ~ − − p ~ ) 2 ≥ 0 ⟹ r ~ − ≥ B / A . (17) A\tilde{r}_{-}-B=(\tilde{r}_{-}-\tilde{p})^2\geq0\quad\Longrightarrow\quad\tilde{r}_{-}\geq B/A. \tag*{(17)} A r ~ − − B = ( r ~ − − p ~ ) 2 ≥ 0 ⟹ r ~ − ≥ B / A . (17)
Thus, the solution set of (15) contains at least the interval [ r ~ − , r ~ + ] [\tilde{r}_{-},\tilde{r}_{+}] [ r ~ − , r ~ + ] . Additionally, if the interval I = [ B / A , p ~ ) I=[B/A,\tilde{p}) I = [ B / A , p ~ ) is non-empty, then the unsquared inequality (15) will also be satisfied for any r ~ ∈ I \tilde{r}\in I r ~ ∈ I , since such an r ~ \tilde{r} r ~ satisfies A r ~ − B ≥ 0 > r ~ − p ~ A\tilde{r}-B\geq0>\tilde{r}-\tilde{p} A r ~ − B ≥ 0 > r ~ − p ~ . Indeed, I I I is non-empty if and only if
B A < p ~ ⟺ 0 < A p ~ − B = 5 p ~ 2 − p ~ 4 ⟺ p ~ < 5 . (18) \frac{B}{A}<\tilde{p}\quad\Longleftrightarrow\quad0<A\tilde{p}-B=5\tilde{p}^{2}-\tilde{p}^{4}\quad\Longleftrightarrow\quad\tilde{p}<\sqrt{5}. \tag*{(18)} A B < p ~ ⟺ 0 < A p ~ − B = 5 p ~ 2 − p ~ 4 ⟺ p ~ < 5 . (18)
Thus, when p ~ < 5 \tilde{p}<\sqrt{5} p ~ < 5 , the solution set of (15) is [ B / A , p ~ ) ∪ [ r ~ − , r ~ + ] = [ B / A , r ~ + ] [B/A,\tilde{p})\cup[\tilde{r}_{-},\tilde{r}_{+}]=[B/A,\tilde{r}_{+}] [ B / A , p ~ ) ∪ [ r ~ − , r ~ + ] = [ B / A , r ~ + ] , since we have B / A < r ~ − < p ~ B/A<\tilde{r}_{-}<\tilde{p} B / A < r ~ − < p ~ in this case. On the contrary, if p ~ ≥ 5 \tilde{p}\geq\sqrt{5} p ~ ≥ 5 , then r ~ − p ~ ≥ 0 \tilde{r}-\tilde{p}\geq0 r ~ − p ~ ≥ 0 whenever A r ~ − B ≥ 0 A\tilde{r}-B \ge0 A r ~ − B ≥ 0 , since p ~ ≤ B / A \tilde{p} \le B/A p ~ ≤ B / A ; in this case, squaring the inequality (15) does not introduce additional solutions, so the solution set is simply [ r ~ − , r ~ + ] [\tilde{r}_{-},\tilde{r}_{+}] [ r ~ − , r ~ + ] .
To summarize, let us define4
H 0 ( r ~ , p ~ , ν ) = h ( r ~ , r ~ , p ~ , ν ) , H ^ ( r ~ , p ~ , ν ) = h ( x ~ + ( r ~ ) , r ~ , p ~ , ν ) . H_{0}(\tilde{r},\tilde{p},\nu)=h(\sqrt{\tilde{r}},\tilde{r},\tilde{p},\nu), \qquad\hat{H}(\tilde{r},\tilde{p},\nu)=h(\tilde{x}_{+}(\tilde{r}),\tilde{r},\tilde{p},\nu). H 0 ( r ~ , p ~ , ν ) = h ( r ~ , r ~ , p ~ , ν ) , H ^ ( r ~ , p ~ , ν ) = h ( x ~ + ( r ~ ) , r ~ , p ~ , ν ) .
It follows that for each r ~ ≥ 0 \tilde{r}\ge0 r ~ ≥ 0 , the supremum in (11) is given by
M ( r ~ , p ~ , ν ) = { max ( H 0 ( r ~ , p ~ , ν ) , H ^ ( r ~ , p ~ , ν ) ) , r ~ min ≤ r ~ ≤ r ~ + , H 0 ( r ~ , p ~ , ν ) , otherwise . (19) M(\tilde{r},\tilde{p},\nu)=
\begin{cases}
\max\left(H_{0}(\tilde{r},\tilde{p},\nu),\hat{H}(\tilde{r},\tilde{p},\nu)\right), & \tilde{r}_{\min}\le\tilde{r}\le\tilde{r}_{+},\\
H_{0}(\tilde{r},\tilde{p},\nu), & \text{otherwise}.
\end{cases} \tag*{(19)} M ( r ~ , p ~ , ν ) = { max ( H 0 ( r ~ , p ~ , ν ) , H ^ ( r ~ , p ~ , ν ) ) , H 0 ( r ~ , p ~ , ν ) , r ~ m i n ≤ r ~ ≤ r ~ + , otherwise . (19)
with
r ~ min = { max ( 0 , B / A ) , p ~ < 5 , r ~ − , p ~ ≥ 5 . (20) \tilde{r}_{\min}=
\begin{cases}
\max(0,B/A), & \tilde{p}<\sqrt{5},\\
\tilde{r}_{-}, & \tilde{p}\ge\sqrt{5}.
\end{cases} \tag*{(20)} r ~ m i n = { max ( 0 , B / A ) , r ~ − , p ~ < 5 , p ~ ≥ 5 . (20)
where r ~ ± \tilde{r}_{\pm} r ~ ± are defined in (16) . Note that there is no need to include the additional restrictions r ~ min ≥ 0 \tilde{r}_{\min}\ge0 r ~ m i n ≥ 0 nor r ~ min ≥ B / A \tilde{r}_{\min}\ge B/A r ~ m i n ≥ B / A in the case of p ~ ≥ 5 \tilde{p}\ge\sqrt{5} p ~ ≥ 5 , since r ~ − r ~ + = B + p ~ 2 = p ~ 4 + 2 p ~ 3 > 0 \tilde{r}_{-}\tilde{r}_{+}=B+\tilde{p}^{2}=\tilde{p}^{4}+2\tilde{p}^{3}>0 r ~ − r ~ + = B + p ~ 2 = p ~ 4 + 2 p ~ 3 > 0 and r ~ + > 0 \tilde{r}_{+}>0 r ~ + > 0 , so we automatically have r ~ min ≥ 0 \tilde{r}_{\min}\ge0 r ~ m i n ≥ 0 ; moreover, r ~ − \tilde{r}_{-} r ~ − satisfies A r ~ − − B ≥ ( r ~ − − p ~ ) 2 ≥ 0 A\tilde{r}_{-}-B\ge(\tilde{r}_{-}-\tilde{p})^{2}\ge0 A r ~ − − B ≥ ( r ~ − − p ~ ) 2 ≥ 0 , so we must have r ~ − ≥ B / A \tilde{r}_{-}\ge B/A r ~ − ≥ B / A . Also note that when p ~ = 5 \tilde{p}=\sqrt{5} p ~ = 5 , we have r ~ − = B / A > 0 \tilde{r}_{-}=B/A>0 r ~ − = B / A > 0 , so r ~ min \tilde{r}_{\min} r ~ m i n is a continuous function of p ~ \tilde{p} p ~ .
The problem of calculating the supremum is therefore reduced to comparing H 0 H_{0} H 0 and H ^ \hat{H} H ^ inside the interval [ r ~ min , r ~ + ] [\tilde{r}_{\min},\tilde{r}_{+}] [ r ~ m i n , r ~ + ] and taking the larger of the two. The next lemma allows us to characterize more precisely the maximum in (19) .
Lemma 3.1. Let ν \nu ν and p ~ \tilde{p} p ~ be positive real constants. Then the curves y = H 0 ( r ~ , p ~ , ν ) y=H_{0}(\tilde{r},\tilde{p},\nu) y = H 0 ( r ~ , p ~ , ν ) and y = H ^ ( r ~ , p ~ , ν ) y=\hat{H}(\tilde{r},\tilde{p},\nu) y = H ^ ( r ~ , p ~ , ν ) are tangent at r ~ = r ~ + \tilde{r}=\tilde{r}_{+} r ~ = r ~ + , and the same is true at r ~ = r ~ − \tilde{r}=\tilde{r}_{-} r ~ = r ~ − if p ~ ≥ 5 \tilde{p}\ge\sqrt{5} p ~ ≥ 5 . Moreover, for any p ~ > 0 \tilde{p}>0 p ~ > 0 , we have H 0 ( r ~ , p ~ , ν ) < H ^ ( r ~ , p ~ , ν ) H_{0}(\tilde{r},\tilde{p},\nu)<\hat{H}(\tilde{r},\tilde{p},\nu) H 0 ( r ~ , p ~ , ν ) < H ^ ( r ~ , p ~ , ν ) for r ~ ∈ ( r ~ − , r ~ + ) \tilde{r}\in(\tilde{r}_{-},\tilde{r}_{+}) r ~ ∈ ( r ~ − , r ~ + ) .
Proof. We have already shown in (17) that r ~ − ≥ B / A \tilde{r}_{-}\ge B/A r ~ − ≥ B / A . Since r ~ + ≥ r ~ − \tilde{r}_{+}\ge\tilde{r}_{-} r ~ + ≥ r ~ − , we deduce that A r ~ ± − B ≥ 0 A\tilde{r}_{\pm}-B\ge0 A r ~ ± − B ≥ 0 , and x ~ + ( r ~ ± ) \tilde{x}_{+}(\tilde{r}_{\pm}) x ~ + ( r ~ ± ) is well defined. The definition of r ~ ± \tilde{r}_{\pm} r ~ ± implies A r ~ ± − B = ∣ r ~ ± − p ~ ∣ \sqrt{A\tilde{r}_{\pm}-B}=\lvert\tilde{r}_{\pm}-\tilde{p}\rvert A r ~ ± − B = ∣ r ~ ± − p ~ ∣ . Since r ~ + ≥ p ~ \tilde{r}_{+}\ge\tilde{p} r ~ + ≥ p ~ , we must have
A r ~ + − B = r ~ + − p ~ ⟹ (3.5) x ~ + ( r ~ + ) = r ~ + . \sqrt{A\tilde{r}_{+}-B}=\tilde{r}_{+}-\tilde{p}\ \overset{\text{(3.5)}}{\Longrightarrow}\ \tilde{x}_{+}(\tilde{r}_{+})=\sqrt{\tilde{r}_{+}}. A r ~ + − B = r ~ + − p ~ ⟹ (3.5) x ~ + ( r ~ + ) = r ~ + .
Thus, H 0 ( r ~ + , p ~ , ν ) = H ^ ( r ~ + , p ~ , ν ) H_{0}(\tilde{r}_{+},\tilde{p},\nu)=\hat{H}(\tilde{r}_{+},\tilde{p},\nu) H 0 ( r ~ + , p ~ , ν ) = H ^ ( r ~ + , p ~ , ν ) . In addition, since x ~ + ( r ) \tilde{x}_{+}(r) x ~ + ( r ) is a root of ∂ h ∂ x ~ \frac{\partial h}{\partial\tilde{x}} ∂ x ~ ∂ h by definition, we have
∂ H ^ ∂ r ~ ( r ~ + , p ~ , ν ) = ∂ h ∂ r ~ ( x ~ + ( r ~ + ) , r ~ + , p ~ , ν ) + ∂ h ∂ x ~ ( x ~ + ( r ~ + ) , r ~ + , p ~ , ν ) ⋅ d x ~ + d r ~ ( r ~ + ) = ∂ h ∂ r ~ ( r ~ + , r ~ + , p ~ , ν ) + ∂ h ∂ x ~ ( r ~ + , r ~ + , p ~ , ν ) ⋅ 1 2 r ~ + = ∂ H 0 ∂ r ~ ( r ~ + , p ~ , ν ) . \frac{\partial\hat{H}}{\partial\tilde{r}}(\tilde{r}_{+},\tilde{p},\nu)
=
\frac{\partial h}{\partial\tilde{r}}(\tilde{x}_{+}(\tilde{r}_{+}),\tilde{r}_{+},\tilde{p},\nu)
+
\frac{\partial h}{\partial\tilde{x}}(\tilde{x}_{+}(\tilde{r}_{+}),\tilde{r}_{+},\tilde{p},\nu)\cdot\frac{d\tilde{x}_{+}}{d\tilde{r}}(\tilde{r}_{+})
=
\frac{\partial h}{\partial\tilde{r}}(\sqrt{\tilde{r}_{+}},\tilde{r}_{+},\tilde{p},\nu)
+
\frac{\partial h}{\partial\tilde{x}}(\sqrt{\tilde{r}_{+}},\tilde{r}_{+},\tilde{p},\nu)\cdot\frac{1}{2\sqrt{\tilde{r}_{+}}}
=
\frac{\partial H_{0}}{\partial\tilde{r}}(\tilde{r}_{+},\tilde{p},\nu). ∂ r ~ ∂ H ^ ( r ~ + , p ~ , ν ) = ∂ r ~ ∂ h ( x ~ + ( r ~ + ) , r ~ + , p ~ , ν ) + ∂ x ~ ∂ h ( x ~ + ( r ~ + ) , r ~ + , p ~ , ν ) ⋅ d r ~ d x ~ + ( r ~ + ) = ∂ r ~ ∂ h ( r ~ + , r ~ + , p ~ , ν ) + ∂ x ~ ∂ h ( r ~ + , r ~ + , p ~ , ν ) ⋅ 2 r ~ + 1 = ∂ r ~ ∂ H 0 ( r ~ + , p ~ , ν ) .
Therefore, the curves y = H 0 ( p ~ , r ~ , ν ) y=H_{0}(\tilde{p},\tilde{r},\nu) y = H 0 ( p ~ , r ~ , ν ) and y = H ^ ( p ~ , r ~ , ν ) y=\hat{H}(\tilde{p},\tilde{r},\nu) y = H ^ ( p ~ , r ~ , ν ) are in fact tangent at r = r ~ + r=\tilde{r}_{+} r = r ~ + .
If p ~ ≥ 5 \tilde{p}\ge\sqrt{5} p ~ ≥ 5 , then A p ~ − B ≤ 0 A\tilde{p}-B\le0 A p ~ − B ≤ 0 by (18) , which implies r ~ − ≥ p ~ \tilde{r}_{-}\ge\tilde{p} r ~ − ≥ p ~ by (16) . Thus, we again have x ~ + ( r ~ − ) = r ~ − \tilde{x}_{+}(\tilde{r}_{-})=\sqrt{\tilde{r}_{-}} x ~ + ( r ~ − ) = r ~ − , so tangency also holds at r ~ = r ~ − \tilde{r}=\tilde{r}_{-} r ~ = r ~ − .
Finally, we show that H ^ ( r ~ , p ~ , ν ) > H 0 ( r ~ , p ~ , ν ) \hat{H}(\tilde{r},\tilde{p},\nu)>H_{0}(\tilde{r},\tilde{p},\nu) H ^ ( r ~ , p ~ , ν ) > H 0 ( r ~ , p ~ , ν ) for r ~ ∈ ( r ~ − , r ~ + ) \tilde{r}\in(\tilde{r}_{-},\tilde{r}_{+}) r ~ ∈ ( r ~ − , r ~ + ) , regardless of the value of p ~ \tilde{p} p ~ . First, substituting x ~ = r ~ \tilde{x}=\sqrt{\tilde{r}} x ~ = r ~ into (12) leads to
∂ h ∂ x ~ ( r ~ , r ~ , p ~ , ν ) = 2 ν ( A r ~ − B − ( r ~ − p ~ ) 2 ( p ~ 2 − r ~ ) 2 ) . \frac{\partial h}{\partial\tilde{x}}(\sqrt{\tilde{r}},\tilde{r},\tilde{p},\nu)
=
2\nu\left(\frac{A\tilde{r}-B-(\tilde{r}-\tilde{p})^{2}}{(\tilde{p}^{2}-\tilde{r})^{2}}\right). ∂ x ~ ∂ h ( r ~ , r ~ , p ~ , ν ) = 2 ν ( ( p ~ 2 − r ~ ) 2 A r ~ − B − ( r ~ − p ~ ) 2 ) .
By the definition of r ~ ± \tilde r_{\pm} r ~ ± , we have A r ~ − B > ( r ~ − p ~ ) 2 A\tilde r-B>(\tilde r-\tilde p)^2 A r ~ − B > ( r ~ − p ~ ) 2 whenever r ~ ∈ ( r ~ − , r ~ + ) \tilde r\in(\tilde r_{-},\tilde r_{+}) r ~ ∈ ( r ~ − , r ~ + ) . Hence, x ~ ↦ h ( x ~ , r ~ , p ~ , ν ) \tilde x\mapsto h(\tilde x,\tilde r,\tilde p,\nu) x ~ ↦ h ( x ~ , r ~ , p ~ , ν ) is strictly increasing in a neighbourhood of x ~ = r ~ \tilde x=\sqrt{\tilde r} x ~ = r ~ , meaning that h ( ⋅ , r ~ , p ~ , ν ) h(\cdot,\tilde r,\tilde p,\nu) h ( ⋅ , r ~ , p ~ , ν ) cannot achieve its maximum at x ~ = r ~ \tilde x=\sqrt{\tilde r} x ~ = r ~ . Therefore, the maximum must occur at x ~ + ( r ) \tilde x_{+}(r) x ~ + ( r ) , so we must have H ^ ( r ~ , p ~ , ν ) > H 0 ( r ~ , p ~ , ν ) \widehat H(\tilde r,\tilde p,\nu)>H_{0}(\tilde r,\tilde p,\nu) H ( r ~ , p ~ , ν ) > H 0 ( r ~ , p ~ , ν ) .
Using the definitions of A A A and B B B in (13) , one can verify by direct calculation that r ~ − ≤ p ~ 2 ≤ r ~ + \tilde r_{-}\leq\tilde p^{2}\leq\tilde r_{+} r ~ − ≤ p ~ 2 ≤ r ~ + . The above lemma implies that when r ~ = p ~ 2 \tilde r=\tilde p^{2} r ~ = p ~ 2 , the supremum is always given by H ^ ( r ~ , p ~ , ν ) \widehat H(\tilde r,\tilde p,\nu) H ( r ~ , p ~ , ν ) , so the fact that H 0 ( r ~ , p ~ , ν ) = − ∞ H_{0}(\tilde r,\tilde p,\nu)=-\infty H 0 ( r ~ , p ~ , ν ) = − ∞ when r ~ = p ~ 2 \tilde r=\tilde p^{2} r ~ = p ~ 2 poses no difficulty. It also implies that when p ~ ≥ 5 \tilde p\geq\sqrt{5} p ~ ≥ 5 , the maximum in (19) is always given by H ^ \widehat H H . We now treat the case of p ~ < 5 \tilde p<\sqrt{5} p ~ < 5 .
Lemma 3.2. Let 0 < p ~ < 5 0<\tilde p<\sqrt{5} 0 < p ~ < 5 . Then H 0 ( r ~ min , p ~ , ν ) ≥ H ^ ( r ~ min , p ~ , ν ) H_{0}(\tilde r_{\min},\tilde p,\nu)\geq\widehat H(\tilde r_{\min},\tilde p,\nu) H 0 ( r ~ m i n , p ~ , ν ) ≥ H ( r ~ m i n , p ~ , ν ) . Moreover, the function r ~ ↦ H ^ ( r ~ , p ~ , ν ) − H 0 ( r ~ , p ~ , ν ) \tilde r\mapsto\widehat H(\tilde r,\tilde p,\nu)-H_{0}(\tilde r,\tilde p,\nu) r ~ ↦ H ( r ~ , p ~ , ν ) − H 0 ( r ~ , p ~ , ν ) is strictly increasing on r ~ ∈ [ r ~ min , r ~ − ] \tilde r\in[\tilde r_{\min},\tilde r_{-}] r ~ ∈ [ r ~ m i n , r ~ − ] , and there is a unique crossing point r ~ ∗ ∈ [ r ~ min , r ~ − ] \tilde r^{*}\in[\tilde r_{\min},\tilde r_{-}] r ~ ∗ ∈ [ r ~ m i n , r ~ − ] at which H 0 ( r ~ ∗ , p ~ , ν ) = H ^ ( r ~ ∗ , p ~ , ν ) H_{0}(\tilde r^{*},\tilde p,\nu)=\widehat H(\tilde r^{*},\tilde p,\nu) H 0 ( r ~ ∗ , p ~ , ν ) = H ( r ~ ∗ , p ~ , ν ) .
Proof. We have either r ~ min = B / A \tilde r_{\min}=B/A r ~ m i n = B / A or r ~ min = 0 \tilde r_{\min}=0 r ~ m i n = 0 , depending on the sign of B B B . If B > 0 B>0 B > 0 , then r ~ min = B / A \tilde r_{\min}=B/A r ~ m i n = B / A , and we have
∂ h ∂ x ~ ( r ~ min , r ~ min , p ~ , ν ) = 2 ν ( 0 − ( r ~ min − p ~ ) 2 ( p ~ 2 − r ~ min ) 2 ) < 0 , \frac{\partial h}{\partial\tilde x}\left(\sqrt{\tilde r_{\min}},\tilde r_{\min},\tilde p,\nu\right)=2\nu\left(\frac{0-(\tilde r_{\min}-\tilde p)^2}{(\tilde p^2-\tilde r_{\min})^2}\right)<0, ∂ x ~ ∂ h ( r ~ m i n , r ~ m i n , p ~ , ν ) = 2 ν ( ( p ~ 2 − r ~ m i n ) 2 0 − ( r ~ m i n − p ~ ) 2 ) < 0 ,
so that x ~ ↦ h ( x ~ , r ~ min , p ~ , ν ) \tilde x\mapsto h(\tilde x,\tilde r_{\min},\tilde p,\nu) x ~ ↦ h ( x ~ , r ~ m i n , p ~ , ν ) is decreasing over all x ~ ∈ [ r ~ min , ∞ ) \tilde x\in[\tilde r_{\min},\infty) x ~ ∈ [ r ~ m i n , ∞ ) . Therefore, we must have H ^ ( r ~ min , p ~ , ν ) = h ( x ~ + ( r ~ min ) , r ~ min , p ~ , ν ) ≤ h ( r ~ min , r ~ min , p ~ , ν ) = H 0 ( r ~ min , p ~ , ν ) \widehat H(\tilde r_{\min},\tilde p,\nu)=h(\tilde x_{+}(\tilde r_{\min}),\tilde r_{\min},\tilde p,\nu)\leq h(\tilde r_{\min},\tilde r_{\min},\tilde p,\nu)=H_{0}(\tilde r_{\min},\tilde p,\nu) H ( r ~ m i n , p ~ , ν ) = h ( x ~ + ( r ~ m i n ) , r ~ m i n , p ~ , ν ) ≤ h ( r ~ m i n , r ~ m i n , p ~ , ν ) = H 0 ( r ~ m i n , p ~ , ν ) . If r ~ min = 0 \tilde r_{\min}=0 r ~ m i n = 0 , i.e., if B ≤ 0 B\leq0 B ≤ 0 , then a direct calculation shows for all x ~ ≥ 0 \tilde x\geq0 x ~ ≥ 0 that
h ( x ~ , 0 , p ~ , ν ) = − 2 ν x ~ + ν log ( ( p ~ − x ~ ) 2 + x ~ 2 ( p ~ + x ~ ) 2 + x ~ 2 ) ≤ 0 = h ( 0 , 0 , p ~ , ν ) , h(\tilde x,0,\tilde p,\nu)=-2\nu\tilde x+\nu\log\left(\frac{(\tilde p-\tilde x)^2+\tilde x^2}{(\tilde p+\tilde x)^2+\tilde x^2}\right)\leq0=h(0,0,\tilde p,\nu), h ( x ~ , 0 , p ~ , ν ) = − 2 ν x ~ + ν log ( ( p ~ + x ~ ) 2 + x ~ 2 ( p ~ − x ~ ) 2 + x ~ 2 ) ≤ 0 = h ( 0 , 0 , p ~ , ν ) ,
since the argument of the logarithm is less than or equal to 1 for any positive p ~ \tilde p p ~ and non-negative x ~ \tilde x x ~ . Letting x ~ = x ~ + ( 0 ) \tilde x=\tilde x_{+}(0) x ~ = x ~ + ( 0 ) shows that H ^ ( 0 , p ~ , ν ) ≤ H 0 ( 0 , p ~ , ν ) \widehat H(0,\tilde p,\nu)\leq H_{0}(0,\tilde p,\nu) H ( 0 , p ~ , ν ) ≤ H 0 ( 0 , p ~ , ν ) , as required. Finally, Lemma 3.1 shows that H ^ ( r ~ − , p ~ , ν ) ≥ H 0 ( r ~ − , p ~ , ν ) \widehat H(\tilde r_{-},\tilde p,\nu)\geq H_{0}(\tilde r_{-},\tilde p,\nu) H ( r ~ − , p ~ , ν ) ≥ H 0 ( r ~ − , p ~ , ν ) ; since both r ~ ↦ H ^ ( r ~ , p ~ , ν ) \tilde r\mapsto\widehat H(\tilde r,\tilde p,\nu) r ~ ↦ H ( r ~ , p ~ , ν ) and r ~ ↦ H 0 ( r ~ , p ~ , ν ) \tilde r\mapsto H_{0}(\tilde r,\tilde p,\nu) r ~ ↦ H 0 ( r ~ , p ~ , ν ) are continuous functions, the intermediate value theorem implies the curves must cross at least once at r ~ ∗ \tilde r^{*} r ~ ∗ , somewhere between r ~ min \tilde r_{\min} r ~ m i n and r ~ − \tilde r_{-} r ~ − .
We now show that r ~ ↦ H ^ ( r ~ , p ~ , ν ) − H 0 ( r ~ , p ~ , ν ) \tilde r\mapsto\widehat H(\tilde r,\tilde p,\nu)-H_{0}(\tilde r,\tilde p,\nu) r ~ ↦ H ( r ~ , p ~ , ν ) − H 0 ( r ~ , p ~ , ν ) is strictly increasing on r ~ ∈ [ r ~ min , r ~ − ] \tilde r\in[\tilde r_{\min},\tilde r_{-}] r ~ ∈ [ r ~ m i n , r ~ − ] , which implies that the function cannot vanish for more than one value of r ~ \tilde r r ~ . A direct calculation shows that
∂ H ^ ∂ r ~ ( r ~ , p ~ , ν ) − ∂ H 0 ∂ r ~ ( r ~ , p ~ , ν ) = ν r ~ + 2 ν p ~ r ~ ( p ~ 2 − r ~ ) − 4 ν p ~ x ~ + G , (21) \frac{\partial\widehat H}{\partial\tilde r}(\tilde r,\tilde p,\nu)-\frac{\partial H_{0}}{\partial\tilde r}(\tilde r,\tilde p,\nu)=\frac{\nu}{\sqrt{\tilde r}}+\frac{2\nu\tilde p}{\sqrt{\tilde r}(\tilde p^{2}-\tilde r)}-\frac{4\nu\tilde p\tilde x_{+}}{G}, \tag*{(21)} ∂ r ~ ∂ H ( r ~ , p ~ , ν ) − ∂ r ~ ∂ H 0 ( r ~ , p ~ , ν ) = r ~ ν + r ~ ( p ~ 2 − r ~ ) 2 ν p ~ − G 4 ν p ~ x ~ + , (21)
where x ~ + = x ~ + ( r ~ ) \tilde x_{+}=\tilde x_{+}(\tilde r) x ~ + = x ~ + ( r ~ ) and G = G ( x ~ + , r ~ , p ~ ) = ( p ~ 2 − r ~ ) 2 + 4 x ~ + 2 ( x ~ + 2 − r ~ ) > 0 G=G(\tilde x_{+},\tilde r,\tilde p)=(\tilde p^{2}-\tilde r)^{2}+4\tilde x_{+}^{2}(\tilde x_{+}^{2}-\tilde r)>0 G = G ( x ~ + , r ~ , p ~ ) = ( p ~ 2 − r ~ ) 2 + 4 x ~ + 2 ( x ~ + 2 − r ~ ) > 0 , since x ~ + ≥ r ~ \tilde x_{+}\geq\sqrt{\tilde r} x ~ + ≥ r ~ . Since x ~ + \tilde x_{+} x ~ + is a local maximum of x ~ ↦ h ( p ~ , r ~ , x ~ ) \tilde x\mapsto h(\tilde p,\tilde r,\tilde x) x ~ ↦ h ( p ~ , r ~ , x ~ ) , we have
0 = ∂ h ∂ x ~ ( x ~ + , r ~ , p ~ , ν ) = − 2 ν − 4 ν p ~ ( p ~ 2 − 2 x ~ + 2 − r ~ ) G ⟹ G = 2 p ~ ( 2 x ~ + 2 + r ~ − p ~ 2 ) . 0=\frac{\partial h}{\partial\tilde x}(\tilde x_{+},\tilde r,\tilde p,\nu)=-2\nu-\frac{4\nu\tilde p(\tilde p^{2}-2\tilde x_{+}^{2}-\tilde r)}{G}\quad\Longrightarrow\quad G=2\tilde p(2\tilde x_{+}^{2}+\tilde r-\tilde p^{2}). 0 = ∂ x ~ ∂ h ( x ~ + , r ~ , p ~ , ν ) = − 2 ν − G 4 ν p ~ ( p ~ 2 − 2 x ~ + 2 − r ~ ) ⟹ G = 2 p ~ ( 2 x ~ + 2 + r ~ − p ~ 2 ) .
Substituting this into (21) leads to
∂ H ^ ∂ r ~ ( r ~ , p ~ , ν ) − ∂ H 0 ∂ r ~ ( r ~ , p ~ , ν ) = 2 ν p ~ G r ~ ( 2 x ~ + 2 + r ~ − p ~ 2 + G p ~ 2 − r ~ − 2 x ~ + r ~ ) = 2 ν p ~ G r ~ ( p ~ 2 − r ~ ) ( G − ( p ~ 2 − r ~ ) 2 + ( 2 x ~ + 2 − 2 x ~ + r ~ ) ( p ~ 2 − r ~ ) ) = 2 ν p ~ G r ~ ( p ~ 2 − r ~ ) ( 4 x ~ + 2 ( x ~ + 2 − r ~ ) + 2 x ~ + ( x ~ + − r ~ ) ( p ~ 2 − r ~ ) ) > 0 , \begin{aligned}
\frac{\partial\widehat H}{\partial\tilde r}(\tilde r,\tilde p,\nu)-\frac{\partial H_{0}}{\partial\tilde r}(\tilde r,\tilde p,\nu)
&=\frac{2\nu\tilde p}{G\sqrt{\tilde r}}\left(2\tilde x_{+}^{2}+\tilde r-\tilde p^{2}+\frac{G}{\tilde p^{2}-\tilde r}-2\tilde x_{+}\sqrt{\tilde r}\right)\\
&=\frac{2\nu\tilde p}{G\sqrt{\tilde r}(\tilde p^{2}-\tilde r)}\left(G-(\tilde p^{2}-\tilde r)^{2}+(2\tilde x_{+}^{2}-2\tilde x_{+}\sqrt{\tilde r})(\tilde p^{2}-\tilde r)\right)\\
&=\frac{2\nu\tilde p}{G\sqrt{\tilde r}(\tilde p^{2}-\tilde r)}\left(4\tilde x_{+}^{2}(\tilde x_{+}^{2}-\tilde r)+2\tilde x_{+}(\tilde x_{+}-\sqrt{\tilde r})(\tilde p^{2}-\tilde r)\right)>0,
\end{aligned} ∂ r ~ ∂ H ( r ~ , p ~ , ν ) − ∂ r ~ ∂ H 0 ( r ~ , p ~ , ν ) = G r ~ 2 ν p ~ ( 2 x ~ + 2 + r ~ − p ~ 2 + p ~ 2 − r ~ G − 2 x ~ + r ~ ) = G r ~ ( p ~ 2 − r ~ ) 2 ν p ~ ( G − ( p ~ 2 − r ~ ) 2 + ( 2 x ~ + 2 − 2 x ~ + r ~ ) ( p ~ 2 − r ~ ) ) = G r ~ ( p ~ 2 − r ~ ) 2 ν p ~ ( 4 x ~ + 2 ( x ~ + 2 − r ~ ) + 2 x ~ + ( x ~ + − r ~ ) ( p ~ 2 − r ~ ) ) > 0 ,
since r ~ ≤ r ~ − < p ~ 2 \tilde r\leq\tilde r_{-}<\tilde p^{2} r ~ ≤ r ~ − < p ~ 2 . Thus, r ~ ↦ H ^ ( p ~ , r ~ ) − H 0 ( p ~ , r ~ ) \tilde r\mapsto\widehat H(\tilde p,\tilde r)-H_{0}(\tilde p,\tilde r) r ~ ↦ H ( p ~ , r ~ ) − H 0 ( p ~ , r ~ ) is strictly increasing, as required.
The following theorem summarizes the behaviour of M ( r ~ , p ~ , ν ) M(\tilde{r},\tilde{p},\nu) M ( r ~ , p ~ , ν ) for all values of p ~ \tilde{p} p ~ .
Theorem 3.3. Let ν > 0 \nu> 0 ν > 0 and p ~ > 0 \tilde{p} > 0 p ~ > 0 be given. Let M ( r ~ , p ~ , ν ) M(\tilde{r},\tilde{p},\nu) M ( r ~ , p ~ , ν ) be the bound on the convergence factor defined in (11) . Then for p ~ ≥ 5 \tilde{p} \ge\sqrt{5} p ~ ≥ 5 , we have
M ( r ~ , p ~ , ν ) = { H ^ ( r ~ , p ~ , ν ) = h ( x ~ + ( r ~ ) , r ~ , p ~ , ν ) , r ~ − ≤ r ~ ≤ r ~ + , H 0 ( r ~ , p ~ , ν ) = h ( r ~ , r ~ , p ~ , ν ) , otherwise , M(\tilde{r},\tilde{p},\nu)=
\begin{cases}
\widehat{H}(\tilde{r},\tilde{p},\nu)=h(\tilde{x}_{+}(\tilde{r}),\tilde{r},\tilde{p},\nu), & \tilde{r}_{-}\le\tilde{r}\le\tilde{r}_{+},\\
H_{0}(\tilde{r},\tilde{p},\nu)=h(\sqrt{\tilde{r}},\tilde{r},\tilde{p},\nu), & \text{otherwise},
\end{cases} M ( r ~ , p ~ , ν ) = { H ( r ~ , p ~ , ν ) = h ( x ~ + ( r ~ ) , r ~ , p ~ , ν ) , H 0 ( r ~ , p ~ , ν ) = h ( r ~ , r ~ , p ~ , ν ) , r ~ − ≤ r ~ ≤ r ~ + , otherwise , where r ~ − \tilde{r}_{-} r ~ − and r ~ + \tilde{r}_{+} r ~ + are defined in (16) . Moreover, r ~ ↦ M ( p ~ , r ~ ) \tilde{r}\mapsto M(\tilde{p},\tilde{r}) r ~ ↦ M ( p ~ , r ~ ) is continuously differentiable on r ~ ∈ [ 0 , ∞ ) \tilde{r}\in[0,\infty) r ~ ∈ [ 0 , ∞ ) . If p ~ < 5 \tilde{p}<\sqrt{5} p ~ < 5 , then
M ( r ~ , p ~ , ν ) = { H ^ ( r ~ , p ~ , ν ) = h ( x ~ + ( r ~ ) , r ~ , p ~ , ν ) , r ~ × ≤ r ~ ≤ r ~ + , H 0 ( r ~ , p ~ , ν ) = h ( r ~ , r ~ , p ~ , ν ) , otherwise , M(\tilde{r},\tilde{p},\nu)=
\begin{cases}
\widehat{H}(\tilde{r},\tilde{p},\nu)=h(\tilde{x}_{+}(\tilde{r}),\tilde{r},\tilde{p},\nu), & \tilde{r}^{\times}\le\tilde{r}\le\tilde{r}_{+},\\
H_{0}(\tilde{r},\tilde{p},\nu)=h(\sqrt{\tilde{r}},\tilde{r},\tilde{p},\nu), & \text{otherwise},
\end{cases} M ( r ~ , p ~ , ν ) = { H ( r ~ , p ~ , ν ) = h ( x ~ + ( r ~ ) , r ~ , p ~ , ν ) , H 0 ( r ~ , p ~ , ν ) = h ( r ~ , r ~ , p ~ , ν ) , r ~ × ≤ r ~ ≤ r ~ + , otherwise , where r ~ × \tilde{r}^{\times} r ~ × is the unique point in [ r ~ min , r ~ − ] [\tilde{r}_{\min},\tilde{r}_{-}] [ r ~ m i n , r ~ − ] that satisfies H ^ ( r ~ × , p ~ , ν ) = H 0 ( r ~ × , p ~ , ν ) \widehat{H}(\tilde{r}^{\times},\tilde{p},\nu)=H_{0}(\tilde{r}^{\times},\tilde{p},\nu) H ( r ~ × , p ~ , ν ) = H 0 ( r ~ × , p ~ , ν ) . Moreover, r ~ ↦ M ( r ~ , p ~ , ν ) \tilde{r}\mapsto M(\tilde{r},\tilde{p},\nu) r ~ ↦ M ( r ~ , p ~ , ν ) is continuous on r ~ ∈ [ 0 , ∞ ) \tilde{r}\in[0,\infty) r ~ ∈ [ 0 , ∞ ) and continuously differentiable everywhere except at r ~ = r ~ × \tilde{r}=\tilde{r}^{\times} r ~ = r ~ × .
Minimizing the bound for fixed p ~ \tilde{p} p ~ and ν \nu ν
We now turn our attention to finding the parameter r ~ \tilde{r} r ~ that yields the sharpest bound, i.e., the one that minimizes M ( r ~ , p ~ , ν ) M(\tilde{r},\tilde{p},\nu) M ( r ~ , p ~ , ν ) . To show the existence of such an r ~ \tilde{r} r ~ , we observe that r ~ ↦ M ( r ~ , p ~ , ν ) \tilde{r}\mapsto M(\tilde{r},\tilde{p},\nu) r ~ ↦ M ( r ~ , p ~ , ν ) is increasing for large enough r ~ \tilde{r} r ~ . Indeed, since
∂ H 0 ∂ r ~ = ∂ ∂ r ~ [ r ~ − 2 ν r ~ + ν log ( ( p ~ − r ~ ) 2 ( p ~ + r ~ ) 2 ) ] = 1 − ν r ~ + 2 ν p ~ r ~ ( r ~ − p ~ 2 ) . (22) \frac{\partial H_{0}}{\partial\tilde{r}}=\frac{\partial}{\partial\tilde{r}}\left[\tilde{r}-2\nu\sqrt{\tilde{r}}+\nu\log\left(\frac{(\tilde{p}-\sqrt{\tilde{r}})^{2}}{(\tilde{p}+\sqrt{\tilde{r}})^{2}}\right)\right]=1-\frac{\nu}{\sqrt{\tilde{r}}}+\frac{2\nu\tilde{p}}{\sqrt{\tilde{r}}(\tilde{r}-\tilde{p}^{2})}. \tag*{(22)} ∂ r ~ ∂ H 0 = ∂ r ~ ∂ [ r ~ − 2 ν r ~ + ν log ( ( p ~ + r ~ ) 2 ( p ~ − r ~ ) 2 ) ] = 1 − r ~ ν + r ~ ( r ~ − p ~ 2 ) 2 ν p ~ . (22)
we have ∂ M ∂ r ~ > 0 \frac{\partial M}{\partial\tilde{r}}>0 ∂ r ~ ∂ M > 0 for all r ~ ≥ max ( ν 2 , r ~ + ) \tilde{r}\ge\max(\nu^{2},\tilde{r}_{+}) r ~ ≥ max ( ν 2 , r ~ + ) . Moreover, r ~ ↦ H 0 ( r ~ , p ~ , ν ) \tilde{r}\mapsto H_{0}(\tilde{r},\tilde{p},\nu) r ~ ↦ H 0 ( r ~ , p ~ , ν ) is decreasing in a neighbourhood containing r ~ = 0 \tilde{r}=0 r ~ = 0 (in fact, ∂ M ∂ r ~ → − ∞ \frac{\partial M}{\partial\tilde{r}}\to-\infty ∂ r ~ ∂ M → − ∞ as r ~ → 0 + \tilde{r}\to0^{+} r ~ → 0 + ). Therefore, r ~ ↦ M ( p ~ , r ~ ) \tilde{r}\mapsto M(\tilde{p},\tilde{r}) r ~ ↦ M ( p ~ , r ~ ) has a global minimum somewhere on the interval ( 0 , max ( ν 2 , r ~ + ) ) (0,\max(\nu^{2},\tilde{r}_{+})) ( 0 , max ( ν 2 , r ~ + )) . If p ~ ≥ 5 \tilde{p}\ge\sqrt{5} p ~ ≥ 5 , the minimum must be either a critical point of H 0 H_{0} H 0 or H ^ \widehat{H} H , since M M M is continuously differentiable. If p ~ < 5 \tilde{p}<\sqrt{5} p ~ < 5 , the minimum can be a critical point, or it can be the intersection point r ~ × ∈ [ r ~ min , r ~ − ] \tilde{r}^{\times}\in[\tilde{r}_{\min},\tilde{r}_{-}] r ~ × ∈ [ r ~ m i n , r ~ − ] , where H 0 ( r ~ × , p ~ , ν ) = H ^ ( r ~ × , p ~ , ν ) H_{0}(\tilde{r}^{\times},\tilde{p},\nu)=\widehat{H}(\tilde{r}^{\times},\tilde{p},\nu) H 0 ( r ~ × , p ~ , ν ) = H ( r ~ × , p ~ , ν ) . At such a point, we have equioscillation between the contraction factor at x ~ = r ~ \tilde{x}=\sqrt{\tilde{r}} x ~ = r ~ and that at the local maximum x ~ = x ~ + ( r ~ ) \tilde{x}=\tilde{x}_{+}(\tilde{r}) x ~ = x ~ + ( r ~ ) . Figure 3.2 illustrates the different possible locations for the minimum, depending on p ~ \tilde{p} p ~ and ν \nu ν . In all but one case, the minimum corresponds to a critical point of H 0 H_{0} H 0 or H ^ \widehat{H} H (or both, when they coincide); the remaining case is characterized by equioscillation, i.e., we have r ~ = r ~ × \tilde{r}=\tilde{r}^{\times} r ~ = r ~ × . This leads us to study the critical points of H 0 H_{0} H 0 and H ^ \widehat{H} H . To study the critical points of H 0 H_{0} H 0 , we let z = r ~ z=\sqrt{\tilde{r}} z = r ~ and rewrite (22) as follows:
Figure 3.2. Different configurations for the r r r -dependent error bound. The bound is defined by the union of the solid curves, and the dashed curves represent portions of H 0 H_0 H 0 or H ^ \hat{H} H ^ that do not contribute to the error bound. The black dot indicates the global minimum of r ↦ M ( r , p , n ) r \mapsto M(r,p,n) r ↦ M ( r , p , n ) . The left column illustrates cases where p ~ = p δ < 5 \tilde{p}=p\delta<\sqrt{5} p ~ = p δ < 5 , while the right column corresponds to p ~ = p δ ≥ 5 \tilde{p}=p\delta\ge\sqrt{5} p ~ = p δ ≥ 5 . Only the bottom left panel is characterized by equioscillation; the other minima are all critical points of H 0 H_0 H 0 or H ^ \hat{H} H ^ (or both, as is the case for the bottom right panel).
∂ H 0 ∂ r ~ = ν p ~ z ( p ~ z ν − [ p ~ + 2 p ~ 2 p ~ 2 − z 2 ] ) ≔ ν p ~ z ( p ~ z ν − F p ~ ( z ) ) . \frac{\partial H_{0}}{\partial\tilde{r}}=\frac{\nu}{\tilde{p}z}\left(\frac{\tilde{p}z}{\nu}-\left[\tilde{p}+\frac{2\tilde{p}^{2}}{\tilde{p}^{2}-z^{2}}\right]\right)\coloneqq\frac{\nu}{\tilde{p}z}\left(\frac{\tilde{p}z}{\nu}-F_{\tilde{p}}(z)\right). ∂ r ~ ∂ H 0 = p ~ z ν ( ν p ~ z − [ p ~ + p ~ 2 − z 2 2 p ~ 2 ] ) : = p ~ z ν ( ν p ~ z − F p ~ ( z ) ) .
where we observe that F p ~ F_{\tilde{p}} F p ~ is independent of ν \nu ν . We now have the following lemma.
Lemma 3.4 (Critical points of H 0 H_{0} H 0 ). Let r ~ − \tilde{r}_{-} r ~ − and r ~ + \tilde{r}_{+} r ~ + be defined as in (16) and ν > 0 \nu>0 ν > 0 .
If ν < p ~ / F p ~ ( r ~ − ) \nu<\tilde{p}/F_{\tilde{p}}(\sqrt{\tilde{r}_{-}}) ν < p ~ / F p ~ ( r ~ − ) , then H 0 H_{0} H 0 has a local minimum r ~ ∗ \tilde{r}_{*} r ~ ∗ and a local maximum r ~ ∗ \tilde{r}^{*} r ~ ∗ with r ~ ∗ < r ~ − < r ~ ∗ < p ~ 2 \tilde{r}_{*}<\tilde{r}_{-}<\tilde{r}^{*}<\tilde{p}^{2} r ~ ∗ < r ~ − < r ~ ∗ < p ~ 2 and no other critical points in [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) .
If ν > p ~ / F p ~ ( r ~ + ) \nu>\tilde{p}/F_{\tilde{p}}(\sqrt{\tilde{r}_{+}}) ν > p ~ / F p ~ ( r ~ + ) , then H 0 H_{0} H 0 has a local minimum r ~ ∗ \tilde{r}_{*} r ~ ∗ and a local maximum r ~ ∗ \tilde{r}^{*} r ~ ∗ with p ~ 2 < r ~ ∗ < r ~ + < r ~ ∗ \tilde{p}^{2}<\tilde{r}^{*}<\tilde{r}_{+}<\tilde{r}_{*} p ~ 2 < r ~ ∗ < r ~ + < r ~ ∗ and no other critical points in [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) .
If ν = p ~ / F p ~ ′ ( r ~ − ) \nu=\tilde p/F'_{\tilde p}(\sqrt{\tilde r_{-}}) ν = p ~ / F p ~ ′ ( r ~ − ) (resp. ν = p ~ / F p ~ ′ ( r ~ + ) \nu=\tilde p/F'_{\tilde p}(\sqrt{\tilde r_{+}}) ν = p ~ / F p ~ ′ ( r ~ + ) ), then H 0 H_0 H 0 has a unique critical point that is also an inflection point at r ~ − \tilde r_{-} r ~ − (resp. r ~ + \tilde r_{+} r ~ + ).
If p ~ / F p ~ ′ ( r ~ − ) < ν < p ~ / F p ~ ′ ( r ~ + ) \tilde p/F'_{\tilde p}(\sqrt{\tilde r_{-}})<\nu<\tilde p/F'_{\tilde p}(\sqrt{\tilde r_{+}}) p ~ / F p ~ ′ ( r ~ − ) < ν < p ~ / F p ~ ′ ( r ~ + ) , then H 0 H_0 H 0 has no critical points in [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) .
Proof. We start by rewriting r ~ ± \tilde r_{\pm} r ~ ± . Substituting the definitions of A A A and B B B in (13) into (16) leads to
r ~ ± = 3 p ~ + p ~ 2 ± ( p ~ 4 + 4 p ~ 3 + 4 p ~ 2 ) + 4 p ~ 2 + 2 p ~ 3 − ( p ~ 4 + 2 p ~ 3 − p ~ 2 ) = p ~ ( 3 + p ~ ± 9 + 4 p ~ ) . (23) \begin{aligned}
\tilde r_{\pm}=3\tilde p+\tilde p^2\pm\sqrt{(\tilde p^4+4\tilde p^3+4\tilde p^2)+4\tilde p^2+2\tilde p^3-(\tilde p^4+2\tilde p^3-\tilde p^2)} \\
&=\tilde p\left(3+\tilde p\pm\sqrt{9+4\tilde p}\right). \tag*{(23)}
\end{aligned} r ~ ± = 3 p ~ + p ~ 2 ± ( p ~ 4 + 4 p ~ 3 + 4 p ~ 2 ) + 4 p ~ 2 + 2 p ~ 3 − ( p ~ 4 + 2 p ~ 3 − p ~ 2 ) = p ~ ( 3 + p ~ ± 9 + 4 p ~ ) . (23)
Next, we consider the zeros of ∂ H 0 ∂ r \frac{\partial H_0}{\partial r} ∂ r ∂ H 0 , which occur when there are intersections between y = F p ~ ( z ) y=F_{\tilde p}(z) y = F p ~ ( z ) and the straight line y = p ~ z / ν y=\tilde p z/\nu y = p ~ z / ν , see Figure 3.3 . As ν \nu ν varies from 0 to ∞ \infty ∞ , we see that the straight line intersects the upper branch of y = F p ~ ( z ) y=F_{\tilde p}(z) y = F p ~ ( z ) if ν \nu ν is small; it intersects the lower branch if ν \nu ν is large, and there are no intersections for intermediate values of ν \nu ν . The transition between these three regimes occurs when the straight line is tangent to one of the branches, see the red lines in Figure 3.3 . Therefore, it suffices to show that these points of tangency occur at z = r ~ ± z=\sqrt{\tilde r_{\pm}} z = r ~ ± . Indeed, for a given z 0 ≠ p ~ z_0\ne\tilde p z 0 = p ~ , the tangent line is of the form
Figure 3.3. Plot of y = F p ~ ( z ) y=F_{\tilde p}(z) y = F p ~ ( z ) and y = p ~ z / ν y=\tilde p z/\nu y = p ~ z / ν in the case where p ~ / F p ~ ′ ( r ~ − ) < ν < p ~ / F p ~ ′ ( r ~ + ) \tilde p/F'_{\tilde p}(\sqrt{\tilde r_{-}})<\nu<\tilde p/F'_{\tilde p}(\sqrt{\tilde r_{+}}) p ~ / F p ~ ′ ( r ~ − ) < ν < p ~ / F p ~ ′ ( r ~ + ) .
y = F p ~ ( z 0 ) + F p ~ ′ ( z 0 ) ( z − z 0 ) = p ~ + 2 p ~ 2 p ~ 2 − z 0 2 + 4 p ~ 2 z 0 2 ( p ~ 2 − z 0 2 ) 2 ( z − z 0 ) . y=F_{\tilde p}(z_0)+F'_{\tilde p}(z_0)(z-z_0)=\tilde p+\frac{2\tilde p^2}{\tilde p^2-z_0^2}+\frac{4\tilde p^2z_0^2}{(\tilde p^2-z_0^2)^2}(z-z_0). y = F p ~ ( z 0 ) + F p ~ ′ ( z 0 ) ( z − z 0 ) = p ~ + p ~ 2 − z 0 2 2 p ~ 2 + ( p ~ 2 − z 0 2 ) 2 4 p ~ 2 z 0 2 ( z − z 0 ) .
This straight line passes through the origin if
0 = p ~ + 2 p ~ 2 p ~ 2 − z 0 2 − 4 p ~ 2 z 0 2 ( p ~ 2 − z 0 2 ) 2 = p ~ ( p ~ 2 − z 0 2 ) 2 + 2 p ~ 2 ( p ~ 2 − z 0 2 ) − 4 p ~ 2 z 0 2 ( p ~ 2 − z 0 2 ) 2 . 0=\tilde p+\frac{2\tilde p^2}{\tilde p^2-z_0^2}-\frac{4\tilde p^2z_0^2}{(\tilde p^2-z_0^2)^2}
=\frac{\tilde p(\tilde p^2-z_0^2)^2+2\tilde p^2(\tilde p^2-z_0^2)-4\tilde p^2z_0^2}{(\tilde p^2-z_0^2)^2}. 0 = p ~ + p ~ 2 − z 0 2 2 p ~ 2 − ( p ~ 2 − z 0 2 ) 2 4 p ~ 2 z 0 2 = ( p ~ 2 − z 0 2 ) 2 p ~ ( p ~ 2 − z 0 2 ) 2 + 2 p ~ 2 ( p ~ 2 − z 0 2 ) − 4 p ~ 2 z 0 2 .
Thus, we see that the numerator is zero whenever
z 0 4 − 2 z 0 2 ( p ~ 2 + 3 p ~ ) + ( p ~ 4 + 2 p ~ 3 ) = 0. z_0^4-2z_0^2(\tilde p^2+3\tilde p)+(\tilde p^4+2\tilde p^3)=0. z 0 4 − 2 z 0 2 ( p ~ 2 + 3 p ~ ) + ( p ~ 4 + 2 p ~ 3 ) = 0.
Solving this quadratic equation in z 0 2 z_0^2 z 0 2 leads to
z 0 2 = p ~ 2 + 3 p ~ ± ( p ~ 4 + 6 p ~ 3 + 9 p ~ 2 ) − ( p ~ 4 + 2 p ~ 3 ) = p ~ ( 3 + p ~ ± 9 + 4 p ~ ) = r ~ ± . z_0^2=\tilde p^2+3\tilde p\pm\sqrt{(\tilde p^4+6\tilde p^3+9\tilde p^2)-(\tilde p^4+2\tilde p^3)}
=\tilde p\left(3+\tilde p\pm\sqrt{9+4\tilde p}\right)=\tilde r_{\pm}. z 0 2 = p ~ 2 + 3 p ~ ± ( p ~ 4 + 6 p ~ 3 + 9 p ~ 2 ) − ( p ~ 4 + 2 p ~ 3 ) = p ~ ( 3 + p ~ ± 9 + 4 p ~ ) = r ~ ± .
The critical values of ν \nu ν are therefore given by p ~ / ν = F p ~ ′ ( z 0 ) \tilde{p}/\nu= F_{\tilde{p}}'(z_0) p ~ / ν = F p ~ ′ ( z 0 ) with z 0 = r ~ ± z_0 = \sqrt{\tilde{r}_{\pm}} z 0 = r ~ ± , and the result follows.
We now turn our attention to the function r ~ ↦ H ^ ( r ~ , p ~ , ν ) \tilde{r} \mapsto\widehat{H}(\tilde{r},\tilde{p},\nu) r ~ ↦ H ( r ~ , p ~ , ν ) .
Lemma 3.5 (Critical points of H ^ \widehat{H} H ). For each fixed p ~ > 0 \tilde{p} > 0 p ~ > 0 , r ~ ↦ H ^ ( r ~ , p ~ , ν ) \tilde{r} \mapsto\widehat{H}(\tilde{r},\tilde{p},\nu) r ~ ↦ H ( r ~ , p ~ , ν ) has at most one critical point r ~ ∗ ∈ [ r ~ min , r ~ + ] \tilde{r}^{*} \in[\tilde{r}_{\min},\tilde{r}_{+}] r ~ ∗ ∈ [ r ~ m i n , r ~ + ] such that ∂ H ^ ∂ r ~ ( r ~ ∗ , p ~ , ν ) = 0 \frac{\partial\widehat{H}}{\partial\tilde{r}}(\tilde{r}^{*},\tilde{p},\nu) = 0 ∂ r ~ ∂ H ( r ~ ∗ , p ~ , ν ) = 0 .
Proof. If r ~ ∗ \tilde{r}^{*} r ~ ∗ is a critical point of H ^ ( r ~ , p ~ , ν ) = h ( x ~ + ( r ~ ) , r ~ , p ~ , ν ) \widehat{H}(\tilde{r},\tilde{p},\nu) = h(\tilde{x}_{+}(\tilde{r}),\tilde{r},\tilde{p},\nu) H ( r ~ , p ~ , ν ) = h ( x ~ + ( r ~ ) , r ~ , p ~ , ν ) , then
∂ H ^ ∂ r ~ ( r ~ ∗ , p ~ , ν ) = ∂ h ∂ r ~ ( x ~ + ( r ~ ∗ ) , r ~ ∗ , p ~ , ν ) + ∂ h ∂ x ~ ( x ~ + ( r ~ ∗ ) , r ~ ∗ , p ~ , ν ) ⏟ = 0 d x ~ + d r ~ ( r ~ ∗ ) = 0. (24) \frac{\partial\widehat{H}}{\partial\tilde{r}}(\tilde{r}^{*},\tilde{p},\nu) = \frac{\partial h}{\partial\tilde{r}}(\tilde{x}_{+}(\tilde{r}^{*}),\tilde{r}^{*},\tilde{p},\nu) + \underbrace{\frac{\partial h}{\partial\tilde{x}}(\tilde{x}_{+}(\tilde{r}^{*}),\tilde{r}^{*},\tilde{p},\nu)}_{=0} \frac{\mathrm{d}\tilde{x}_{+}}{\mathrm{d}\tilde{r}}(\tilde{r}^{*}) = 0. \tag*{(24)} ∂ r ~ ∂ H ( r ~ ∗ , p ~ , ν ) = ∂ r ~ ∂ h ( x ~ + ( r ~ ∗ ) , r ~ ∗ , p ~ , ν ) + = 0 ∂ x ~ ∂ h ( x ~ + ( r ~ ∗ ) , r ~ ∗ , p ~ , ν ) d r ~ d x ~ + ( r ~ ∗ ) = 0. (24)
Therefore, ( x ~ + , r ~ ∗ ) (\tilde{x}_{+},\tilde{r}^{*}) ( x ~ + , r ~ ∗ ) is a solution of the system
{ ∂ h ∂ r ~ = 1 − 4 ν p ~ x ~ G ( x ~ , r ~ , p ~ ) = 0 , ∂ h ∂ x ~ = − 2 ν − 4 ν p ~ ( p ~ 2 − 2 x ~ 2 − r ~ ) G ( x ~ , r ~ , p ~ ) = 0. (25) \begin{cases}
\frac{\partial h}{\partial\tilde{r}} = 1 - \frac{4\nu\tilde{p}\tilde{x}}{G(\tilde{x},\tilde{r},\tilde{p})} = 0,\\
\frac{\partial h}{\partial\tilde{x}} = -2\nu- \frac{4\nu\tilde{p}(\tilde{p}^{2} - 2\tilde{x}^{2} - \tilde{r})}{G(\tilde{x},\tilde{r},\tilde{p})} = 0. \tag*{(25)}
\end{cases} { ∂ r ~ ∂ h = 1 − G ( x ~ , r ~ , p ~ ) 4 ν p ~ x ~ = 0 , ∂ x ~ ∂ h = − 2 ν − G ( x ~ , r ~ , p ~ ) 4 ν p ~ ( p ~ 2 − 2 x ~ 2 − r ~ ) = 0. (25)
where G ( x ~ , r ~ , p ~ ) = ( p ~ 2 − r ~ ) 2 + 4 x ~ 2 ( x ~ 2 − r ~ ) > 0 G(\tilde{x},\tilde{r},\tilde{p}) = (\tilde{p}^{2} - \tilde{r})^{2} + 4\tilde{x}^{2}(\tilde{x}^{2} - \tilde{r}) > 0 G ( x ~ , r ~ , p ~ ) = ( p ~ 2 − r ~ ) 2 + 4 x ~ 2 ( x ~ 2 − r ~ ) > 0 for all x ~ > r ~ \tilde{x} > \sqrt{\tilde{r}} x ~ > r ~ . At the solution ( x ~ + , r ~ ∗ ) (\tilde{x}_{+},\tilde{r}^{*}) ( x ~ + , r ~ ∗ ) , we can isolate G ( x ~ + , r ~ ∗ , p ~ ) G(\tilde{x}_{+},\tilde{r}^{*},\tilde{p}) G ( x ~ + , r ~ ∗ , p ~ ) from each equation to get
G ( x ~ + , r ~ ∗ , p ~ ) = 4 ν p ~ x ~ + = − 2 p ~ ( p ~ 2 − 2 x ~ + 2 − r ~ ∗ ) , (26) G(\tilde{x}_{+},\tilde{r}^{*},\tilde{p}) = 4\nu\tilde{p}\tilde{x}_{+} = -2\tilde{p}(\tilde{p}^{2} - 2\tilde{x}_{+}^{2} - \tilde{r}^{*}), \tag*{(26)} G ( x ~ + , r ~ ∗ , p ~ ) = 4 ν p ~ x ~ + = − 2 p ~ ( p ~ 2 − 2 x ~ + 2 − r ~ ∗ ) , (26)
from which we obtain
r ~ ∗ = p ~ 2 + 2 ν x ~ + − 2 x ~ + 2 . (27) \tilde{r}^{*} = \tilde{p}^{2} + 2\nu\tilde{x}_{+} - 2\tilde{x}_{+}^{2}. \tag*{(27)} r ~ ∗ = p ~ 2 + 2 ν x ~ + − 2 x ~ + 2 . (27)
Hence, we define r ~ = φ ( x ~ ) = p ~ 2 + 2 ν x ~ − 2 x ~ 2 \tilde{r} = \varphi(\tilde{x}) = \tilde{p}^{2} + 2\nu\tilde{x} - 2\tilde{x}^{2} r ~ = φ ( x ~ ) = p ~ 2 + 2 ν x ~ − 2 x ~ 2 . Next, we use the explicit formula of x ~ + \tilde{x}_{+} x ~ + in (14) to define the function r ~ = ψ ( x ~ ) \tilde{r} = \psi(\tilde{x}) r ~ = ψ ( x ~ ) , such that
x ~ + = ( ψ ( x ~ + ) + p ~ + A ψ ( x ~ + ) − B 2 ) 1 / 2 = : X ( ψ ( x ~ + ) ) . \tilde{x}_{+} = \left(\frac{\psi(\tilde{x}_{+}) + \tilde{p} + \sqrt{A\psi(\tilde{x}_{+}) - B}}{2}\right)^{1/2} =: X(\psi(\tilde{x}_{+})). x ~ + = ( 2 ψ ( x ~ + ) + p ~ + A ψ ( x ~ + ) − B ) 1/2 =: X ( ψ ( x ~ + )) .
In other words, ψ \psi ψ and X X X are inverse functions. We claim that ψ \psi ψ is an increasing convex function. Indeed, since X ( r ~ ) > 0 X(\tilde{r}) > 0 X ( r ~ ) > 0 and
2 ( X ( r ~ ) ) 2 = r ~ + p ~ + A r ~ − B ⟹ 4 X ( r ~ ) X ′ ( r ~ ) = 1 + A 2 A r ~ − B > 0 ⟹ X ′ ( r ~ ) > 0 , 2(X(\tilde{r}))^{2} = \tilde{r} + \tilde{p} + \sqrt{A\tilde{r} - B} \Longrightarrow4X(\tilde{r})X'(\tilde{r}) = 1 + \frac{A}{2\sqrt{A\tilde{r} - B}} > 0 \Longrightarrow X'(\tilde{r}) > 0, 2 ( X ( r ~ ) ) 2 = r ~ + p ~ + A r ~ − B ⟹ 4 X ( r ~ ) X ′ ( r ~ ) = 1 + 2 A r ~ − B A > 0 ⟹ X ′ ( r ~ ) > 0 ,
Differentiating once more gives
4 ( X ′ ( r ~ ) ) 2 + X ( r ~ ) X ′ ′ ( r ~ ) = − A 2 4 ( A r ~ − B ) 3 / 2 ⟹ X ′ ′ ( r ~ ) = 1 X ( r ~ ) ( − A 2 4 ( A r ~ − B ) 3 / 2 − 4 ( X ′ ( r ~ ) ) 2 ) < 0. 4(X'(\tilde{r}))^{2} + X(\tilde{r})X''(\tilde{r}) = \frac{-A^{2}}{4(A\tilde{r} - B)^{3/2}} \Longrightarrow X''(\tilde{r}) = \frac{1}{X(\tilde{r})}\left(\frac{-A^{2}}{4(A\tilde{r} - B)^{3/2}} - 4(X'(\tilde{r}))^{2}\right) < 0. 4 ( X ′ ( r ~ ) ) 2 + X ( r ~ ) X ′′ ( r ~ ) = 4 ( A r ~ − B ) 3/2 − A 2 ⟹ X ′′ ( r ~ ) = X ( r ~ ) 1 ( 4 ( A r ~ − B ) 3/2 − A 2 − 4 ( X ′ ( r ~ ) ) 2 ) < 0.
Since the inverse of an increasing concave function is increasing and convex, we conclude that ψ ( x ~ ) \psi(\tilde{x}) ψ ( x ~ ) is increasing and convex. We now show that there can be at most one intersection between r ~ = φ ( x ~ ) \tilde{r} = \varphi(\tilde{x}) r ~ = φ ( x ~ ) and r ~ = ψ ( x ~ ) \tilde{r} = \psi(\tilde{x}) r ~ = ψ ( x ~ ) within the interval x ~ ∈ [ x ~ min , x ~ + ] : = [ X ( r ~ min ) , X ( r ~ ∗ ) ] \tilde{x} \in[\tilde{x}_{\min},\tilde{x}_{+}] := [X(\tilde{r}_{\min}), X(\tilde{r}^{*})] x ~ ∈ [ x ~ m i n , x ~ + ] := [ X ( r ~ m i n ) , X ( r ~ ∗ )] . First, we either have r ~ min = 0 \tilde{r}_{\min} = 0 r ~ m i n = 0 or
r ~ min = B A = p ~ 4 + 2 p ~ 3 − p ~ 2 2 p ~ 2 + 4 p ~ < p ~ 4 + 2 p ~ 3 − p ~ 2 2 p ~ 2 + 4 p ~ p ~ 2 2 . \tilde{r}_{\min} = \frac{B}{A} = \frac{\tilde{p}^{4} + 2\tilde{p}^{3} - \tilde{p}^{2}}{2\tilde{p}^{2} + 4\tilde{p}} < \frac{\tilde{p}^{4} + 2\tilde{p}^{3} - \tilde{p}^{2}}{2\tilde{p}^{2} + 4\tilde{p}} \frac{\tilde{p}^{2}}{2}. r ~ m i n = A B = 2 p ~ 2 + 4 p ~ p ~ 4 + 2 p ~ 3 − p ~ 2 < 2 p ~ 2 + 4 p ~ p ~ 4 + 2 p ~ 3 − p ~ 2 2 p ~ 2 .
So either way, r ~ min < p ~ 2 = φ ( 0 ) \tilde{r}_{\min}<\tilde{p}^{2}=\varphi(0) r ~ m i n < p ~ 2 = φ ( 0 ) . If r ~ min > φ ( x ~ min ) \tilde{r}_{\min}>\varphi(\tilde{x}_{\min}) r ~ m i n > φ ( x ~ m i n ) , then
p ~ 2 > p ~ 2 + 2 ν x ~ min − 2 x ~ min 2 ⟹ x ~ min > ν . \tilde{p}^{2}>\tilde{p}^{2}+2\nu\tilde{x}_{\min}-2\tilde{x}_{\min}^{2}\quad\Longrightarrow\quad\tilde{x}_{\min}>\nu. p ~ 2 > p ~ 2 + 2 ν x ~ m i n − 2 x ~ m i n 2 ⟹ x ~ m i n > ν .
Since φ ′ ( x ~ ) = 2 ν − 4 x ~ < 0 \varphi'(\tilde{x})=2\nu-4\tilde{x}<0 φ ′ ( x ~ ) = 2 ν − 4 x ~ < 0 for x ~ > ν \tilde{x}>\nu x ~ > ν , we see that ψ − φ \psi-\varphi ψ − φ is strictly increasing on [ x ~ min , x ~ + ] [\tilde{x}_{\min},\tilde{x}_{+}] [ x ~ m i n , x ~ + ] , so any x ~ \tilde{x} x ~ in this interval must satisfy
ψ ( x ~ ) − φ ( x ~ ) ≥ ψ ( x ~ min ) − φ ( x ~ min ) = r ~ min − φ ( x ~ min ) > 0. \psi(\tilde{x})-\varphi(\tilde{x})\geq\psi(\tilde{x}_{\min})-\varphi(\tilde{x}_{\min})=\tilde{r}_{\min}-\varphi(\tilde{x}_{\min})>0. ψ ( x ~ ) − φ ( x ~ ) ≥ ψ ( x ~ m i n ) − φ ( x ~ m i n ) = r ~ m i n − φ ( x ~ m i n ) > 0.
In other words, no intersection can exist between r ~ = ψ ( x ~ ) \tilde{r}=\psi(\tilde{x}) r ~ = ψ ( x ~ ) and r ~ = φ ( x ~ ) \tilde{r}=\varphi(\tilde{x}) r ~ = φ ( x ~ ) . Now suppose r ~ min ≤ φ ( x ~ min ) \tilde{r}_{\min}\leq\varphi(\tilde{x}_{\min}) r ~ m i n ≤ φ ( x ~ m i n ) . Since ψ ( x ~ min ) − φ ( x ~ min ) ≤ 0 \psi(\tilde{x}_{\min})-\varphi(\tilde{x}_{\min})\leq0 ψ ( x ~ m i n ) − φ ( x ~ m i n ) ≤ 0 and ψ − φ \psi-\varphi ψ − φ is convex on [ x ~ min , x ~ + ] [\tilde{x}_{\min},\tilde{x}_{+}] [ x ~ m i n , x ~ + ] , it can have at most one root x ~ + ∈ [ x ~ min , x ~ + ] \tilde{x}_{+}\in[\tilde{x}_{\min},\tilde{x}_{+}] x ~ + ∈ [ x ~ m i n , x ~ + ] .
In summary, since there is at most one intersection between r ~ = φ ( x ~ ) \tilde{r}=\varphi(\tilde{x}) r ~ = φ ( x ~ ) and r ~ = ψ ( x ~ ) \tilde{r}=\psi(\tilde{x}) r ~ = ψ ( x ~ ) , there is at most one solution to the optimality system (25) , so H ^ \hat{H} H ^ can have at most one critical point.
Given the different ways in which r ~ ↦ M ( r ~ , p ~ , ν ) \tilde{r}\mapsto M(\tilde{r},\tilde{p},\nu) r ~ ↦ M ( r ~ , p ~ , ν ) can achieve its minimum (see Figure 3.2 ), it is difficult to obtain a closed form expression for the optimal r ~ \tilde{r} r ~ and the resulting convergence bound analytically. Nonetheless, the optimal bound is easy to calculate numerically for a given p ~ = δ p \tilde{p}=\delta p p ~ = δ p :
If p ~ ≥ 5 \tilde{p}\geq\sqrt{5} p ~ ≥ 5 :
Compute r ~ ± \tilde{r}_{\pm} r ~ ± , defined in (16) . If H 0 H_0 H 0 has a local minimum r ~ 0 \tilde{r}_0 r ~ 0 by Lemma 3.4, then this is also the global minimum of r ~ ↦ M ( r ~ , p ~ , ν ) \tilde{r}\mapsto M(\tilde{r},\tilde{p},\nu) r ~ ↦ M ( r ~ , p ~ , ν ) .
If p ~ < 5 \tilde{p}<\sqrt{5} p ~ < 5 :
Compute the local minimum r ^ \hat{r} r ^ of H ^ \hat{H} H ^ by solving (14) and (27) simultaneously. If r ^ \hat{r} r ^ exists and satisfies H 0 ( r ^ , p ~ , ν ~ ) ≤ H ^ ( r ^ , p ~ , ν ~ ) H_0(\hat{r},\tilde{p},\tilde{\nu})\leq\hat{H}(\hat{r},\tilde{p},\tilde{\nu}) H 0 ( r ^ , p ~ , ν ~ ) ≤ H ^ ( r ^ , p ~ , ν ~ ) , then r ^ \hat{r} r ^ is the global minimum of r ~ ↦ M ( r ~ , p ~ , ν ~ ) \tilde{r}\mapsto M(\tilde{r},\tilde{p},\tilde{\nu}) r ~ ↦ M ( r ~ , p ~ , ν ~ ) .
Otherwise, compute the equioscillation point r ~ × \tilde{r}^{\times} r ~ × by setting H 0 H_0 H 0 equal to H ^ \hat{H} H ^ (e.g. using r ~ − \tilde{r}_{-} r ~ − as the starting point for Newton). If ∂ H 0 ∂ r ~ ∂ H ^ ∂ r ~ < 0 \frac{\partial H_0}{\partial\tilde{r}}\frac{\partial\hat{H}}{\partial\tilde{r}}<0 ∂ r ~ ∂ H 0 ∂ r ~ ∂ H ^ < 0 at r ~ = r ~ × \tilde{r}=\tilde{r}^{\times} r ~ = r ~ × , then r ~ × \tilde{r}^{\times} r ~ × is the global minimum of M M M .
If neither r ^ \hat{r} r ^ nor r ~ × \tilde{r}^{\times} r ~ × works, then H 0 H_0 H 0 must have a local minimum r ~ 0 \tilde{r}_0 r ~ 0 that coincides with that of M M M . Calculate r ~ 0 \tilde{r}_0 r ~ 0 by finding the root of ∂ H 0 ∂ r ~ = 0 \frac{\partial H_0}{\partial\tilde{r}}=0 ∂ r ~ ∂ H 0 = 0 that is outside the interval [ r ~ − , r ~ + ] [\tilde{r}_{-},\tilde{r}_{+}] [ r ~ − , r ~ + ] , see (22) for an expression of ∂ H 0 ∂ r ~ \frac{\partial H_0}{\partial\tilde{r}} ∂ r ~ ∂ H 0 .
We use this procedure to compute bounds for the four test cases shown in Figure 3.4 . The convergence curves correspond to OSWR applied to the 1D homogeneous heat equation posed on the finite domain y ∈ Ω = ( − 4 , 4 ) y\in\Omega=(-4,4) y ∈ Ω = ( − 4 , 4 ) and integrated on the time interval t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . We impose homogeneous Dirichlet conditions at y = ± 4 y=\pm4 y = ± 4 and zero initial conditions. The problem is discretized using backward Euler in time and centered finite differences in space, with Δ y = 1 / 800 \Delta y=1/800 Δ y = 1/800 and Δ t = 1 / 2400 \Delta t=1/2400 Δ t = 1/2400 . The two subdomains are of equal size, with the overlap indicated on the plot title; the Robin parameters used in the transmission conditions are also shown in each case. Random initial guesses are used for the interface at iteration 0. We see that in all cases, our theoretical bounds track the convergence curves closely and are able to capture the superlinear behaviour reliably.
Figure 3.4. Comparison between convergence curves for OSWR and the superlinear bound for different values of p p p and δ \delta δ . We have T = 1 T=1 T = 1 in all cases.
Optimizing the Robin parameter
Interestingly, the analysis becomes more tractable when we seek an optimal Robin parameter p ~ \tilde{p} p ~ , since optimality yields an extra equation that can be used to obtain asymptotic estimates under various scalings of ν \nu ν . We start by showing that such an optimal p ~ \tilde{p} p ~ exists.
Theorem 4.1. Let ν > 0 \nu>0 ν > 0 be given. For any p ~ > 0 \tilde{p}>0 p ~ > 0 , let r ~ ∗ = r ~ ∗ ( p ~ , ν ) \tilde{r}_*=\tilde{r}_*(\tilde{p},\nu) r ~ ∗ = r ~ ∗ ( p ~ , ν ) be a global minimum of the function r ~ ↦ M ( r ~ , p ~ , ν ) \tilde{r}\mapsto M(\tilde{r},\tilde{p},\nu) r ~ ↦ M ( r ~ , p ~ , ν ) . Then there exists a value P ~ \tilde{P} P ~ , which depends on ν \nu ν , such that for any p ~ > P ~ \tilde{p}>\tilde{P} p ~ > P ~ , we have M ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) > M ( r ~ ∗ ( P ~ , ν ) , P ~ , ν ) M(\tilde{r}_*(\tilde{p},\nu),\tilde{p},\nu)>M(\tilde{r}_*(\tilde{P},\nu),\tilde{P},\nu) M ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) > M ( r ~ ∗ ( P ~ , ν ) , P ~ , ν ) . As a result, the function p ~ ↦ M ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) \tilde{p}\mapsto M(\tilde{r}_*(\tilde{p},\nu),\tilde{p},\nu) p ~ ↦ M ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) attains its minimum on the interval [ 0 , P ~ ] [0,\tilde{P}] [ 0 , P ~ ] .
Proof. Without loss of generality, assume that P ~ ≥ 5 \tilde{P}\geq\sqrt{5} P ~ ≥ 5 . Then for any p ~ ≥ P ~ \tilde{p}\geq\tilde{P} p ~ ≥ P ~ , the function r ~ ↦ M ( r ~ , p ~ , ν ) \tilde{r}\mapsto M(\tilde{r},\tilde{p},\nu) r ~ ↦ M ( r ~ , p ~ , ν ) is differentiable, so r ~ ∗ \tilde{r}_* r ~ ∗ must be a root of ∂ M ∂ r ~ ( r ~ , p ~ , ν ) \frac{\partial M}{\partial\tilde{r}}(\tilde{r},\tilde{p},\nu) ∂ r ~ ∂ M ( r ~ , p ~ , ν ) . We will show for p ~ \tilde{p} p ~ large enough that r ~ ∗ \tilde{r}_* r ~ ∗ must be a local minimum of r ~ ↦ H 0 ( r ~ , p ~ , ν ) \tilde{r}\mapsto H_0(\tilde{r},\tilde{p},\nu) r ~ ↦ H 0 ( r ~ , p ~ , ν ) . Indeed, by Lemma 3.4, we know that if ν < p ~ / F p ~ ′ ( r ~ − ) \nu<\tilde{p}/F'_{\tilde{p}}(\sqrt{\tilde{r}_{-}}) ν < p ~ / F p ~ ′ ( r ~ − ) , then H 0 H_0 H 0 has a local minimum at r ~ ∗ ∈ ( 0 , r ~ − ) \tilde{r}_*\in(0,\tilde{r}_{-}) r ~ ∗ ∈ ( 0 , r ~ − ) ; in fact, this is also the global minimum of r ~ ↦ M ( r ~ , p ~ , ν ) \tilde{r}\mapsto M(\tilde{r},\tilde{p},\nu) r ~ ↦ M ( r ~ , p ~ , ν ) , since r ~ ↦ H ^ ( r ~ , p ~ , ν ) \tilde{r}\mapsto\hat{H}(\tilde{r},\tilde{p},\nu) r ~ ↦ H ^ ( r ~ , p ~ , ν ) must be monotonically increasing on [ r ~ − , r ~ + ] [\tilde{r}_{-},\tilde{r}_{+}] [ r ~ − , r ~ + ] (it is increasing at both end points and cannot have more than one critical point inside this interval). Thus, if we can show that p ~ / F p ~ ′ ( r ~ − ) → ∞ \tilde{p}/F'_{\tilde{p}}(\sqrt{\tilde{r}_{-}})\to\infty p ~ / F p ~ ′ ( r ~ − ) → ∞ as p ~ → ∞ \tilde{p}\to\infty p ~ → ∞ , then we can find P ~ = P ~ ( ν ) \tilde{P}=\tilde{P}(\nu) P ~ = P ~ ( ν ) such that ν < p ~ / F p ~ ′ ( r ~ − ) \nu<\tilde{p}/F'_{\tilde{p}}(\sqrt{\tilde{r}_{-}}) ν < p ~ / F p ~ ′ ( r ~ − ) whenever p ~ > P ~ \tilde{p}>\tilde{P} p ~ > P ~ .
To this end, let us show that p ~ ↦ F p ~ ′ ( r ~ − ) \tilde{p}\mapsto F'_{\tilde{p}}(\sqrt{\tilde{r}_{-}}) p ~ ↦ F p ~ ′ ( r ~ − ) is a positive, strictly decreasing function of p ~ \tilde{p} p ~ . Recall from (23) that
r ~ − = p ~ ( 3 + p ~ − 9 + 4 p ~ ) = 2 p ~ 2 + p ~ 3 3 + p ~ + 9 + 4 p ~ . \tilde{r}_{-}=\tilde{p}\left(3+\tilde{p}-\sqrt{9+4\tilde{p}}\right)=\frac{2\tilde{p}^{2}+\tilde{p}^{3}}{3+\tilde{p}+\sqrt{9+4\tilde{p}}}. r ~ − = p ~ ( 3 + p ~ − 9 + 4 p ~ ) = 3 + p ~ + 9 + 4 p ~ 2 p ~ 2 + p ~ 3 .
Substituting this into the definition of F p ~ ′ F'_{\tilde{p}} F p ~ ′ and simplifying leads to
F p ~ ′ ( r ~ − ) = 4 p ~ 2 r ~ − ( p ~ 2 − r ~ − ) 2 = ( ( 3 + 9 + 4 p ~ 2 ) ( 2 + p ~ 2 ) 4 p ~ 2 ( 1 + 9 + 4 p ~ 2 ) ) 1 / 2 . F'_{\tilde{p}}(\sqrt{\tilde{r}_{-}})=\frac{4\tilde{p}^{2}\sqrt{\tilde{r}_{-}}}{(\tilde{p}^{2}-\tilde{r}_{-})^{2}}=\left(\frac{(3+\sqrt{9+4\tilde{p}^{2}})(2+\tilde{p}^{2})}{4\tilde{p}^{2}(1+\sqrt{9+4\tilde{p}^{2}})}\right)^{1/2}. F p ~ ′ ( r ~ − ) = ( p ~ 2 − r ~ − ) 2 4 p ~ 2 r ~ − = ( 4 p ~ 2 ( 1 + 9 + 4 p ~ 2 ) ( 3 + 9 + 4 p ~ 2 ) ( 2 + p ~ 2 ) ) 1/2 .
The latter function is positive and strictly decreasing with p ~ \tilde{p} p ~ ; in particular, we have 1 < F p ~ ′ ( r ~ − ) ≤ 8 1<F'_{\tilde{p}}(\sqrt{\tilde{r}_{-}})\leq\sqrt{8} 1 < F p ~ ′ ( r ~ − ) ≤ 8 whenever p ~ ≥ 4 \tilde{p}\geq4 p ~ ≥ 4 . Thus, by choosing P ~ = max ( 4 , 8 ν ) \tilde{P}=\max(4,\sqrt{8}\nu) P ~ = max ( 4 , 8 ν ) , we guarantee that ν < p ~ / F p ~ ′ ( r ~ − ) \nu<\tilde{p}/F'_{\tilde{p}}(\sqrt{\tilde{r}_{-}}) ν < p ~ / F p ~ ′ ( r ~ − ) for all p ~ > P ~ \tilde{p}>\tilde{P} p ~ > P ~ , meaning that r ~ ∗ ( p ~ , ν ) \tilde{r}_{*}(\tilde{p},\nu) r ~ ∗ ( p ~ , ν ) will always be a local minimum of H 0 H_{0} H 0 for these values of p ~ \tilde{p} p ~ .
It remains to show that H 0 ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) > H 0 ( r ~ ∗ ( P ~ , ν ) , P ~ , ν ) H_{0}(\tilde{r}_{*}(\tilde{p},\nu),\tilde{p},\nu)>H_{0}(\tilde{r}_{*}(\tilde{P},\nu),\tilde{P},\nu) H 0 ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) > H 0 ( r ~ ∗ ( P ~ , ν ) , P ~ , ν ) for any p ~ > P ~ \tilde{p}>\tilde{P} p ~ > P ~ . To do so, we write K ( p ~ , ν ) : = H 0 ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) K(\tilde{p},\nu):=H_{0}(\tilde{r}_{*}(\tilde{p},\nu),\tilde{p},\nu) K ( p ~ , ν ) := H 0 ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) and show that ∂ K ∂ p ~ > 0 \frac{\partial K}{\partial\tilde{p}}>0 ∂ p ~ ∂ K > 0 for all p ~ > P ~ \tilde{p}>\tilde{P} p ~ > P ~ . Indeed, we have
∂ K ∂ p ~ ( p ~ , ν ) = ∂ H 0 ∂ r ~ ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) ⏟ = 0 ∂ r ~ ∗ ∂ p ~ ( p ~ , ν ) + ∂ H 0 ∂ p ~ ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) = 4 ν r ~ ∗ p ~ 2 − r ~ ∗ > 0 , \frac{\partial K}{\partial\tilde{p}}(\tilde{p},\nu)=\underbrace{\frac{\partial H_{0}}{\partial\tilde{r}}(\tilde{r}_{*}(\tilde{p},\nu),\tilde{p},\nu)}_{=0}\frac{\partial\tilde{r}_{*}}{\partial\tilde{p}}(\tilde{p},\nu)+\frac{\partial H_{0}}{\partial\tilde{p}}(\tilde{r}_{*}(\tilde{p},\nu),\tilde{p},\nu)=\frac{4\nu\sqrt{\tilde{r}_{*}}}{\tilde{p}^{2}-\tilde{r}_{*}}>0, ∂ p ~ ∂ K ( p ~ , ν ) = = 0 ∂ r ~ ∂ H 0 ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) ∂ p ~ ∂ r ~ ∗ ( p ~ , ν ) + ∂ p ~ ∂ H 0 ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) = p ~ 2 − r ~ ∗ 4 ν r ~ ∗ > 0 ,
since r ~ ∗ \tilde{r}_{*} r ~ ∗ is critical point of r ~ ↦ H 0 ( r ~ , p ~ , ν ) \tilde{r}\mapsto H_{0}(\tilde{r},\tilde{p},\nu) r ~ ↦ H 0 ( r ~ , p ~ , ν ) and r ~ ∗ < r ~ − < p ~ 2 \tilde{r}_{*}<\tilde{r}_{-}<\tilde{p}^{2} r ~ ∗ < r ~ − < p ~ 2 . The mean-value theorem then implies that K ( p ~ , ν ) > K ( P ~ , ν ) K(\tilde{p},\nu)>K(\tilde{P},\nu) K ( p ~ , ν ) > K ( P ~ , ν ) for any p ~ > P ~ \tilde{p}>\tilde{P} p ~ > P ~ , as claimed.
Finally, for a given ν > 0 \nu>0 ν > 0 , the function p ~ ↦ M ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) \tilde{p}\mapsto M(\tilde{r}_{*}(\tilde{p},\nu),\tilde{p},\nu) p ~ ↦ M ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) is continuous everywhere. Since M ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) = K ( p ~ , ν ) M(\tilde{r}_{*}(\tilde{p},\nu),\tilde{p},\nu)=K(\tilde{p},\nu) M ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) = K ( p ~ , ν ) for all p ~ ≥ P ~ \tilde{p}\geq\tilde{P} p ~ ≥ P ~ , the above result implies that
inf p ~ ≥ 0 M ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) = inf p ~ ∈ [ 0 , P ~ ] M ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) = min p ~ ∈ [ 0 , P ~ ] M ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) . \inf_{\tilde{p}\geq0}M(\tilde{r}_{*}(\tilde{p},\nu),\tilde{p},\nu)=\inf_{\tilde{p}\in[0,\tilde{P}]}M(\tilde{r}_{*}(\tilde{p},\nu),\tilde{p},\nu)=\min_{\tilde{p}\in[0,\tilde{P}]}M(\tilde{r}_{*}(\tilde{p},\nu),\tilde{p},\nu). p ~ ≥ 0 inf M ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) = p ~ ∈ [ 0 , P ~ ] inf M ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) = p ~ ∈ [ 0 , P ~ ] min M ( r ~ ∗ ( p ~ , ν ) , p ~ , ν ) .
Thus, the minimum is attained at some optimal parameter p ~ ∗ ∈ [ 0 , P ~ ] \tilde{p}^{*}\in[0,\tilde{P}] p ~ ∗ ∈ [ 0 , P ~ ] , as required. □
The above theorem implies that ( r ~ , p ~ ) ↦ M ( r ~ , p ~ , ν ) (\tilde{r},\tilde{p})\mapsto M(\tilde{r},\tilde{p},\nu) ( r ~ , p ~ ) ↦ M ( r ~ , p ~ , ν ) has a global minimum. Since M M M is a smooth function everywhere except along the equioscillation curve H 0 ( r ~ , p ~ , ν ) = H ^ ( r ~ , p ~ , ν ) H_{0}(\tilde{r},\tilde{p},\nu)=\hat{H}(\tilde{r},\tilde{p},\nu) H 0 ( r ~ , p ~ , ν ) = H ^ ( r ~ , p ~ , ν ) for p ~ < 5 \tilde{p}<\sqrt{5} p ~ < 5 , this global minimum must occur at one of the following points:
A critical point of H 0 H_{0} H 0 , i.e., where ∂ H 0 ∂ r ~ = ∂ H 0 ∂ p ~ = 0 \frac{\partial H_{0}}{\partial\tilde{r}}=\frac{\partial H_{0}}{\partial\tilde{p}}=0 ∂ r ~ ∂ H 0 = ∂ p ~ ∂ H 0 = 0 ;
A critical point of H ^ \hat{H} H ^ , i.e., where ∂ H ^ ∂ r ~ = ∂ H ^ ∂ p ~ = 0 \frac{\partial\hat{H}}{\partial\tilde{r}}=\frac{\partial\hat{H}}{\partial\tilde{p}}=0 ∂ r ~ ∂ H ^ = ∂ p ~ ∂ H ^ = 0 ;
A constrainted minimum along the equioscillation curve r ~ = r ~ × ( p ~ ) \tilde{r}=\tilde{r}^{\times}(\tilde{p}) r ~ = r ~ × ( p ~ ) , where H 0 ( r ~ , p ~ , ν ) = H ^ ( r ~ , p ~ , ν ) H_{0}(\tilde{r},\tilde{p},\nu)=\hat{H}(\tilde{r},\tilde{p},\nu) H 0 ( r ~ , p ~ , ν ) = H ^ ( r ~ , p ~ , ν ) . (Note that this curve is independent of ν \nu ν .)
As a result, we will need expressions for the partial derivatives of H 0 H_{0} H 0 and H ^ \hat{H} H ^ , where ∂ H 0 ∂ r ~ \frac{\partial H_{0}}{\partial\tilde{r}} ∂ r ~ ∂ H 0 is in (3.13), ∂ H ^ ∂ r ~ \frac{\partial\hat{H}}{\partial\tilde{r}} ∂ r ~ ∂ H ^ is in (3.15), and the remaining two are
∂ H 0 ∂ p ~ ( r ~ , p ~ , ν ) = ∂ h ∂ p ~ ( r ~ , r ~ , p ~ , ν ) = 4 ν r ~ ( p ~ 2 − r ~ ) 2 ( p ~ 2 − r ~ ) = 4 ν r ~ p ~ 2 − r ~ . (28) \frac{\partial H_{0}}{\partial\tilde{p}}(\tilde{r},\tilde{p},\nu)=\frac{\partial h}{\partial\tilde{p}}(\sqrt{\tilde{r}},\tilde{r},\tilde{p},\nu)=\frac{4\nu\sqrt{\tilde{r}}}{(\tilde{p}^{2}-\tilde{r})^{2}}(\tilde{p}^{2}-\tilde{r})=\frac{4\nu\sqrt{\tilde{r}}}{\tilde{p}^{2}-\tilde{r}}. \tag*{(28)} ∂ p ~ ∂ H 0 ( r ~ , p ~ , ν ) = ∂ p ~ ∂ h ( r ~ , r ~ , p ~ , ν ) = ( p ~ 2 − r ~ ) 2 4 ν r ~ ( p ~ 2 − r ~ ) = p ~ 2 − r ~ 4 ν r ~ . (28)
∂ H ^ ∂ p ~ ( r ~ , p ~ , ν ) = ∂ h ∂ p ~ ( x ~ + ( r ~ , p ~ ) , r ~ , p ~ , ν ) + ∂ h ∂ x ~ ( x ~ + ( r ~ , p ~ ) , r ~ , p ~ , ν ) ⏟ = 0 ∂ x ~ + ∂ p ~ ( x ~ + ( r ~ , p ~ ) , r ~ , p ~ , ν ) = 4 ν x ~ + G ( x ~ + , r ~ , p ~ ) ( p ~ 2 + r ~ − 2 x ~ + 2 ) . (29) \begin{aligned}
\frac{\partial\hat{H}}{\partial\tilde{p}}(\tilde{r},\tilde{p},\nu)
&=\frac{\partial h}{\partial\tilde{p}}(\tilde{x}_{+}(\tilde{r},\tilde{p}),\tilde{r},\tilde{p},\nu)
+\underbrace{\frac{\partial h}{\partial\tilde{x}}(\tilde{x}_{+}(\tilde{r},\tilde{p}),\tilde{r},\tilde{p},\nu)}_{=0}
\frac{\partial\tilde{x}_{+}}{\partial\tilde{p}}(\tilde{x}_{+}(\tilde{r},\tilde{p}),\tilde{r},\tilde{p},\nu) \\
&=\frac{4\nu\tilde{x}_{+}}{G(\tilde{x}_{+},\tilde{r},\tilde{p})}(\tilde{p}^{2}+\tilde{r}-2\tilde{x}_{+}^{2}). \tag*{(29)}
\end{aligned} ∂ p ~ ∂ H ^ ( r ~ , p ~ , ν ) = ∂ p ~ ∂ h ( x ~ + ( r ~ , p ~ ) , r ~ , p ~ , ν ) + = 0 ∂ x ~ ∂ h ( x ~ + ( r ~ , p ~ ) , r ~ , p ~ , ν ) ∂ p ~ ∂ x ~ + ( x ~ + ( r ~ , p ~ ) , r ~ , p ~ , ν ) = G ( x ~ + , r ~ , p ~ ) 4 ν x ~ + ( p ~ 2 + r ~ − 2 x ~ + 2 ) . (29)
Since ∂ H 0 ∂ p ~ ≠ 0 \frac{\partial H_{0}}{\partial\tilde{p}}\neq0 ∂ p ~ ∂ H 0 = 0 unless r ~ = 0 \tilde{r}=0 r ~ = 0 and ∂ H 0 ∂ r ~ → − ∞ \frac{\partial H_{0}}{\partial\tilde{r}}\to-\infty ∂ r ~ ∂ H 0 → − ∞ as r ~ → 0 + \tilde{r}\to0+ r ~ → 0 + , we conclude that H 0 H_{0} H 0 has no critical points at all. We therefore focus on calculating the critical points of H ^ \hat{H} H ^ and the constrained minima along the curve r ~ = r ~ ∗ ( p ~ ) \tilde r=\tilde r^*(\tilde p) r ~ = r ~ ∗ ( p ~ ) . This will be done in the next two subsections; the results will culminate in the theorem below, which is stated in terms of the unscaled quantities δ \delta δ , p p p , T T T and n n n and is the main theorem of this paper.
Theorem 4.2. Consider the alternating OSWR method with Robin transmission conditions B i = ( − 1 ) i + 1 ∂ y + p B_i=(-1)^{i+1}\partial_y+p B i = ( − 1 ) i + 1 ∂ y + p , where p > 0 p>0 p > 0 . Let ν = n δ 2 / T \nu=n\delta^2/T ν = n δ 2 / T , where δ > 0 \delta>0 δ > 0 is the overlap, T > 0 T>0 T > 0 is the length of the time horizon, and n ≥ 1 n\ge1 n ≥ 1 is the number of iterations taken.
For ν ≫ 1 \nu\gg1 ν ≫ 1 (e.g., for T → 0 T\to0 T → 0 , or for fixed T T T and δ \delta δ and letting n → ∞ n\to\infty n → ∞ ), the optimized parameter p ∗ p^* p ∗ and the corresponding n n n -step contraction factor ρ ∗ \rho^* ρ ∗ behave asymptotically as
p ∗ ∼ p ∞ ∗ : = n δ T + 3 2 δ , ρ ∗ ∼ e − n 2 δ 2 / T ( C 1 n δ 2 T ) − n , p^*\sim p_\infty^*:=\frac{n\delta}{T}+\frac{3}{2\delta},\qquad\rho^*\sim e^{-n^2\delta^2/T}\left(\frac{C_1n\delta^2}{T}\right)^{-n}, p ∗ ∼ p ∞ ∗ := T n δ + 2 δ 3 , ρ ∗ ∼ e − n 2 δ 2 / T ( T C 1 n δ 2 ) − n , where C 1 = 4 e ≈ 10.87 C_1=4e\approx10.87 C 1 = 4 e ≈ 10.87 .
For ν ≪ 1 \nu\ll1 ν ≪ 1 (e.g., for fixed n n n and letting either δ → 0 \delta\to0 δ → 0 or T → ∞ T\to\infty T → ∞ ), the optimized parameter p ∗ p^* p ∗ and the corresponding n n n -step contraction factor ρ ∗ \rho^* ρ ∗ behave asymptotically as
p ∗ ∼ p 0 ∗ : = ( 4 9 T 2 δ ) 1 / 5 n 2 / 5 + C 2 ( δ 3 T 4 ) 1 / 5 n 4 / 5 , ρ ∗ ∼ e − C 3 n 6 / 5 ( δ 2 / T ) 1 / 5 , p^*\sim p_0^*:=\left(\frac{4}{9T^2\delta}\right)^{1/5}n^{2/5}+C_2\left(\frac{\delta^3}{T^4}\right)^{1/5}n^{4/5},\qquad\rho^*\sim e^{-C_3n^{6/5}(\delta^2/T)^{1/5}}, p ∗ ∼ p 0 ∗ := ( 9 T 2 δ 4 ) 1/5 n 2/5 + C 2 ( T 4 δ 3 ) 1/5 n 4/5 , ρ ∗ ∼ e − C 3 n 6/5 ( δ 2 / T ) 1/5 , where C 2 ≈ 0.6346 C_2\approx0.6346 C 2 ≈ 0.6346 and C 3 = 10 3 ( 2 3 ) 1 / 5 ≈ 3.07 C_3=\frac{10}{3}\left(\frac{2}{3}\right)^{1/5}\approx3.07 C 3 = 3 10 ( 3 2 ) 1/5 ≈ 3.07 .
Critical points of H ^ \hat H H ^
The zeros of ∂ H ^ ∂ p ~ \frac{\partial\hat H}{\partial\tilde p} ∂ p ~ ∂ H ^ are characterized by
p ~ 2 + r ~ − 2 x ~ + 2 = 0. (30) \tilde p^2+\tilde r-2\tilde x_+^2=0. \tag*{(30)} p ~ 2 + r ~ − 2 x ~ + 2 = 0. (30)
From the definition of x ~ + \tilde x_+ x ~ + in (14) , we get 2 x ~ + 2 − r ~ = p ~ + A r ~ − B 2\tilde x_+^2-\tilde r=\tilde p+\sqrt{A\tilde r-B} 2 x ~ + 2 − r ~ = p ~ + A r ~ − B , which we insert into (30) to deduce
p ~ 2 = p ~ + A r ~ − B ⟺ A r ~ − B = p ~ 2 ( p ~ − 1 ) 2 . \tilde p^2=\tilde p+\sqrt{A\tilde r-B}\Longleftrightarrow A\tilde r-B=\tilde p^2(\tilde p-1)^2. p ~ 2 = p ~ + A r ~ − B ⟺ A r ~ − B = p ~ 2 ( p ~ − 1 ) 2 .
Using the definitions of A A A and B B B in (13) , we can simplify the above equation to
p ~ 4 = r ~ ( p ~ 2 + 2 p ~ ) . (31) \tilde p^4=\tilde r(\tilde p^2+2\tilde p). \tag*{(31)} p ~ 4 = r ~ ( p ~ 2 + 2 p ~ ) . (31)
We now recall that we also have ∂ H ^ ∂ r ~ ( r ~ , p ~ , ν ) = 0 \frac{\partial\hat H}{\partial\tilde r}(\tilde r,\tilde p,\nu)=0 ∂ r ~ ∂ H ^ ( r ~ , p ~ , ν ) = 0 , which by (26) implies that
4 ν p ~ x ~ + = − 2 p ~ ( p ~ 2 − 2 x ~ + 2 − r ~ ) = 4 p ~ r ~ , 4\nu\tilde p\tilde x_+=-2\tilde p(\tilde p^2-2\tilde x_+^2-\tilde r)=4\tilde p\tilde r, 4 ν p ~ x ~ + = − 2 p ~ ( p ~ 2 − 2 x ~ + 2 − r ~ ) = 4 p ~ r ~ ,
since p ~ 2 − 2 x ~ + 2 + r ~ = 0 \tilde p^2-2\tilde x_+^2+\tilde r=0 p ~ 2 − 2 x ~ + 2 + r ~ = 0 at the critical point. Thus, at the critical point, we have x ~ + = r ~ / ν \tilde x_+=\tilde r/\nu x ~ + = r ~ / ν , which we can again insert into (30) to get
p ~ 2 = 2 r ~ 2 ν 2 − r ~ = r ~ ( 2 r ~ ν 2 − 1 ) . (32) \tilde p^2=\frac{2\tilde r^2}{\nu^2}-\tilde r=\tilde r\left(\frac{2\tilde r}{\nu^2}-1\right). \tag*{(32)} p ~ 2 = ν 2 2 r ~ 2 − r ~ = r ~ ( ν 2 2 r ~ − 1 ) . (32)
We now seek to eliminate r ~ \tilde r r ~ from (31) and (32) . First, dividing (32) by (31) gives
p ~ − 2 = 2 r ~ / ν 2 − 1 p ~ 2 + 2 p ~ ⟹ r ~ = ν 2 ( 1 + 1 p ~ ) . \tilde p^{-2}=\frac{2\tilde r/\nu^2-1}{\tilde p^2+2\tilde p}\Longrightarrow\tilde r=\nu^2\left(1+\frac{1}{\tilde p}\right). p ~ − 2 = p ~ 2 + 2 p ~ 2 r ~ / ν 2 − 1 ⟹ r ~ = ν 2 ( 1 + p ~ 1 ) .
Substituting into (31) yields
p ~ 4 = ν 2 ( 1 + 1 p ~ ) ( p ~ 2 + 2 p ~ ) = ν 2 ( p ~ + 1 ) ( p ~ + 2 ) . (33) \tilde p^4=\nu^2\left(1+\frac{1}{\tilde p}\right)(\tilde p^2+2\tilde p)=\nu^2(\tilde p+1)(\tilde p+2). \tag*{(33)} p ~ 4 = ν 2 ( 1 + p ~ 1 ) ( p ~ 2 + 2 p ~ ) = ν 2 ( p ~ + 1 ) ( p ~ + 2 ) . (33)
We consider the case where T T T is very small, or where we wish to take many iterations of SWR, i.e., as ν → ∞ \nu\to\infty ν → ∞ . Letting u = p ~ / ν u=\tilde{p}/\nu u = p ~ / ν in (33) and dividing both sides by ν 4 \nu^4 ν 4 leads to
ν 4 u 4 = ν 2 ( ν u + 1 ) ( ν u + 2 ) ⟹ u 4 = ( u + ν − 1 ) ( u + 2 ν − 1 ) . \nu^4u^4=\nu^2(\nu u+1)(\nu u+2)\Longrightarrow u^4=(u+\nu^{-1})(u+2\nu^{-1}). ν 4 u 4 = ν 2 ( ν u + 1 ) ( ν u + 2 ) ⟹ u 4 = ( u + ν − 1 ) ( u + 2 ν − 1 ) .
We therefore seek a positive solution to u 4 = ( u + ϵ ) ( u + 2 ϵ ) u^4=(u+\epsilon)(u+2\epsilon) u 4 = ( u + ϵ ) ( u + 2 ϵ ) with ϵ : = ν − 1 \epsilon:=\nu^{-1} ϵ := ν − 1 , which approaches zero as ν → ∞ \nu\to\infty ν → ∞ . Since the right-hand side is greater than u 2 u^2 u 2 , we must have
u 4 > u 2 ⟹ u 2 ( u 2 − 1 ) > 0 ⟹ u > 1. u^4>u^2\Longrightarrow u^2(u^2-1)>0\Longrightarrow u>1. u 4 > u 2 ⟹ u 2 ( u 2 − 1 ) > 0 ⟹ u > 1.
Moreover, u → 1 + u\to1+ u → 1 + as ϵ → 0 + \epsilon\to0+ ϵ → 0 + . In the neighbourhood of u = 1 u=1 u = 1 , the implicit function theorem implies that u u u is a differentiable function of ϵ \epsilon ϵ and can be written as
u = 1 + 3 2 ϵ + O ( ϵ 2 ) . u=1+\frac{3}{2}\epsilon+O(\epsilon^2). u = 1 + 2 3 ϵ + O ( ϵ 2 ) .
Recalling that p ~ = ν u \tilde{p}=\nu u p ~ = ν u , we arrive at an asymptotic expansion of p ~ \tilde{p} p ~ for large ν \nu ν , which in turn yields those for r ~ \tilde{r} r ~ and x ~ \tilde{x} x ~ at this critical point ( r ~ ∗ , p ~ ∗ ) (\tilde{r}^{*},\tilde{p}^{*}) ( r ~ ∗ , p ~ ∗ ) , namely:
p ~ ∗ = ν ( 1 + 3 2 ϵ + O ( ϵ 2 ) ) , (34) \tilde{p}^{*}=\nu\left(1+\frac{3}{2}\epsilon+O(\epsilon^2)\right), \tag*{(34)} p ~ ∗ = ν ( 1 + 2 3 ϵ + O ( ϵ 2 ) ) , (34)
r ~ ∗ = ν 2 ( 1 + ϵ + O ( ϵ 2 ) ) , (35) \tilde{r}^{*}=\nu^2\left(1+\epsilon+O(\epsilon^2)\right), \tag*{(35)} r ~ ∗ = ν 2 ( 1 + ϵ + O ( ϵ 2 ) ) , (35)
x ~ + = ν ( 1 + ϵ + O ( ϵ 2 ) ) . (36) \tilde{x}_{+}=\nu\left(1+\epsilon+O(\epsilon^2)\right). \tag*{(36)} x ~ + = ν ( 1 + ϵ + O ( ϵ 2 ) ) . (36)
We are now ready to calculate the asymptotic expression for H ^ ( r ~ , p ~ , ν ) \hat{H}(\tilde{r},\tilde{p},\nu) H ^ ( r ~ , p ~ , ν ) at the critical point ( r ~ ∗ , p ~ ∗ ) (\tilde{r}^{*},\tilde{p}^{*}) ( r ~ ∗ , p ~ ∗ ) . Since
H ^ ( r ~ ∗ , p ~ ∗ , ν ) = h ( x ~ + , r ~ ∗ , p ~ ∗ , ν ) = r ~ ∗ − 2 ν x ~ + + ν log [ ( p ~ ∗ − x ~ + ) 2 + x ~ + 2 − r ~ ∗ ( p ~ ∗ + x ~ + ) 2 + x ~ + 2 − r ~ ∗ ] , \hat{H}(\tilde{r}^{*},\tilde{p}^{*},\nu)=h(\tilde{x}_{+},\tilde{r}^{*},\tilde{p}^{*},\nu)=\tilde{r}^{*}-2\nu\tilde{x}_{+}+\nu\log\left[\frac{(\tilde{p}^{*}-\tilde{x}_{+})^2+\tilde{x}_{+}^2-\tilde{r}^{*}}{(\tilde{p}^{*}+\tilde{x}_{+})^2+\tilde{x}_{+}^2-\tilde{r}^{*}}\right], H ^ ( r ~ ∗ , p ~ ∗ , ν ) = h ( x ~ + , r ~ ∗ , p ~ ∗ , ν ) = r ~ ∗ − 2 ν x ~ + + ν log [ ( p ~ ∗ + x ~ + ) 2 + x ~ + 2 − r ~ ∗ ( p ~ ∗ − x ~ + ) 2 + x ~ + 2 − r ~ ∗ ] ,
we compute the relevant parts of the expression as follows:
r ~ ∗ − 2 ν x ~ + = − r ~ ∗ = − ν 2 ( 1 + ϵ + O ( ϵ 2 ) ) , p ~ ∗ − x ~ + = ν ϵ ( 1 2 + O ( ϵ ) ) = 1 2 ϵ + O ( ϵ ) , p ~ ∗ + x ~ + = ν ( 2 + 5 2 ϵ + O ( ϵ 2 ) ) , x ~ + 2 − r ~ ∗ = x ~ + ( x ~ + − ν ) = ν 2 ϵ ( 1 + ϵ + O ( ϵ 2 ) ) = ν ( 1 + ϵ + O ( ϵ 2 ) ) , ( p ~ ∗ − x ~ + ) 2 + x ~ + 2 − r ~ ∗ ( p ~ ∗ + x ~ + ) 2 + x ~ + 2 − r ~ ∗ = ν ( 1 + 5 4 ϵ + O ( ϵ 2 ) ) ν 2 ( 4 + 11 ϵ + O ( ϵ 2 ) ) = 1 4 ϵ ( 1 − 3 2 ϵ + O ( ϵ 2 ) ) . \begin{aligned}
\tilde{r}^{*}-2\nu\tilde{x}_{+}&=-\tilde{r}^{*}=-\nu^2(1+\epsilon+O(\epsilon^2)),\\
\tilde{p}^{*}-\tilde{x}_{+}&=\nu\epsilon\left(\frac{1}{2}+O(\epsilon)\right)=\frac{1}{2}\epsilon+O(\epsilon),\\
\tilde{p}^{*}+\tilde{x}_{+}&=\nu\left(2+\frac{5}{2}\epsilon+O(\epsilon^2)\right),\\
\tilde{x}_{+}^2-\tilde{r}^{*}&=\tilde{x}_{+}(\tilde{x}_{+}-\nu)=\nu^2\epsilon(1+\epsilon+O(\epsilon^2))=\nu(1+\epsilon+O(\epsilon^2)),\\
\frac{(\tilde{p}^{*}-\tilde{x}_{+})^2+\tilde{x}_{+}^2-\tilde{r}^{*}}{(\tilde{p}^{*}+\tilde{x}_{+})^2+\tilde{x}_{+}^2-\tilde{r}^{*}}&=\frac{\nu(1+\frac{5}{4}\epsilon+O(\epsilon^2))}{\nu^2(4+11\epsilon+O(\epsilon^2))}=\frac{1}{4}\epsilon\left(1-\frac{3}{2}\epsilon+O(\epsilon^2)\right).
\end{aligned} r ~ ∗ − 2 ν x ~ + p ~ ∗ − x ~ + p ~ ∗ + x ~ + x ~ + 2 − r ~ ∗ ( p ~ ∗ + x ~ + ) 2 + x ~ + 2 − r ~ ∗ ( p ~ ∗ − x ~ + ) 2 + x ~ + 2 − r ~ ∗ = − r ~ ∗ = − ν 2 ( 1 + ϵ + O ( ϵ 2 )) , = ν ϵ ( 2 1 + O ( ϵ ) ) = 2 1 ϵ + O ( ϵ ) , = ν ( 2 + 2 5 ϵ + O ( ϵ 2 ) ) , = x ~ + ( x ~ + − ν ) = ν 2 ϵ ( 1 + ϵ + O ( ϵ 2 )) = ν ( 1 + ϵ + O ( ϵ 2 )) , = ν 2 ( 4 + 11 ϵ + O ( ϵ 2 )) ν ( 1 + 4 5 ϵ + O ( ϵ 2 )) = 4 1 ϵ ( 1 − 2 3 ϵ + O ( ϵ 2 ) ) .
H ^ ( r ~ ∗ , p ~ ∗ , ν ) = − ν 2 ( 1 + ϵ + O ( ϵ 2 ) ) + ν log ( ϵ / 4 ) − ν ( 3 2 ϵ + O ( ϵ 2 ) ) = − ν 2 ( 1 + ϵ ∣ log ϵ ∣ + ϵ ( 1 + ln 4 ) + O ( ϵ 2 ) ) . \begin{aligned}
\hat{H}(\tilde{r}^{*},\tilde{p}^{*},\nu)&=-\nu^2(1+\epsilon+O(\epsilon^2))+\nu\log(\epsilon/4)-\nu\left(\frac{3}{2}\epsilon+O(\epsilon^2)\right)\\
&=-\nu^2\left(1+\epsilon|\log\epsilon|+\epsilon(1+\ln4)+O(\epsilon^2)\right).
\end{aligned} H ^ ( r ~ ∗ , p ~ ∗ , ν ) = − ν 2 ( 1 + ϵ + O ( ϵ 2 )) + ν log ( ϵ /4 ) − ν ( 2 3 ϵ + O ( ϵ 2 ) ) = − ν 2 ( 1 + ϵ ∣ log ϵ ∣ + ϵ ( 1 + ln 4 ) + O ( ϵ 2 ) ) .
In order to see whether H ^ ( r ~ ∗ , p ~ ∗ , ν ) \hat{H}(\tilde{r}^{*},\tilde{p}^{*},\nu) H ^ ( r ~ ∗ , p ~ ∗ , ν ) yields a correct convergence bound, we need to compare this value with H 0 ( r ~ ∗ , p ~ ∗ , ν ) H_0(\tilde{r}^{*},\tilde{p}^{*},\nu) H 0 ( r ~ ∗ , p ~ ∗ , ν ) , which we can calculate by simply replacing x ~ + \tilde{x}_{+} x ~ + by r ~ ∗ \sqrt{\widetilde{r}^{*}} r ∗ in the expression for H ^ \widehat{H} H :
r ~ ∗ − 2 ν r ~ ∗ = − ν 2 ( 1 + O ( ϵ 2 ) ) , \widetilde{r}^{*}-2\nu\sqrt{\widetilde{r}^{*}}=-\nu^{2}\left(1+O(\epsilon^{2})\right), r ∗ − 2 ν r ∗ = − ν 2 ( 1 + O ( ϵ 2 ) ) ,
p ~ ∗ − r ~ ∗ = ν ϵ ( 1 + O ( ϵ ) ) , \widetilde{p}^{*}-\sqrt{\widetilde{r}^{*}}=\nu\epsilon\left(1+O(\epsilon)\right), p ∗ − r ∗ = ν ϵ ( 1 + O ( ϵ ) ) ,
p ~ ∗ + r ~ ∗ = ν ( 2 + 2 ϵ + O ( ϵ 2 ) ) , \widetilde{p}^{*}+\sqrt{\widetilde{r}^{*}}=\nu\left(2+2\epsilon+O(\epsilon^{2})\right), p ∗ + r ∗ = ν ( 2 + 2 ϵ + O ( ϵ 2 ) ) ,
p ~ ∗ − r ~ ∗ p ~ ∗ + r ~ ∗ = ϵ 2 + O ( ϵ 2 ) \frac{\widetilde{p}^{*}-\sqrt{\widetilde{r}^{*}}}{\widetilde{p}^{*}+\sqrt{\widetilde{r}^{*}}}=\frac{\epsilon}{2}+O(\epsilon^{2}) p ∗ + r ∗ p ∗ − r ∗ = 2 ϵ + O ( ϵ 2 )
H 0 ( r ~ ∗ , p ~ ∗ , ν ) = − ν 2 ( 1 + O ( ϵ 2 ) ) − 2 ν ∣ log ( ϵ / 2 + O ( ϵ 2 ) ) ∣ = − ν 2 ( 1 + 2 ϵ ∣ log ϵ ∣ + O ( ϵ ) ) . \begin{aligned}
H_{0}(\widetilde{r}^{*},\widetilde{p}^{*},\nu)=-\nu^{2}\left(1+O(\epsilon^{2})\right)-2\nu\left|\log\left(\epsilon/2+O(\epsilon^{2})\right)\right| \\
&=-\nu^{2}\left(1+2\epsilon|\log\epsilon|+O(\epsilon)\right).
\end{aligned} H 0 ( r ∗ , p ∗ , ν ) = − ν 2 ( 1 + O ( ϵ 2 ) ) − 2 ν log ( ϵ /2 + O ( ϵ 2 ) ) = − ν 2 ( 1 + 2 ϵ ∣ log ϵ ∣ + O ( ϵ ) ) .
Since H 0 ( r ~ ∗ , p ~ ∗ , ν ) < H ^ ( r ~ ∗ , p ~ ∗ , ν ) H_{0}(\widetilde{r}^{*},\widetilde{p}^{*},\nu)<\widehat{H}(\widetilde{r}^{*},\widetilde{p}^{*},\nu) H 0 ( r ∗ , p ∗ , ν ) < H ( r ∗ , p ∗ , ν ) asymptotically, we conclude that H ^ ( r ~ ∗ , p ~ ∗ , ν ) \widehat{H}(\widetilde{r}^{*},\widetilde{p}^{*},\nu) H ( r ∗ , p ∗ , ν ) is a valid convergence bound. Note that by choosing this p ∗ = δ p ~ ∗ p^{*}=\delta\widetilde{p}^{*} p ∗ = δ p ∗ as the Robin parameter, we obtain an n n n -step contraction factor of the form
ρ ( r ∗ , p ∗ , n ) = exp ( − T δ 2 ( ν 2 + ν log ν + ν ( 1 + ln 4 ) + O ( 1 ) ) ) ≃ e − n δ 2 / T ( C n δ 2 T ) − n . \rho(r^{*},p^{*},n)=\exp\left(-\frac{T}{\delta^{2}}\left(\nu^{2}+\nu\log\nu+\nu(1+\ln4)+O(1)\right)\right)\simeq e^{-n\delta^{2}/T}\left(\frac{C n\delta^{2}}{T}\right)^{-n}. ρ ( r ∗ , p ∗ , n ) = exp ( − δ 2 T ( ν 2 + ν log ν + ν ( 1 + ln 4 ) + O ( 1 ) ) ) ≃ e − n δ 2 / T ( T C n δ 2 ) − n .
where C = 4 e ≈ 10.87 C=4e\approx10.87 C = 4 e ≈ 10.87 . We therefore gain an extra factor of ( C n δ 2 T ) − n \left(\frac{C n\delta^{2}}{T}\right)^{-n} ( T C n δ 2 ) − n relative to the Dirichlet bound, which indicates faster convergence than for Dirichlet transmission conditions.
A different scaling regime consists of letting ν → 0 \nu\to0 ν → 0 , i.e., fixing the number of iterations n n n and considering what happens in the limit of a vanishing overlap δ → 0 \delta\to0 δ → 0 (or, equivalently, that of a long time horizon T → ∞ T\to\infty T → ∞ ). In this case, making the substitution p ~ = ν 1 / 2 u \widetilde{p}=\nu^{1/2}u p = ν 1/2 u in (33) leads to
u 4 = ( ν 1 / 2 u + 1 ) ( ν 1 / 2 u + 2 ) = 2 + ϵ ( 3 u + ϵ u 2 ) , where ϵ = ν 1 / 2 . u^{4}=(\nu^{1/2}u+1)(\nu^{1/2}u+2)=2+\epsilon(3u+\epsilon u^{2}),\qquad\text{where }\epsilon=\nu^{1/2}. u 4 = ( ν 1/2 u + 1 ) ( ν 1/2 u + 2 ) = 2 + ϵ ( 3 u + ϵ u 2 ) , where ϵ = ν 1/2 .
Taking the limit as ν → 0 \nu\to0 ν → 0 leads to u 4 = 2 u^{4}=2 u 4 = 2 , whose only positive solution is u ˉ = 2 1 / 4 \bar{u}=2^{1/4} u ˉ = 2 1/4 . By Rouché’s theorem, (33) will possess a simple root near u ˉ \bar{u} u ˉ for ν \nu ν small enough; moreover, the trajectory of this root is differentiable with respect to ϵ \epsilon ϵ , since u ˉ \bar{u} u ˉ is a simple root of the limiting equation. Thus, to obtain the asymptotics, we write u ∗ = u ˉ + c 1 ϵ + O ( ϵ 2 ) u^{*}=\bar{u}+c_{1}\epsilon+O(\epsilon^{2}) u ∗ = u ˉ + c 1 ϵ + O ( ϵ 2 ) and insert it into (33) . Solving for c 1 c_{1} c 1 leads to
u ∗ = 2 1 / 4 + 3 2 5 / 2 ϵ + O ( ϵ 2 ) . u^{*}=2^{1/4}+\frac{3}{2^{5/2}}\epsilon+O(\epsilon^{2}). u ∗ = 2 1/4 + 2 5/2 3 ϵ + O ( ϵ 2 ) .
We can now compute the asymptotic expansions of p ~ ∗ , r ~ ∗ \widetilde{p}^{*},\widetilde{r}^{*} p ∗ , r ∗ as before. This leads to
p ~ ∗ = ν 1 / 2 ( 2 1 / 4 + 3 2 5 / 2 ν 1 / 2 + O ( ν ) ) , r ~ ∗ = ν 3 / 2 ( 2 − 1 / 4 − 5 8 ν 1 / 2 + O ( ν ) ) . \widetilde{p}^{*}=\nu^{1/2}\left(2^{1/4}+\frac{3}{2^{5/2}}\nu^{1/2}+O(\nu)\right),\qquad\widetilde{r}^{*}=\nu^{3/2}\left(2^{-1/4}-\frac{5}{8}\nu^{1/2}+O(\nu)\right). p ∗ = ν 1/2 ( 2 1/4 + 2 5/2 3 ν 1/2 + O ( ν ) ) , r ∗ = ν 3/2 ( 2 − 1/4 − 8 5 ν 1/2 + O ( ν ) ) .
We now continue the calculation to obtain the asymptotics for H ^ ( x ~ + , r ~ ∗ , p ~ ∗ , ν ) \widehat{H}(\widetilde{x}_{+},\widetilde{r}^{*},\widetilde{p}^{*},\nu) H ( x + , r ∗ , p ∗ , ν ) . Note that the coefficients are given as floating point numbers in order to facilitate comparisons with H 0 ( x ~ + , r ~ ∗ , p ~ ∗ , ν ) H_0(\tilde{x}_+,\tilde{r}^*,\tilde{p}^*,\nu) H 0 ( x ~ + , r ~ ∗ , p ~ ∗ , ν ) later:
r ~ ∗ − 2 ν x ~ + = − r ~ ∗ = − ν 3 / 2 ( 0.8409 − 0.625 ν 1 / 2 + O ( ν ) ) , \tilde{r}^* - 2\nu\tilde{x}_+ = -\tilde{r}^* = -\nu^{3/2}\left(0.8409 - 0.625\nu^{1/2} + O(\nu)\right), r ~ ∗ − 2 ν x ~ + = − r ~ ∗ = − ν 3/2 ( 0.8409 − 0.625 ν 1/2 + O ( ν ) ) ,
p ~ ∗ − x ~ + = ν 1 / 2 ( 0.3483 + 1.1553 ν 1 / 2 + O ( ν ) ) , \tilde{p}^* - \tilde{x}_+ = \nu^{1/2}\left(0.3483 + 1.1553\nu^{1/2} + O(\nu)\right), p ~ ∗ − x ~ + = ν 1/2 ( 0.3483 + 1.1553 ν 1/2 + O ( ν ) ) ,
p ~ ∗ + x ~ + = ν 1 / 2 ( 2.0301 − 0.0947 ν 1 / 2 + O ( ν ) ) , \tilde{p}^* + \tilde{x}_+ = \nu^{1/2}\left(2.0301 - 0.0947\nu^{1/2} + O(\nu)\right), p ~ ∗ + x ~ + = ν 1/2 ( 2.0301 − 0.0947 ν 1/2 + O ( ν ) ) ,
x ~ + 2 − r ~ = ν ( 0.707 − 1.892 ν 1 / 2 + O ( ν ) ) , \tilde{x}_+^2 - \tilde{r} = \nu\left(0.707 - 1.892\nu^{1/2} + O(\nu)\right), x ~ + 2 − r ~ = ν ( 0.707 − 1.892 ν 1/2 + O ( ν ) ) ,
( p ~ ∗ − x ~ + ) 2 + x ~ + 2 − r ~ ∗ ( p ~ ∗ + x ~ + ) 2 + x ~ + 2 − r ~ ∗ = ν ( 0.8283 − 1.0872 ν 1 / 2 + O ( ν ) ) ν ( 4.8283 − 2.2765 ν 1 / 2 + O ( ν ) ) = 0.1716 + 0.2463 ν 1 / 2 + O ( ν ) , \frac{(\tilde{p}^*-\tilde{x}_+)^2+\tilde{x}_+^2-\tilde{r}^*}{(\tilde{p}^*+\tilde{x}_+)^2+\tilde{x}_+^2-\tilde{r}^*}
=\frac{\nu(0.8283-1.0872\nu^{1/2}+O(\nu))}{\nu(4.8283-2.2765\nu^{1/2}+O(\nu))}
=0.1716+0.2463\nu^{1/2}+O(\nu), ( p ~ ∗ + x ~ + ) 2 + x ~ + 2 − r ~ ∗ ( p ~ ∗ − x ~ + ) 2 + x ~ + 2 − r ~ ∗ = ν ( 4.8283 − 2.2765 ν 1/2 + O ( ν )) ν ( 0.8283 − 1.0872 ν 1/2 + O ( ν )) = 0.1716 + 0.2463 ν 1/2 + O ( ν ) ,
H ^ ( r ~ ∗ , p ~ ∗ , ν ) = − ν 3 / 2 ( 0.8409 + O ( ν 1 / 2 ) ) + ν ( log ( 0.1716 ) + 1.4353 ν 1 / 2 + O ( ν ) ) = − 1.7626 ν + 0.5944 ν 3 / 2 + O ( ν 2 ) . \begin{aligned}
\hat{H}(\tilde{r}^*,\tilde{p}^*,\nu)
=-\nu^{3/2}\left(0.8409+O(\nu^{1/2})\right)
+\nu\left(\log(0.1716)+1.4353\nu^{1/2}+O(\nu)\right) \\
&=-1.7626\nu+0.5944\nu^{3/2}+O(\nu^2).
\end{aligned} H ^ ( r ~ ∗ , p ~ ∗ , ν ) = − ν 3/2 ( 0.8409 + O ( ν 1/2 ) ) + ν ( log ( 0.1716 ) + 1.4353 ν 1/2 + O ( ν ) ) = − 1.7626 ν + 0.5944 ν 3/2 + O ( ν 2 ) .
As before, we compare the above value with H 0 ( r ∗ , p ∗ , k ) H_0(r^*,p^*,k) H 0 ( r ∗ , p ∗ , k ) :
r ~ ∗ − 2 ν r ~ ∗ = ν 3 / 2 ( 2 − 1 / 4 + O ( ν 1 / 4 ) ) , \tilde{r}^*-2\nu\sqrt{\tilde{r}^*}=\nu^{3/2}\left(2^{-1/4}+O(\nu^{1/4})\right), r ~ ∗ − 2 ν r ~ ∗ = ν 3/2 ( 2 − 1/4 + O ( ν 1/4 ) ) ,
p ~ ∗ ± r ~ ∗ = ν 1 / 2 2 1 / 4 ( 1 ± 2 − 3 / 8 ν 1 / 4 + 3 ⋅ 2 − 11 / 4 ν 1 / 2 + O ( ν 3 / 4 ) ) , \tilde{p}^*\pm\sqrt{\tilde{r}^*}
=\nu^{1/2}2^{1/4}\left(1\pm2^{-3/8}\nu^{1/4}+3\cdot2^{-11/4}\nu^{1/2}+O(\nu^{3/4})\right), p ~ ∗ ± r ~ ∗ = ν 1/2 2 1/4 ( 1 ± 2 − 3/8 ν 1/4 + 3 ⋅ 2 − 11/4 ν 1/2 + O ( ν 3/4 ) ) ,
H 0 ( r ~ ∗ , p ~ ∗ , ν ) = − 3.0844 ν 5 / 4 + O ( ν 7 / 4 ) . H_0(\tilde{r}^*,\tilde{p}^*,\nu)=-3.0844\nu^{5/4}+O(\nu^{7/4}). H 0 ( r ~ ∗ , p ~ ∗ , ν ) = − 3.0844 ν 5/4 + O ( ν 7/4 ) .
This time we have H 0 ( r ~ ∗ , p ~ ∗ , ν ) > H ^ ( r ~ ∗ , p ~ ∗ , ν ) H_0(\tilde{r}^*,\tilde{p}^*,\nu)>\hat{H}(\tilde{r}^*,\tilde{p}^*,\nu) H 0 ( r ~ ∗ , p ~ ∗ , ν ) > H ^ ( r ~ ∗ , p ~ ∗ , ν ) whenever ν = n δ 2 / T ≲ 0.0728 \nu=n\delta^2/T\lesssim0.0728 ν = n δ 2 / T ≲ 0.0728 , meaning that this critical point does not yield a valid convergence bound as ν → 0 \nu\to0 ν → 0 . We need to look instead for a constrained minimum along the equioscillation curve to obtain a valid bound in this case.
Constrained minimum along the equioscillation curve
We start by noting that the equioscillation curve r ~ × = r ~ × ( p ~ ) \tilde{r}^{\times}=\tilde{r}^{\times}(\tilde{p}) r ~ × = r ~ × ( p ~ ) is independent of ν \nu ν , since
γ ( r ~ , p ~ ) ≔ ν − 1 ( H ^ ( r ~ , p ~ , ν ) − H 0 ( r ~ , p ~ , ν ) ) = 2 ( r ~ − x ~ + ) + log ( ( p ~ − x ~ + ) 2 + x ~ + 2 − r ~ ( p ~ + x ~ + ) 2 + x ~ + 2 − r ~ ) − log ( ( p ~ − r ~ ) 2 ( p ~ + r ~ ) 2 ) (37) \begin{aligned}
\gamma(\tilde{r},\tilde{p})\coloneqq\nu^{-1}\left(\hat{H}(\tilde{r},\tilde{p},\nu)-H_0(\tilde{r},\tilde{p},\nu)\right) \\
&=2(\sqrt{\tilde{r}}-\tilde{x}_+)+\log\left(\frac{(\tilde{p}-\tilde{x}_+)^2+\tilde{x}_+^2-\tilde{r}}{(\tilde{p}+\tilde{x}_+)^2+\tilde{x}_+^2-\tilde{r}}\right)-\log\left(\frac{(\tilde{p}-\sqrt{\tilde{r}})^2}{(\tilde{p}+\sqrt{\tilde{r}})^2}\right) \tag*{(37)}
\end{aligned} γ ( r ~ , p ~ ) : = ν − 1 ( H ^ ( r ~ , p ~ , ν ) − H 0 ( r ~ , p ~ , ν ) ) = 2 ( r ~ − x ~ + ) + log ( ( p ~ + x ~ + ) 2 + x ~ + 2 − r ~ ( p ~ − x ~ + ) 2 + x ~ + 2 − r ~ ) − log ( ( p ~ + r ~ ) 2 ( p ~ − r ~ ) 2 ) (37)
is independent of ν \nu ν (including x ~ + = x ~ + ( r ~ , p ~ ) \tilde{x}_+=\tilde{x}_+(\tilde{r},\tilde{p}) x ~ + = x ~ + ( r ~ , p ~ ) ). However, the constrained minimum will depend on ν \nu ν , since we are minimizing the function H 0 ( r ~ , p ~ , ν ) H_0(\tilde{r},\tilde{p},\nu) H 0 ( r ~ , p ~ , ν ) there, and the latter depends on ν \nu ν . Our strategy, as before, is to find the correct scaling that reveals the asymptotic behaviour of p ~ \tilde{p} p ~ , r ~ \tilde{r} r ~ and x ~ + \tilde{x}_+ x ~ + as ν → 0 \nu\to0 ν → 0 . To do so, we will need the following lemma, whose proof appears in Appendix A.
Lemma 4.3. Let p ~ ∈ ( 0 , 5 ) \tilde{p}\in(0,\sqrt{5}) p ~ ∈ ( 0 , 5 ) be given, and let r ~ ∈ [ 0 , r ~ − ( p ~ ) ] \tilde{r}\in[0,\tilde{r}_-(\tilde{p})] r ~ ∈ [ 0 , r ~ − ( p ~ )] be such that 0 ≤ r ~ × ≤ r ~ − 0\leq\tilde{r}^{\times}\leq\tilde{r}_- 0 ≤ r ~ × ≤ r ~ − and H 0 ( r ~ × , p ~ , ν ) = H ^ ( r ~ × , p ~ , ν ) H_0(\tilde{r}^{\times},\tilde{p},\nu)=\hat{H}(\tilde{r}^{\times},\tilde{p},\nu) H 0 ( r ~ × , p ~ , ν ) = H ^ ( r ~ × , p ~ , ν ) (cf. Theorem 3.3 ). Then there exist positive constants C 1 C_1 C 1 and C 2 C_2 C 2 , independent of p ~ \tilde{p} p ~ , such that C 1 p ~ 3 ≤ r ~ × ≤ C 2 p ~ 3 C_1\tilde{p}^3\leq\tilde{r}^{\times}\leq C_2\tilde{p}^3 C 1 p ~ 3 ≤ r ~ × ≤ C 2 p ~ 3 .
Motivated by the above lemma, we make the substitution r ~ × = c 2 p ~ 3 \tilde{r}^{\times}=c^2\tilde{p}^3 r ~ × = c 2 p ~ 3 and x ~ + = ξ p ~ \tilde{x}_+=\xi\sqrt{\tilde{p}} x ~ + = ξ p ~ into γ ( r ~ , p ~ ) = 0 \gamma(\tilde{r},\tilde{p})=0 γ ( r ~ , p ~ ) = 0 to get
0 = 2 ( c p ~ 3 / 2 − ξ p ~ 1 / 2 ) + log ( p ~ [ ( p ~ 1 / 2 − ξ ) 2 + ξ 2 − c 2 p ~ 2 ] p ~ [ ( p ~ 1 / 2 + ξ ) 2 + ξ 2 − c 2 p ~ 2 ] ) − log ( p ~ 2 ( 1 − c p ~ 1 / 2 ) 2 p ~ 2 ( 1 + c p ~ 1 / 2 ) 2 ) \begin{aligned}
0={}&2(c\tilde{p}^{3/2}-\xi\tilde{p}^{1/2})+\log\left(\frac{\tilde{p}\left[(\tilde{p}^{1/2}-\xi)^2+\xi^2-c^2\tilde{p}^2\right]}{\tilde{p}\left[(\tilde{p}^{1/2}+\xi)^2+\xi^2-c^2\tilde{p}^2\right]}\right)\\
&-\log\left(\frac{\tilde{p}^2(1-c\tilde{p}^{1/2})^2}{\tilde{p}^2(1+c\tilde{p}^{1/2})^2}\right)
\end{aligned} 0 = 2 ( c p ~ 3/2 − ξ p ~ 1/2 ) + log ( p ~ [ ( p ~ 1/2 + ξ ) 2 + ξ 2 − c 2 p ~ 2 ] p ~ [ ( p ~ 1/2 − ξ ) 2 + ξ 2 − c 2 p ~ 2 ] ) − log ( p ~ 2 ( 1 + c p ~ 1/2 ) 2 p ~ 2 ( 1 − c p ~ 1/2 ) 2 )
= 2 p ~ 1 / 2 ( c p ~ − ξ ) + log ( ( p ~ 1 / 2 − ξ ) 2 + ξ 2 − c 2 p ~ 2 ( p ~ 1 / 2 + ξ ) 2 + ξ 2 − c 2 p ~ 2 ) − log ( ( 1 − c p ~ 1 / 2 ) 2 ( 1 + c p ~ 1 / 2 ) 2 ) . (38) \begin{aligned}
&=2\tilde{p}^{1/2}(c\tilde{p}-\xi)+\log\left(\frac{(\tilde{p}^{1/2}-\xi)^2+\xi^2-c^2\tilde{p}^2}{(\tilde{p}^{1/2}+\xi)^2+\xi^2-c^2\tilde{p}^2}\right)-\log\left(\frac{(1-c\tilde{p}^{1/2})^2}{(1+c\tilde{p}^{1/2})^2}\right). \tag*{(38)}
\end{aligned} = 2 p ~ 1/2 ( c p ~ − ξ ) + log ( ( p ~ 1/2 + ξ ) 2 + ξ 2 − c 2 p ~ 2 ( p ~ 1/2 − ξ ) 2 + ξ 2 − c 2 p ~ 2 ) − log ( ( 1 + c p ~ 1/2 ) 2 ( 1 − c p ~ 1/2 ) 2 ) . (38)
Additionally, the definition of x ~ + \tilde{x}_+ x ~ + in (14) leads to the relation
2 x ~ + 2 = 2 ξ 2 p ~ = c 2 p ~ 3 + p ~ + ( 4 p ~ + 2 p ~ 2 ) c 2 p ~ 3 − ( p ~ 4 + 2 p ~ 3 − p ~ 2 ) , 2\tilde{x}_+^2=2\xi^2\tilde{p}=c^2\tilde{p}^3+\tilde{p}+\sqrt{(4\tilde{p}+2\tilde{p}^2)c^2\tilde{p}^3-(\tilde{p}^4+2\tilde{p}^3-\tilde{p}^2)}, 2 x ~ + 2 = 2 ξ 2 p ~ = c 2 p ~ 3 + p ~ + ( 4 p ~ + 2 p ~ 2 ) c 2 p ~ 3 − ( p ~ 4 + 2 p ~ 3 − p ~ 2 ) ,
which implies
2 ξ 2 = 1 + c 2 p ~ 2 + c 2 ( 4 p ~ 2 + 2 p ~ 3 ) + ( 1 − 2 p ~ − p ~ 2 ) . (39) 2\xi^2 = 1 + c^2\tilde{p}^2 + \sqrt{c^2(4\tilde{p}^2 + 2\tilde{p}^3) + (1 - 2\tilde{p} - \tilde{p}^2)}. \tag*{(39)} 2 ξ 2 = 1 + c 2 p ~ 2 + c 2 ( 4 p ~ 2 + 2 p ~ 3 ) + ( 1 − 2 p ~ − p ~ 2 ) . (39)
We wish to find the simultaneous solution ( c , ξ ) (c,\xi) ( c , ξ ) of (38) –(39) , and particularly in the limit of p ~ → 0 \tilde{p} \to0 p ~ → 0 , since p ~ ∗ → 0 \tilde{p}^{*} \to0 p ~ ∗ → 0 as ν → 0 \nu\to0 ν → 0 . Indeed, p ~ ∗ \tilde{p}^{*} p ~ ∗ must be bounded because we are in the regime where 0 < p ~ < 5 0 < \tilde{p} < \sqrt{5} 0 < p ~ < 5 (otherwise the equioscillation curve plays no role in the minimization problem). If the optimal p ~ \tilde{p} p ~ approaches a positive constant as ν → 0 \nu\to0 ν → 0 , then r ~ ∗ \tilde{r}^{*} r ~ ∗ and x ~ + \tilde{x}_{+} x ~ + would also approach a positive constant, meaning that H 0 ( r ~ , p ~ , ν ) H_{0}(\tilde{r},\tilde{p},\nu) H 0 ( r ~ , p ~ , ν ) approaches a constant as well. This would lead to an asymptotic bound that does not contain ν \nu ν , i.e., it does not contain the iteration number n n n . Such a bound would be useless in practice.
Letting p ~ → 0 \tilde{p} \to0 p ~ → 0 in (39) leads to ξ = 1 \xi= 1 ξ = 1 . Next, we rewrite (38) as
ξ − c p ~ = 1 2 p ~ 1 / 2 log ( ( p ~ 1 / 2 − ξ ) 2 + ξ 2 − c 2 p ~ 2 ( p ~ 1 / 2 + ξ ) 2 + ξ 2 − c 2 p ~ 2 ) − 1 2 p ~ 1 / 2 log ( ( 1 − c p ~ 1 / 2 ) 2 ( 1 + c p ~ 1 / 2 ) 2 ) . (40) \xi-c\tilde{p}=\frac{1}{2\tilde{p}^{1/2}}\log\left(\frac{(\tilde{p}^{1/2}-\xi)^2+\xi^2-c^2\tilde{p}^2}{(\tilde{p}^{1/2}+\xi)^2+\xi^2-c^2\tilde{p}^2}\right)-\frac{1}{2\tilde{p}^{1/2}}\log\left(\frac{(1-c\tilde{p}^{1/2})^2}{(1+c\tilde{p}^{1/2})^2}\right). \tag*{(40)} ξ − c p ~ = 2 p ~ 1/2 1 log ( ( p ~ 1/2 + ξ ) 2 + ξ 2 − c 2 p ~ 2 ( p ~ 1/2 − ξ ) 2 + ξ 2 − c 2 p ~ 2 ) − 2 p ~ 1/2 1 log ( ( 1 + c p ~ 1/2 ) 2 ( 1 − c p ~ 1/2 ) 2 ) . (40)
Taking the limit p ~ → 0 \tilde{p} \to0 p ~ → 0 on both sides yields
ξ = − 1 ξ + 2 c ⟹ c = 1. \xi=-\frac{1}{\xi}+2c \Longrightarrow c=1. ξ = − ξ 1 + 2 c ⟹ c = 1.
We therefore conclude that r ~ ∗ = p ~ 3 + O ( p ~ 4 ) \tilde{r}^{*}=\tilde{p}^{3}+O(\tilde{p}^{4}) r ~ ∗ = p ~ 3 + O ( p ~ 4 ) . Moreover, after dividing (38) by p ~ 1 / 2 \tilde{p}^{1/2} p ~ 1/2 , we see that the Jacobian of the system (38) –(39) with respect to ( c , ξ ) (c,\xi) ( c , ξ ) is invertible in the neighbourhood of ( c , ξ , p ~ ) = ( 1 , 1 , 0 ) (c,\xi,\tilde{p})=(1,1,0) ( c , ξ , p ~ ) = ( 1 , 1 , 0 ) ; thus, c c c and ξ \xi ξ are locally differentiable (in fact, analytic) with respect to p ~ \tilde{p} p ~ . Substituting the Taylor series of these functions into (38) –(39) and equating coefficients leads to
ξ = 1 − 1 4 p ~ + 15 32 p ~ 2 + O ( p ~ 3 ) , c = 1 − 11 12 p ~ + 219 160 p ~ 2 + O ( p ~ 3 ) . \xi=1-\frac{1}{4}\tilde{p}+\frac{15}{32}\tilde{p}^2+O(\tilde{p}^3),\qquad c=1-\frac{11}{12}\tilde{p}+\frac{219}{160}\tilde{p}^2+O(\tilde{p}^3). ξ = 1 − 4 1 p ~ + 32 15 p ~ 2 + O ( p ~ 3 ) , c = 1 − 12 11 p ~ + 160 219 p ~ 2 + O ( p ~ 3 ) .
We are now ready to prove the asymptotic optimal parameter p ~ \tilde{p} p ~ along the equioscillation curve.
Theorem 4.4. Let p ~ ∈ ( 0 , 5 ) \tilde{p}\in(0,\sqrt{5}) p ~ ∈ ( 0 , 5 ) and r ~ × = r ~ × ( p ~ ) \tilde{r}^{\times}=\tilde{r}^{\times}(\tilde{p}) r ~ × = r ~ × ( p ~ ) be the equioscillation curve satisfying H 0 ( r ~ × , p ~ , ν ) = H ^ ( r ~ × , p ~ , ν ) H_{0}(\tilde{r}^{\times},\tilde{p},\nu)=\hat{H}(\tilde{r}^{\times},\tilde{p},\nu) H 0 ( r ~ × , p ~ , ν ) = H ^ ( r ~ × , p ~ , ν ) . Then the minimum of p ~ ↦ H 0 ( r ~ × ( p ~ ) , p ~ , ν ) \tilde{p}\mapsto H_{0}(\tilde{r}^{\times}(\tilde{p}),\tilde{p},\nu) p ~ ↦ H 0 ( r ~ × ( p ~ ) , p ~ , ν ) has the asymptotic behaviour
p ~ ∗ = ( 2 3 ) 2 / 5 ν 2 / 5 + K ν 4 / 5 + O ( ν 6 / 5 ) , H 0 ( r ~ × ( p ~ ∗ ) , p ~ ∗ , ν ) ∼ − C ν 6 / 5 \tilde{p}^{*}=\left(\frac{2}{3}\right)^{2/5}\nu^{2/5}+K\nu^{4/5}+O(\nu^{6/5}),\qquad H_{0}(\tilde{r}^{\times}(\tilde{p}^{*}),\tilde{p}^{*},\nu)\sim-C\nu^{6/5} p ~ ∗ = ( 3 2 ) 2/5 ν 2/5 + K ν 4/5 + O ( ν 6/5 ) , H 0 ( r ~ × ( p ~ ∗ ) , p ~ ∗ , ν ) ∼ − C ν 6/5 with K ≈ 0.6346 K\approx0.6346 K ≈ 0.6346 and C = 10 3 ( 2 3 ) 1 / 5 ≈ 3.07 C=\frac{10}{3}\left(\frac{2}{3}\right)^{1/5}\approx3.07 C = 3 10 ( 3 2 ) 1/5 ≈ 3.07 , where we write f ∼ α ν β f\sim\alpha\nu^{\beta} f ∼ α ν β whenever f = α ν β + o ( ν β ) f=\alpha\nu^{\beta}+o(\nu^{\beta}) f = α ν β + o ( ν β ) as ν → 0 \nu\to0 ν → 0 .
Proof. Let V ( p ~ ) = H 0 ( r ~ × ( p ~ ) , p ~ , ν ) V(\tilde{p})=H_{0}(\tilde{r}^{\times}(\tilde{p}),\tilde{p},\nu) V ( p ~ ) = H 0 ( r ~ × ( p ~ ) , p ~ , ν ) . We seek the value of p ~ \tilde{p} p ~ minimizing V V V by finding a root of V ′ ( p ~ ) V'(\tilde{p}) V ′ ( p ~ ) . From (22) and (28) , we see that
V ′ ( p ~ ) = ∂ H 0 ∂ p ~ ( r ~ × ( p ~ ) , p ~ , ν ) + ∂ H 0 ∂ r ~ ( r ~ × ( p ~ ) , p ~ , ν ) d r ~ × d p ~ = [ 4 ν r ~ × p ~ 2 − r ~ × + ( 1 − ν r ~ × − 2 ν p ~ r ~ × ( p ~ 2 − r ~ × ) ) d r ~ × d p ~ ] . \begin{aligned}
V'(\tilde{p})=\frac{\partial H_{0}}{\partial\tilde{p}}(\tilde{r}^{\times}(\tilde{p}),\tilde{p},\nu)+\frac{\partial H_{0}}{\partial\tilde{r}}(\tilde{r}^{\times}(\tilde{p}),\tilde{p},\nu)\frac{\mathrm{d}\tilde{r}^{\times}}{\mathrm{d}\tilde{p}} \\
&=\left[\frac{4\nu\sqrt{\tilde{r}^{\times}}}{\tilde{p}^{2}-\tilde{r}^{\times}}+\left(1-\frac{\nu}{\sqrt{\tilde{r}^{\times}}}-\frac{2\nu\tilde{p}}{\sqrt{\tilde{r}^{\times}}(\tilde{p}^{2}-\tilde{r}^{\times})}\right)\frac{\mathrm{d}\tilde{r}^{\times}}{\mathrm{d}\tilde{p}}\right].
\end{aligned} V ′ ( p ~ ) = ∂ p ~ ∂ H 0 ( r ~ × ( p ~ ) , p ~ , ν ) + ∂ r ~ ∂ H 0 ( r ~ × ( p ~ ) , p ~ , ν ) d p ~ d r ~ × = [ p ~ 2 − r ~ × 4 ν r ~ × + ( 1 − r ~ × ν − r ~ × ( p ~ 2 − r ~ × ) 2 ν p ~ ) d p ~ d r ~ × ] .
Setting V ′ ( p ~ ) = 0 V'(\tilde p)=0 V ′ ( p ~ ) = 0 leads to
ν r ~ × ( p ~ 2 − r ~ × ) ( ( ( p ~ 2 − r ~ × ) + 2 p ~ ) d r ~ × d p ~ − 4 r ~ × ) = d r ~ × d p ~ . \frac{\nu}{\sqrt{\tilde r^{\times}(\tilde p^2-\tilde r^{\times})}}\left(((\tilde p^2-\tilde r^{\times})+2\tilde p)\frac{d\tilde r^{\times}}{d\tilde p}-4\tilde r^{\times}\right)=\frac{d\tilde r^{\times}}{d\tilde p}. r ~ × ( p ~ 2 − r ~ × ) ν ( (( p ~ 2 − r ~ × ) + 2 p ~ ) d p ~ d r ~ × − 4 r ~ × ) = d p ~ d r ~ × .
leading to
ν = r ~ × ( p ~ 2 − r ~ × ) ( r ~ × ) ′ ( ( p ~ 2 − r ~ × ) + 2 p ~ ) ( r ~ × ) ′ − 4 r ~ × (41) \nu=\frac{\sqrt{\tilde r^{\times}}(\tilde p^2-\tilde r^{\times})(\tilde r^{\times})'}{((\tilde p^2-\tilde r^{\times})+2\tilde p)(\tilde r^{\times})'-4\tilde r^{\times}} \tag*{(41)} ν = (( p ~ 2 − r ~ × ) + 2 p ~ ) ( r ~ × ) ′ − 4 r ~ × r ~ × ( p ~ 2 − r ~ × ) ( r ~ × ) ′ (41)
Since r ~ × = c 2 p ~ 3 \tilde r^{\times}=c^2\tilde p^3 r ~ × = c 2 p ~ 3 , we have the asymptotic expansions
r ~ × = c p ~ 3 / 2 = p ~ 3 / 2 ( 1 − 11 12 p ~ + 219 160 p ~ 2 + O ( p ~ 3 ) ) , \sqrt{\tilde r^{\times}}=c\tilde p^{3/2}=\tilde p^{3/2}\left(1-\frac{11}{12}\tilde p+\frac{219}{160}\tilde p^2+O(\tilde p^3)\right), r ~ × = c p ~ 3/2 = p ~ 3/2 ( 1 − 12 11 p ~ + 160 219 p ~ 2 + O ( p ~ 3 ) ) ,
( r ~ × ) ′ = 3 p ~ 2 − 22 3 p ~ 3 + 161 9 p ~ 4 + O ( p ~ 5 ) . (\tilde r^{\times})'=3\tilde p^2-\frac{22}{3}\tilde p^3+\frac{161}{9}\tilde p^4+O(\tilde p^5). ( r ~ × ) ′ = 3 p ~ 2 − 3 22 p ~ 3 + 9 161 p ~ 4 + O ( p ~ 5 ) .
Plugging these into (41) yields
ν = p ~ 5 / 2 ( 3 2 − 79 24 p ~ + 3841 576 p ~ 2 + O ( p ~ 3 ) ) . \nu=\tilde p^{5/2}\left(\frac{3}{2}-\frac{79}{24}\tilde p+\frac{3841}{576}\tilde p^2+O(\tilde p^3)\right). ν = p ~ 5/2 ( 2 3 − 24 79 p ~ + 576 3841 p ~ 2 + O ( p ~ 3 ) ) .
To express p ~ \tilde p p ~ as a function of ν \nu ν , we first let p ~ = q 2 / 5 ν 2 / 5 \tilde p=q^{2/5}\nu^{2/5} p ~ = q 2/5 ν 2/5 , so that
ν = q 5 ν ( 3 2 − 79 24 q 2 ν 2 / 5 + 3841 576 q 4 ν 4 / 5 + O ( q 6 ν 6 / 5 ) ) . \nu=q^5\nu\left(\frac{3}{2}-\frac{79}{24}q^2\nu^{2/5}+\frac{3841}{576}q^4\nu^{4/5}+O(q^6\nu^{6/5})\right). ν = q 5 ν ( 2 3 − 24 79 q 2 ν 2/5 + 576 3841 q 4 ν 4/5 + O ( q 6 ν 6/5 ) ) .
Now we divide both sides by ν \nu ν and let z = ν 2 / 5 z=\nu^{2/5} z = ν 2/5 to get
1 = q 5 ( 3 2 − 79 24 z q 2 + 3841 576 z 2 q 4 + O ( z 3 q 6 ) ) . 1=q^5\left(\frac{3}{2}-\frac{79}{24}zq^2+\frac{3841}{576}z^2q^4+O(z^3q^6)\right). 1 = q 5 ( 2 3 − 24 79 z q 2 + 576 3841 z 2 q 4 + O ( z 3 q 6 ) ) .
This equation is now analytic in z z z and q q q and has a simple root at ( q , z ) = ( ( 2 / 3 ) 1 / 5 , 0 ) (q,z)=((2/3)^{1/5},0) ( q , z ) = (( 2/3 ) 1/5 , 0 ) . The implicit function theorem once again allows us to obtain an asymptotic expansion of q q q in terms of z = ν 2 / 5 z=\nu^{2/5} z = ν 2/5 , namely
q = ( 2 3 ) 1 / 5 + q 1 ν 2 / 5 + O ( ν 4 / 5 ) , q=\left(\frac{2}{3}\right)^{1/5}+q_1\nu^{2/5}+O(\nu^{4/5}), q = ( 3 2 ) 1/5 + q 1 ν 2/5 + O ( ν 4/5 ) ,
where q 1 ≈ 0.3441 q_1\approx0.3441 q 1 ≈ 0.3441 .
This finally leads to
p ~ = ν 2 / 5 ( ( 2 3 ) 2 / 5 + K ν 2 / 5 + O ( ν 4 / 5 ) ) , (42) \tilde p=\nu^{2/5}\left(\left(\frac{2}{3}\right)^{2/5}+K\nu^{2/5}+O(\nu^{4/5})\right), \tag*{(42)} p ~ = ν 2/5 ( ( 3 2 ) 2/5 + K ν 2/5 + O ( ν 4/5 ) ) , (42)
where K ≈ 0.6346 K\approx0.6346 K ≈ 0.6346 .
To calculate the asymptotic contraction rate, we substitute p ~ ∗ ∼ ( 2 ν / 3 ) 2 / 5 \tilde p^*\sim(2\nu/3)^{2/5} p ~ ∗ ∼ ( 2 ν /3 ) 2/5 into H 0 H_0 H 0 to obtain
H 0 ( r ~ × , p ~ ∗ , ν ) = r ~ × − 2 ν r ~ × + ν log ( ( p ~ − r ~ × ) 2 ( p ~ + r ~ × ) 2 ) ∼ p ~ 3 − 2 ν p ~ 3 / 2 + 2 ν log ( 1 − p ~ 1 / 2 1 + p ~ 1 / 2 ) ∼ p ~ 3 − 2 ν p ~ 3 / 2 − 4 ν p ~ 1 / 2 ∼ ( 2 ν 3 ) 6 / 5 + O ( ν 8 / 5 ) − 4 ( 2 3 ) 1 / 5 ν 6 / 5 ∼ − C ν 6 / 5 , \begin{aligned}
H_0(\tilde r^{\times},\tilde p^*,\nu)=\tilde r^{\times}-2\nu\sqrt{\tilde r^{\times}}+\nu\log\left(\frac{(\tilde p-\sqrt{\tilde r^{\times}})^2}{(\tilde p+\sqrt{\tilde r^{\times}})^2}\right) \\
&\sim\tilde p^3-2\nu\tilde p^{3/2}+2\nu\log\left(\frac{1-\tilde p^{1/2}}{1+\tilde p^{1/2}}\right) \\
&\sim\tilde p^3-2\nu\tilde p^{3/2}-4\nu\tilde p^{1/2} \\
&\sim\left(\frac{2\nu}{3}\right)^{6/5}+O(\nu^{8/5})-4\left(\frac{2}{3}\right)^{1/5}\nu^{6/5}\sim-C\nu^{6/5},
\end{aligned} H 0 ( r ~ × , p ~ ∗ , ν ) = r ~ × − 2 ν r ~ × + ν log ( ( p ~ + r ~ × ) 2 ( p ~ − r ~ × ) 2 ) ∼ p ~ 3 − 2 ν p ~ 3/2 + 2 ν log ( 1 + p ~ 1/2 1 − p ~ 1/2 ) ∼ p ~ 3 − 2 ν p ~ 3/2 − 4 ν p ~ 1/2 ∼ ( 3 2 ν ) 6/5 + O ( ν 8/5 ) − 4 ( 3 2 ) 1/5 ν 6/5 ∼ − C ν 6/5 ,
where C = 4 ( 2 / 3 ) 1 / 5 − ( 2 / 3 ) 6 / 5 = 10 3 ( 2 / 3 ) 1 / 5 C=4(2/3)^{1/5}-(2/3)^{6/5}=\frac{10}{3}(2/3)^{1/5} C = 4 ( 2/3 ) 1/5 − ( 2/3 ) 6/5 = 3 10 ( 2/3 ) 1/5 . ∎
Numerical experiments
We start by studying the convergence of OSWR using optimized parameters p ∗ = p ∗ ( n ) p^{*}=p^{*}(n) p ∗ = p ∗ ( n ) that minimize the bound M ( r ∗ ( p ) , p , n ) M(r^{*}(p),p,n) M ( r ∗ ( p ) , p , n ) for n = 1 , 2 , 3 , 4 n=1,2,3,4 n = 1 , 2 , 3 , 4 ; we consider T = 0.5 T=0.5 T = 0.5 and two overlap sizes, δ ∈ { 0.14 , 0.5 } \delta\in\{0.14,0.5\} δ ∈ { 0.14 , 0.5 } . In each case, we show two sets of curves: the L 2 L^2 L 2 relative error of OSWR iterations applied to the same discretized problem as in Section 3.3, i.e., the heat equation u t = u y y u_t=u_{yy} u t = u y y with Dirichlet boundary conditions on the finite domain Ω = ( − 4 , 4 ) \Omega=(-4,4) Ω = ( − 4 , 4 ) , with centred difference in space and backward Euler in time with Δ y = 1 / 800 \Delta y=1/800 Δ y = 1/800 and Δ t = 1 / 2400 \Delta t=1/2400 Δ t = 1/2400 . We then plot, next to the convergence curves, the theoretical bounds for the same p ∗ p^{*} p ∗ , see Figure 4.1 . As we mentioned before, the optimal p ∗ p^{*} p ∗ depends on n n n , the target number of iterations, via the parameter ν = n δ 2 / T \nu=n\delta^2/T ν = n δ 2 / T ; indeed, we see from Figure 4.1 that while we have convergence for any choice of parameter p p p , the best parameter does depend on the target number of iterations, as seen by the curves intersecting each other during the first four iterations of OSWR.
Figure 4.1. Relative L 2 L^2 L 2 errors of OSWR iterations and their theoretical bounds for T = 0.5 T = 0.5 T = 0.5 and δ ∈ { 0.14 , 0.5 } \delta\in\{0.14, 0.5\} δ ∈ { 0.14 , 0.5 } , with the optimal parameter p p p chosen for n = 1 , 2 , 3 , 4 n = 1, 2, 3, 4 n = 1 , 2 , 3 , 4 in each case. Left: relative errors, right: superlinear bounds. The second and fourth rows are zooms of the first and third ones to better show the behaviour of the first 5 iterations.
To illustrate the asymptotic behaviour of p ∗ p^{*} p ∗ stated in Theorem 4.2 , we report in Table 5.1 the computed p ∗ p^{*} p ∗ as well as the asymptotic values p ∞ ∗ p_{\infty}^{*} p ∞ ∗ and p 0 ∗ p_{0}^{*} p 0 ∗ for the same value of ν \nu ν . We also report whether the minimum is characterized by equioscillation, rather than by a critical point of H ^ \hat{H} H ^ . We observe for δ = 0.14 \delta=0.14 δ = 0.14 that ∣ p ∗ − p 0 ∗ ∣ p ∗ ≈ C ν 4 / 5 \frac{\lvert p^{*}-p_{0}^{*}\rvert}{p^{*}}\approx C\nu^{4/5} p ∗ ∣ p ∗ − p 0 ∗ ∣ ≈ C ν 4/5 since ν \nu ν is small; this is consistent with the asymptotic expansion (42) . For δ = 0.5 \delta=0.5 δ = 0.5 , we see that we are in a transition zone, with p ∗ p^{*} p ∗ closer to p 0 ∗ p_{0}^{*} p 0 ∗ for n = 1 n=1 n = 1 , but closer to p ∞ ∗ p_{\infty}^{*} p ∞ ∗ for n = 4 n=4 n = 4 . Although (34) suggests that p ∞ ∗ p_{\infty}^{*} p ∞ ∗ should be an O ( ν − 2 ) O(\nu^{-2}) O ( ν − 2 ) approximation of p ∗ p^{*} p ∗ , our ν \nu ν values are too far from the asymptotic regime of ν → ∞ \nu\to\infty ν → ∞ , so we only observe an O ( ν − 1 ) O(\nu^{-1}) O ( ν − 1 ) approximation in this case. To illustrate (34) properly, we show results for δ = 1 \delta=1 δ = 1 , which yields larger values of ν \nu ν ; here, we see that p ∞ ∗ p_{\infty}^{*} p ∞ ∗ is a much better approximation of p ∗ p^{*} p ∗ , and the relative approximation error now scales like O ( ν − 2 ) O(\nu^{-2}) O ( ν − 2 ) , as expected.
Table 5.1. Optimal values of p p p for different numbers of iterations, compared with their asymptotic values p 0 ∗ p_0^* p 0 ∗ and p ∞ ∗ p_\infty^* p ∞ ∗ given by Theorem 4.2 . We also indicate whether the optimum lies on the equioscillation curve and calculate the discrepancy with one of the asymptotic values. The time horizon is T = 0.5 T = 0.5 T = 0.5 for all cases.
In the left panel of Figure 5.1 , we show the level sets of the error measured in the space-time infinity norm when applying OSWR to the error equations for the configuration considered in the top part of Table 5.1 . We vary the number of iterations n n n and the Robin parameter p p p , using the same random initial guess in each case. The error decreases with increasing n n n for all values of p p p . The optimal parameter p p p , computed by solving (33) or (41) , is shown in black and closely follows the valley of minimal error indicated by the level sets. We also plot the asymptotic approximations p 0 ∗ p_{0}^{*} p 0 ∗ and p ∞ ∗ p_{\infty}^{*} p ∞ ∗ in red and blue. These approximations capture the behaviour of the optimal parameter well: p 0 ∗ p_{0}^{*} p 0 ∗ follows the valley for small n n n , while p ∞ ∗ p_{\infty}^{*} p ∞ ∗ will provide a better approximation when n n n becomes large.
Figure 5.1. Left: The best parameter p p p (black curve) calculated by solving (33) or (41) for T = 0.5 T=0.5 T = 0.5 , δ = 0.14 \delta=0.14 δ = 0.14 and various n n n . The asymptotic formulas p 0 ∗ p_0^* p 0 ∗ and p ∞ ∗ p_\infty^* p ∞ ∗ are shown in red and blue, and the iteration errors are shown as contours in the background. Right: comparison between the best parameter p ∗ p^* p ∗ (solid) and the asymptotic formulas (dashed) for various ν \nu ν .
To illustrate the behaviour of the black curve (the valley of minimal error) over a wider range of ν \nu ν , we plot in the right panel of Figure 5.1 the optimal parameter p ~ ∗ \tilde{p}^{*} p ~ ∗ as a function of ν \nu ν ; we further distinguish the portion of the curve given by (33) (critical point of H ^ \hat{H} H ^ , blue solid curve) from the part given by (41) (equioscillation, red solid curve). We see that the two parts of the curve connect very smoothly in the transition zone, so solving either equation will yield a good approximation for p ∗ p^* p ∗ . Moreover, we show in dashed lines the asymptotic formulas p 0 ∗ p_0^* p 0 ∗ and p ∞ ∗ p_\infty^* p ∞ ∗ . We see that p 0 ∗ p_0^* p 0 ∗ is a good approximation of p ∗ p^* p ∗ for ν ≲ 0.5 \nu\lesssim0.5 ν ≲ 0.5 , whereas p ∞ ∗ p_\infty^* p ∞ ∗ approximates p ∗ p^* p ∗ well for ν ≳ 5 \nu\gtrsim5 ν ≳ 5 . Since neither approximation is accurate enough for ν ∈ ( 0.5 , 5 ) \nu\in(0.5,5) ν ∈ ( 0.5 , 5 ) , we recommend solving (33) , namely the quartic equation p ~ 4 = ν 2 ( p ~ + 1 ) ( p ~ + 2 ) \tilde{p}^{4} = \nu^{2}(\tilde{p}+1)(\tilde{p}+2) p ~ 4 = ν 2 ( p ~ + 1 ) ( p ~ + 2 ) , to obtain p ∗ p^* p ∗ directly for such values of ν \nu ν .
Conclusion
We have developed a new analytical technique for capturing the superlinear convergence behaviour of Schwarz waveform relaxation methods. Traditionally, analyzing this regime has relied on technically demanding kernel estimates, which have hindered the optimization of transmission conditions in optimized variants of these algorithms for almost three decades. Yet, the superlinear convergence regime is precisely where these methods achieve their best performance, making the optimization of transmission conditions particularly important.
Our approach exploits the real part of the Laplace parameter in a carefully chosen manner to obtain sharp convergence estimates without resorting to kernel techniques. This enables us, for the first time since the introduction of this class of algorithms three decades ago, to characterize in detail how the optimal transmission parameter depends on the number of iterations, the overlap size, and the time window length. Furthermore, our analysis yields simple asymptotic formulas that can be readily used in practice to select effective transmission parameters.
We have focused on the one-dimensional heat equation to develop and validate our approach. However, the underlying analytical technique is not restricted to this setting and can be extended to more general problems, which we leave for future work.
Appendix A. Proof of Lemma 4.3
First, we need a bound for r ~ − \tilde{r}_{-} r ~ − . By definition, we recall from (23) that
r ~ − = p ~ ( 3 + p ~ − 9 + 4 p ~ ) = p ~ ( 2 p ~ + p ~ 2 3 + p ~ + 9 + 4 p ~ ) ≤ p ~ 2 ( 2 + p ~ 3 + p ~ + 9 + 4 p ~ ) . \tilde{r}_{-}=\tilde{p}\left(3+\tilde{p}-\sqrt{9+4\tilde{p}}\right)=\tilde{p}\left(\frac{2\tilde{p}+\tilde{p}^{2}}{3+\tilde{p}+\sqrt{9+4\tilde{p}}}\right)\leq\tilde{p}^{2}\left(\frac{2+\tilde{p}}{3+\tilde{p}+\sqrt{9+4\tilde{p}}}\right). r ~ − = p ~ ( 3 + p ~ − 9 + 4 p ~ ) = p ~ ( 3 + p ~ + 9 + 4 p ~ 2 p ~ + p ~ 2 ) ≤ p ~ 2 ( 3 + p ~ + 9 + 4 p ~ 2 + p ~ ) .
Since the term inside the parentheses is an increasing function of p ~ \tilde{p} p ~ (verified by differentiation), the fact that p ~ < 5 \tilde{p}<\sqrt{5} p ~ < 5 implies
r ~ − p ~ 2 < 2 + 5 3 + 5 + 9 + 4 5 = 1 5 . (43) \frac{\tilde{r}_{-}}{\tilde{p}^{2}}<\frac{2+\sqrt{5}}{3+\sqrt{5}+\sqrt{9+4\sqrt{5}}}=\frac{1}{\sqrt{5}}. \tag*{(43)} p ~ 2 r ~ − < 3 + 5 + 9 + 4 5 2 + 5 = 5 1 . (43)
Since r ~ × ≤ r ~ − \tilde{r}^{\times}\leq\tilde{r}_{-} r ~ × ≤ r ~ − , r ~ × / p ~ 2 \tilde{r}^{\times}/\tilde{p}^{2} r ~ × / p ~ 2 is bounded by the same constant.
Next, let x ~ + = x ~ + ( r ~ × , p ~ ) \tilde{x}_{+}=\tilde{x}_{+}(\tilde{r}^{\times},\tilde{p}) x ~ + = x ~ + ( r ~ × , p ~ ) and x ~ − = x ~ − ( r ~ × , p ~ ) \tilde{x}_{-}=\tilde{x}_{-}(\tilde{r}^{\times},\tilde{p}) x ~ − = x ~ − ( r ~ × , p ~ ) be the critical points of x ~ ↦ h ( p ~ , r ~ × , x ~ ) \tilde{x}\mapsto h(\tilde{p},\tilde{r}^{\times},\tilde{x}) x ~ ↦ h ( p ~ , r ~ × , x ~ ) defined in (14) . Then H 0 ( r ~ × , p ~ , ν ~ ) = H ^ ( r ~ × , p ~ , ν ~ ) H_{0}(\tilde{r}^{\times},\tilde{p},\tilde{\nu})=\hat{H}(\tilde{r}^{\times},\tilde{p},\tilde{\nu}) H 0 ( r ~ × , p ~ , ν ~ ) = H ^ ( r ~ × , p ~ , ν ~ ) is equivalent to saying that h ( r ~ × , r ~ × , p ~ , ν ~ ) = h ( x ~ + , r ~ × , p ~ , ν ~ ) h(\sqrt{\tilde{r}^{\times}},\tilde{r}^{\times},\tilde{p},\tilde{\nu})=h(\tilde{x}_{+},\tilde{r}^{\times},\tilde{p},\tilde{\nu}) h ( r ~ × , r ~ × , p ~ , ν ~ ) = h ( x ~ + , r ~ × , p ~ , ν ~ ) , so by the mean-value theorem, there is at least one point x ^ ∈ ( r ~ × , x ~ + ) \hat{x}\in(\sqrt{\tilde{r}^{\times}},\tilde{x}_{+}) x ^ ∈ ( r ~ × , x ~ + ) at which ∂ h ∂ x ( x ^ , r ~ × , p ~ , ν ~ ) = 0 \frac{\partial h}{\partial x}(\hat{x},\tilde{r}^{\times},\tilde{p},\tilde{\nu})=0 ∂ x ∂ h ( x ^ , r ~ × , p ~ , ν ~ ) = 0 . But the only positive roots of x ~ ↦ ∂ h ∂ x ( x ~ , r ~ × , p ~ , ν ~ ) \tilde{x}\mapsto\frac{\partial h}{\partial x}(\tilde{x},\tilde{r}^{\times},\tilde{p},\tilde{\nu}) x ~ ↦ ∂ x ∂ h ( x ~ , r ~ × , p ~ , ν ~ ) are x ~ − \tilde{x}_{-} x ~ − and x ~ + \tilde{x}_{+} x ~ + ; therefore, x ^ = x ~ − ∈ ( r ~ × , x ~ + ) \hat{x}=\tilde{x}_{-}\in(\sqrt{\tilde{r}^{\times}},\tilde{x}_{+}) x ^ = x ~ − ∈ ( r ~ × , x ~ + ) , from which it follows that
x ~ − = ( r ~ × + p ~ − A r ~ × − B 2 ) 1 / 2 > r ~ × ⟹ r ~ × + A r ~ × − B < p ~ . \tilde{x}_{-}=\left(\frac{\tilde{r}^{\times}+\tilde{p}-\sqrt{A\tilde{r}^{\times}-B}}{2}\right)^{1/2}>\sqrt{\tilde{r}^{\times}}\quad\Longrightarrow\quad\tilde{r}^{\times}+\sqrt{A\tilde{r}^{\times}-B}<\tilde{p}. x ~ − = ( 2 r ~ × + p ~ − A r ~ × − B ) 1/2 > r ~ × ⟹ r ~ × + A r ~ × − B < p ~ .
Thus, the definition of x ~ + \tilde{x}_{+} x ~ + leads to p ~ / 2 ≤ x ~ + 2 < p ~ \tilde{p}/2\leq\tilde{x}_{+}^{2}<\tilde{p} p ~ /2 ≤ x ~ + 2 < p ~ .
We now rewrite the equation γ ( r ~ × , p ~ ) = 0 \gamma(\tilde{r}^{\times},\tilde{p})=0 γ ( r ~ × , p ~ ) = 0 in (37) as
− 2 r ~ × + log ( ( p ~ − r ~ × ) 2 ( p ~ + r ~ × ) 2 ) = − 2 x ~ + + log ( ( p ~ − x ~ + ) 2 + x ~ + 2 − r ~ × ( p ~ + x ~ + ) 2 + x ~ + 2 − r ~ × ) . (44) -2\sqrt{\tilde{r}^{\times}}+\log\left(\frac{(\tilde{p}-\sqrt{\tilde{r}^{\times}})^{2}}{(\tilde{p}+\sqrt{\tilde{r}^{\times}})^{2}}\right)=-2\tilde{x}_{+}+\log\left(\frac{(\tilde{p}-\tilde{x}_{+})^{2}+\tilde{x}_{+}^{2}-\tilde{r}^{\times}}{(\tilde{p}+\tilde{x}_{+})^{2}+\tilde{x}_{+}^{2}-\tilde{r}^{\times}}\right). \tag*{(44)} − 2 r ~ × + log ( ( p ~ + r ~ × ) 2 ( p ~ − r ~ × ) 2 ) = − 2 x ~ + + log ( ( p ~ + x ~ + ) 2 + x ~ + 2 − r ~ × ( p ~ − x ~ + ) 2 + x ~ + 2 − r ~ × ) . (44)
Notice that both logarithms are negative or zero, since their arguments are less than or equal to 1. This allows us to write for instance
2 x ~ + − 2 r ~ × + log ( ( p ~ − r ~ × ) 2 ( p ~ + r ~ × ) 2 ) < 0 , 2\tilde{x}_+ - 2\sqrt{\tilde{r}^\times} + \log\left(\frac{(\tilde{p}-\sqrt{\tilde{r}^\times})^2}{(\tilde{p}+\sqrt{\tilde{r}^\times})^2}\right) < 0, 2 x ~ + − 2 r ~ × + log ( ( p ~ + r ~ × ) 2 ( p ~ − r ~ × ) 2 ) < 0 ,
or, equivalently
r ~ × − log ( 1 − r ~ × p ~ 1 + r ~ × p ~ ) > x ~ + . \sqrt{\tilde{r}^\times}-\log\left(\frac{1-\frac{\sqrt{\tilde{r}^\times}}{\tilde{p}}}{1+\frac{\sqrt{\tilde{r}^\times}}{\tilde{p}}}\right)>\tilde{x}_+. r ~ × − log 1 + p ~ r ~ × 1 − p ~ r ~ × > x ~ + .
We now recall that the Taylor expansion
− log ( 1 − a 1 + a ) = 2 ( a + a 3 3 + a 5 5 + ⋯ ) -\log\left(\frac{1-a}{1+a}\right)=2\left(a+\frac{a^3}{3}+\frac{a^5}{5}+\cdots\right) − log ( 1 + a 1 − a ) = 2 ( a + 3 a 3 + 5 a 5 + ⋯ )
implies
2 a ≤ − log ( 1 − a 1 + a ) ≤ 2 a 1 − a 2 for 0 ≤ a < 1. (45) 2a\leq-\log\left(\frac{1-a}{1+a}\right)\leq\frac{2a}{1-a^2}\qquad\text{for }0\leq a<1. \tag*{(45)} 2 a ≤ − log ( 1 + a 1 − a ) ≤ 1 − a 2 2 a for 0 ≤ a < 1. (45)
By letting a = r ~ × / p ~ < 1 / 5 1 / 4 < 1 a=\sqrt{\tilde{r}^\times}/\tilde{p}<1/5^{1/4}<1 a = r ~ × / p ~ < 1/ 5 1/4 < 1 , we deduce that
x ~ + < r ~ × ( 1 + 2 / p ~ 1 − 1 / 5 ) = r ~ × p ~ ( p ~ + 5 + 5 2 ) ≤ C r ~ × p ~ , \tilde{x}_+<\sqrt{\tilde{r}^\times}\left(1+\frac{2/\tilde{p}}{1-1/\sqrt{5}}\right)=\frac{\sqrt{\tilde{r}^\times}}{\tilde{p}}\left(\tilde{p}+\frac{5+\sqrt{5}}{2}\right)\leq\frac{C\sqrt{\tilde{r}^\times}}{\tilde{p}}, x ~ + < r ~ × ( 1 + 1 − 1/ 5 2/ p ~ ) = p ~ r ~ × ( p ~ + 2 5 + 5 ) ≤ p ~ C r ~ × ,
where C = 5 + 3 5 2 ≈ 5.8541 C=\frac{5+3\sqrt{5}}{2}\approx5.8541 C = 2 5 + 3 5 ≈ 5.8541 . Squaring both sides and using the fact that x ~ + 2 ≥ p ~ / 2 \tilde{x}_+^2\geq\tilde{p}/2 x ~ + 2 ≥ p ~ /2 leads to
r ~ × ≥ p ~ 2 p ~ 2 C 2 = C 1 p ~ 3 , with C 1 = 1 2 C 2 = 7 − 3 5 20 ≈ 0.0146. \tilde{r}^\times\geq\frac{\tilde{p}}{2}\frac{\tilde{p}^2}{C^2}=C_1\tilde{p}^3,\qquad\text{with }C_1=\frac{1}{2C^2}=\frac{7-3\sqrt{5}}{20}\approx0.0146. r ~ × ≥ 2 p ~ C 2 p ~ 2 = C 1 p ~ 3 , with C 1 = 2 C 2 1 = 20 7 − 3 5 ≈ 0.0146.
Now, to prove the upper bound r ~ × ≤ C 2 p ~ 3 \tilde{r}^\times\leq C_2\tilde{p}^3 r ~ × ≤ C 2 p ~ 3 , we rewrite (A.2) as
r ~ × − log ( 1 − r ~ × p ~ 1 + r ~ × p ~ ) = x ~ + − 1 2 log ( ( p ~ − x ~ + ) 2 + x ~ + 2 − r ~ × ( p ~ + x ~ + ) 2 + x ~ + 2 − r ~ × ) . (46) \sqrt{\tilde{r}^\times}-\log\left(\frac{1-\frac{\sqrt{\tilde{r}^\times}}{\tilde{p}}}{1+\frac{\sqrt{\tilde{r}^\times}}{\tilde{p}}}\right)=\tilde{x}_+-\frac{1}{2}\log\left(\frac{(\tilde{p}-\tilde{x}_+)^2+\tilde{x}_+^2-\tilde{r}^\times}{(\tilde{p}+\tilde{x}_+)^2+\tilde{x}_+^2-\tilde{r}^\times}\right). \tag*{(46)} r ~ × − log 1 + p ~ r ~ × 1 − p ~ r ~ × = x ~ + − 2 1 log ( ( p ~ + x ~ + ) 2 + x ~ + 2 − r ~ × ( p ~ − x ~ + ) 2 + x ~ + 2 − r ~ × ) . (46)
Using (A.3), we see that the left-hand side of (A.4) is bounded below by
r ~ × − log ( 1 − r ~ × p ~ 1 + r ~ × p ~ ) ≥ r ~ × + 2 r ~ × p ~ = r ~ × ( 1 + 2 p ~ ) . (47) \sqrt{\tilde{r}^\times}-\log\left(\frac{1-\frac{\sqrt{\tilde{r}^\times}}{\tilde{p}}}{1+\frac{\sqrt{\tilde{r}^\times}}{\tilde{p}}}\right)\geq\sqrt{\tilde{r}^\times}+\frac{2\sqrt{\tilde{r}^\times}}{\tilde{p}}=\sqrt{\tilde{r}^\times}\left(1+\frac{2}{\tilde{p}}\right). \tag*{(47)} r ~ × − log 1 + p ~ r ~ × 1 − p ~ r ~ × ≥ r ~ × + p ~ 2 r ~ × = r ~ × ( 1 + p ~ 2 ) . (47)
For the right-hand side of (A.4), we observe that
x ~ + − 1 2 log ( ( p ~ − x ~ + ) 2 + x ~ + 2 − r ~ × ( p ~ + x ~ + ) 2 + x ~ + 2 − r ~ × ) = x ~ + − 1 2 log ( 1 − 2 p ~ x ~ + p ~ 2 + 2 x ~ + 2 − r ~ × 1 + 2 p ~ x ~ + p ~ 2 + 2 x ~ + 2 − r ~ × ) . \tilde{x}_+-\frac{1}{2}\log\left(\frac{(\tilde{p}-\tilde{x}_+)^2+\tilde{x}_+^2-\tilde{r}^\times}{(\tilde{p}+\tilde{x}_+)^2+\tilde{x}_+^2-\tilde{r}^\times}\right)=\tilde{x}_+-\frac{1}{2}\log\left(\frac{1-\frac{2\tilde{p}\tilde{x}_+}{\tilde{p}^2+2\tilde{x}_+^2-\tilde{r}^\times}}{1+\frac{2\tilde{p}\tilde{x}_+}{\tilde{p}^2+2\tilde{x}_+^2-\tilde{r}^\times}}\right). x ~ + − 2 1 log ( ( p ~ + x ~ + ) 2 + x ~ + 2 − r ~ × ( p ~ − x ~ + ) 2 + x ~ + 2 − r ~ × ) = x ~ + − 2 1 log ( 1 + p ~ 2 + 2 x ~ + 2 − r ~ × 2 p ~ x ~ + 1 − p ~ 2 + 2 x ~ + 2 − r ~ × 2 p ~ x ~ + ) .
Since r ~ × ≤ p ~ 2 / 5 \tilde{r}^\times\leq\tilde{p}^2/\sqrt{5} r ~ × ≤ p ~ 2 / 5 , we have
p ~ 2 + 2 x ~ + 2 − r ~ × ≥ 0.5528 p ~ 2 + 2 x ~ + 2 ≥ 2 ( 0.5528 ⋅ 2 ) 1 / 2 p ~ x ~ + = 2.1029 p ~ x ~ + , \tilde{p}^2+2\tilde{x}_+^2-\tilde{r}^\times\geq0.5528\tilde{p}^2+2\tilde{x}_+^2\geq2(0.5528\cdot2)^{1/2}\tilde{p}\tilde{x}_+=2.1029\tilde{p}\tilde{x}_+, p ~ 2 + 2 x ~ + 2 − r ~ × ≥ 0.5528 p ~ 2 + 2 x ~ + 2 ≥ 2 ( 0.5528 ⋅ 2 ) 1/2 p ~ x ~ + = 2.1029 p ~ x ~ + ,
where we used the algebraic inequality D 1 2 p ~ 2 + D 2 2 x ~ + 2 ≥ 2 D 1 D 2 p ~ x ~ + D_1^2\widetilde{p}^2+D_2^2\widetilde{x}_+^2\ge2D_1D_2\widetilde{p}\widetilde{x}_+ D 1 2 p 2 + D 2 2 x + 2 ≥ 2 D 1 D 2 p x + . Therefore, we can invoke our Taylor expansion with a = 2 p ~ x ~ + p ~ 2 + 2 x ~ + 2 − r ~ × < 0.9511 < 1 a=\frac{2\widetilde{p}\widetilde{x}_+}{\widetilde{p}^2+2\widetilde{x}_+^2-\widetilde{r}^{\times}}<0.9511<1 a = p 2 + 2 x + 2 − r × 2 p x + < 0.9511 < 1 to deduce that
x ~ + − 1 2 log ( ( p ~ − x ~ + ) 2 + x ~ + 2 − r ~ × ( p ~ + x ~ + ) 2 + x ~ + 2 − r ~ × ) < x ~ + + a 1 − ( 0.9511 ) 2 ≤ x ~ + + 20.9443 p ~ x ~ + 0.5528 p ~ 2 + 2 x ~ + 2 . (48) \widetilde{x}_+-\frac{1}{2}\log\left(\frac{(\widetilde{p}-\widetilde{x}_+)^2+\widetilde{x}_+^2-\widetilde{r}^{\times}}{(\widetilde{p}+\widetilde{x}_+)^2+\widetilde{x}_+^2-\widetilde{r}^{\times}}\right)<\widetilde{x}_++\frac{a}{1-(0.9511)^2}\le\widetilde{x}_++\frac{20.9443\widetilde{p}\widetilde{x}_+}{0.5528\widetilde{p}^2+2\widetilde{x}_+^2}. \tag*{(48)} x + − 2 1 log ( ( p + x + ) 2 + x + 2 − r × ( p − x + ) 2 + x + 2 − r × ) < x + + 1 − ( 0.9511 ) 2 a ≤ x + + 0.5528 p 2 + 2 x + 2 20.9443 p x + . (48)
Inserting the bounds (A.5) and (A.6) into (A.4), we deduce that
r ~ × ( 1 + 2 p ~ ) < x ~ + + 20.9443 p ~ x ~ + 0.5528 p ~ 2 + 2 x ~ + 2 . \sqrt{\widetilde{r}^{\times}}\left(1+\frac{2}{\widetilde{p}}\right)<\widetilde{x}_++\frac{20.9443\widetilde{p}\widetilde{x}_+}{0.5528\widetilde{p}^2+2\widetilde{x}_+^2}. r × ( 1 + p 2 ) < x + + 0.5528 p 2 + 2 x + 2 20.9443 p x + .
Rearranging leads to
r ~ × ≤ x ~ + ( p ~ 2 + p ~ + 20.9443 p ~ 2 ( 2 + p ~ ) ( 0.5528 p ~ 2 + 2 x ~ + 2 ) ) ≤ x ~ + ( p ~ 2 + p ~ + 20.9443 p ~ ( 2 + p ~ ) ( 1 + 0.5528 p ~ ) ) since x ~ + 2 ≥ p ~ / 2 , \begin{aligned}
\sqrt{\widetilde{r}^{\times}}\le\widetilde{x}_+\left(\frac{\widetilde{p}}{2+\widetilde{p}}+\frac{20.9443\widetilde{p}^2}{(2+\widetilde{p})(0.5528\widetilde{p}^2+2\widetilde{x}_+^2)}\right) \\
&\le\widetilde{x}_+\left(\frac{\widetilde{p}}{2+\widetilde{p}}+\frac{20.9443\widetilde{p}}{(2+\widetilde{p})(1+0.5528\widetilde{p})}\right)
\qquad\text{since }\widetilde{x}_+^2\ge\widetilde{p}/2,
\end{aligned} r × ≤ x + ( 2 + p p + ( 2 + p ) ( 0.5528 p 2 + 2 x + 2 ) 20.9443 p 2 ) ≤ x + ( 2 + p p + ( 2 + p ) ( 1 + 0.5528 p ) 20.9443 p ) since x + 2 ≥ p /2 ,
≤ 10.9721 p ~ x ~ + since p ~ ≥ 0 , \le10.9721\widetilde{p}\widetilde{x}_+
\qquad\text{since }\widetilde{p}\ge0, ≤ 10.9721 p x + since p ≥ 0 ,
≤ 10.9721 p ~ 3 / 2 since x ~ + ≤ p ~ . \le10.9721\widetilde{p}^{3/2}
\qquad\text{since }\widetilde{x}_+\le\sqrt{\widetilde{p}}. ≤ 10.9721 p 3/2 since x + ≤ p .
Squaring both sides leads finally to r ~ × ≤ C 2 p ~ 3 \widetilde{r}^{\times}\le C_2\widetilde{p}^3 r × ≤ C 2 p 3 , with C 2 = ( 10.9721 ) 2 ≈ 120.4 C_2=(10.9721)^2\approx120.4 C 2 = ( 10.9721 ) 2 ≈ 120.4 .
References