Using variation of constants, we can express the solution
\(\varphi ^{\pm }\) of the semi-nonrelativistic limit system (
3.1) at time
\(t_n+\tau \) as
$$\begin{aligned} \varphi ^{\pm }(t_n+\tau ) = e^{{{\mp }} \textrm{i}\tau {\mathscr {D}}} \varphi ^{\pm }(t_n) -\textrm{i}I_1^{\pm }(\varphi ^+,\varphi ^-,t_n) -\textrm{i}\gamma I_2^{\pm }(\varphi ^+,\varphi ^-,t_n) \end{aligned}$$
(3.4)
with
\(I_j^{\pm } = I_j^{\pm }(\varphi ^+,\varphi ^-,t_n) \) given by
$$\begin{aligned} I_1^{\pm }&= \int _{0}^{\tau } e^{{{\mp }} \textrm{i}(\tau -s) {\mathscr {D}}} \varPi ^{\pm }\left[ W(t_n+s) \, \varphi ^{\pm }(t_n+s) \right] \, ds, \\ I_2^{\pm }&= \int _{0}^{\tau } e^{{{\mp }} \textrm{i}(\tau -s) {\mathscr {D}}} \varPi ^{\pm }\left[ \left( \left|\varphi ^+(t_n+s) \right|^2 + \left|\varphi ^-(t_n+s) \right|^2 \right) \varphi ^{\pm }(t_n+s) \right] ds. \end{aligned}$$
The operators
\(e^{{{\mp }} \textrm{i}(\tau -s) {\mathscr {D}}}\) and
\(\varPi ^{\pm }\) are both bounded in
\((L_2 ({\mathbb {R}}^3))^4\). Thus, in order to obtain a third-order approximation (in
\(\tau \)) to the integrals, we need a second-order approximation (in
s) to the integrands. Under assumption (IV),
\(W(t_n+s)\) can be replaced by the Taylor expansion
$$\begin{aligned} W(t_n+s) = W(t_n) + s \partial _t W(t_n) + {\mathscr {O}} \! \left( s^2\right) . \end{aligned}$$
(3.5)
Since the second time derivative of
\(\varphi ^{\pm }\) is bounded in
\((L_2({\mathbb {R}}^3))^4\) under assumptions (I)–(III), we can also expand
$$\begin{aligned} \varphi ^{\pm }(t_n+s) = \varphi ^{\pm }(t_n) + s \varTheta ^{\pm }(\varphi ^+(t_n),\varphi ^-(t_n),t_n) + {\mathscr {O}} \! \left( s^2\right) , \end{aligned}$$
(3.6)
with
$$\begin{aligned} \varTheta ^{\pm }= \varTheta ^{\pm }(\varphi ^+(t_n),\varphi ^-(t_n),t_n):= \partial _t \varphi ^{\pm }(t_n) \end{aligned}$$
being the first time derivative of
\(\varphi ^{\pm }\) at time
\(t_n\). It is obtained by evaluating the right-hand side of the PDE (
3.1) at time
\(t_n\):
$$\begin{aligned} \varTheta ^{\pm }(\varphi ^+(t_n),\varphi ^-(t_n),t_n)&= {{\mp }} \textrm{i}{\mathscr {D}}\varphi ^{\pm }(t_n) - \textrm{i}\varPi ^{\pm }\left[ W(t_n) \varphi ^{\pm }(t_n) \right] \nonumber \\&\quad - \textrm{i}\gamma \varPi ^{\pm }\left[ \left( \left|\varphi ^+(t_n) \right|^2 + \left|\varphi ^-(t_n) \right|^2 \right) \varphi ^{\pm }(t_n) \right] . \end{aligned}$$
(3.7)
Before we continue by inserting (
3.5) and (
3.6) into (
3.4), let us quickly comment on an alternative approach to construct a second-order approximation to
\(\varphi ^{\pm }(t_n+s)\). Using variation of constants once again, but now over a time interval of length
s, and fixing
\(\varphi ^{\pm }\) as well as
W at time
\(t_n\) inside the integrals yields
$$\begin{aligned} \varphi ^{\pm }(t_n+s)&= e^{{{\mp }} \text {i}s {\mathscr {D}}} \varphi ^{\pm }(t_n) -\text {i}\int _{0}^{s} e^{{{\mp }} \text {i}(s-r) {\mathscr {D}}} \varPi ^{\pm }\left[ W(t_n) \varphi ^{\pm }(t_n) \right] \, dr \nonumber \\ {}&\quad -\text {i}\int _{0}^{s} e^{{{\mp }} \text {i}(s-r) {\mathscr {D}}} \varPi ^{\pm }\left[ \gamma \left( \left| \varphi ^+(t_n) \right| ^2 + \left| \varphi ^-(t_n) \right| ^2 \right) \varphi ^{\pm }(t_n) \right] \, dr \nonumber \\ {}&\quad + {\mathscr {O}} \! \left( s^2\right) . \end{aligned}$$
(3.8)
This approach does only rely on boundedness of the first time derivative of
\(\varphi ^{\pm }\) and thus is, at first glance, feasible under lower regularity assumptions on the potential
W and the initial data. Unfortunately, inserting (
3.8) into (
3.4) leads to integrals which cannot be computed analytically. In order to avoid this problem, we could use the formal approximations
$$\begin{aligned} e^{{{\mp }} \textrm{i}s {\mathscr {D}}} \varphi ^{\pm }(t_n)&= \varphi ^{\pm }(t_n) {{\mp }} \textrm{i}s {\mathscr {D}}\varphi ^{\pm }(t_n) + {\mathscr {O}} \! \left( s^2\right) , \\ \int _{0}^{s} e^{{{\mp }} \textrm{i}(s-r) {\mathscr {D}}} \varPi ^{\pm }\left[ v \right] \, dr&= s \varPi ^{\pm }\left[ v \right] + {\mathscr {O}} \! \left( s^2\right) \end{aligned}$$
which can be rigorously justified for
\( \varphi ^{\pm }\in \left( H^{4}({\mathbb {R}}^3)\right) ^4 \) and
\( v\in \left( H^{2}({\mathbb {R}}^3)\right) ^4 \). Using these approximations in (
3.8), however, yields exactly the same second-order approximation to
\(\varphi ^{\pm }(t_n+s)\) as (
3.6) together with (
3.7).
They can be computed as
$$\begin{aligned} p_1(z) = {\left\{ \begin{array}{ll} \frac{e^z-1}{z} &{} \text {for } z\not =0, \\ 1 &{} \text {for } z=0 \end{array}\right. } \qquad \text {and} \qquad p_2(z) = {\left\{ \begin{array}{ll} \frac{e^z-z-1}{z^2} &{} \text {for } z \ne 0, \\ \frac{1}{2} &{} \text {for } z=0. \end{array}\right. } \end{aligned}$$
When inserting (
3.6) into
\(I_2^{\pm }\), we can additionally drop all
\({\mathscr {O}} \! \left( s^2\right) \)-terms in the integrand that arise due to the nonlinearity. Overall, we obtain
$$\begin{aligned} I_2^{\pm }(\varphi ^+,\varphi ^-,t_n) = {\widehat{I}}_2^{\pm }(\varphi ^+(t_n),\varphi ^-(t_n),t_n) + {\mathscr {O}} \! \left( \tau ^3\right) \end{aligned}$$
where
\({\widehat{I}}_2^{\pm } = {\widehat{I}}_2^{\pm }(\varphi ^+(t_n),\varphi ^-(t_n),t_n)\) is given by
$$\begin{aligned} {\widehat{I}}_2^{\pm }&= \int _0^\tau e^{{{\mp }} \textrm{i}(\tau -s){\mathscr {D}}} \varPi ^{\pm }\left[ \zeta _{\pm }+ s \zeta '_{\pm } \right] \nonumber \, ds\\&= \tau p_1 \left( {{\mp }} \textrm{i}\tau {\mathscr {D}}\right) \varPi ^{\pm }\left[ \zeta _{\pm } \right] + \tau ^2 p_2 \left( {{\mp }} \textrm{i}\tau {\mathscr {D}}\right) \varPi ^{\pm }\left[ \zeta '_{\pm } \right] \end{aligned}$$
with
\(\zeta _{\pm }= \zeta _{\pm }\left( \varphi ^+(t_n),\varphi ^-(t_n),t_n\right) \) and
\(\zeta '_{\pm }= \zeta '_{\pm }\left( \varphi ^+(t_n),\varphi ^-(t_n),t_n\right) \) defined by
$$\begin{aligned} \zeta _{\pm }&= \left( \left|\varphi ^+(t_n) \right|^2 + \left|\varphi ^-(t_n) \right|^2 \right) \varphi ^{\pm }(t_n) \\ \zeta '_{\pm }&= \left( \left|\varphi ^+(t_n) \right|^2 + \left|\varphi ^-(t_n) \right|^2 \right) \varTheta ^{\pm }+ 2 {\text {Re}} \Big ( (\varTheta ^+)^* \varphi ^+(t_n) + (\varTheta ^-)^* \varphi ^-(t_n) \Big ) \varphi ^{\pm }(t_n) \end{aligned}$$
and
\(\varTheta ^{\pm }\) from (
3.7). A third-order approximation to
\(\varphi ^{\pm }(t_n+\tau )\) is obtained by simply replacing the integrals
\(I_1^{\pm }\) and
\(I_2^{\pm }\) in (
3.4) by their approximations
\({\widehat{I}}_1^{\pm }\) and
\({\widehat{I}}_2^{\pm }\):
$$\begin{aligned} \varphi ^{\pm }(t_n+\tau )&= e^{{{\mp }} \textrm{i}\tau {\mathscr {D}}} \varphi ^{\pm }(t_n) -\textrm{i}{\widehat{I}}_1^{\pm }(\varphi ^+(t_n),\varphi ^-(t_n),t_n) \nonumber \\&\quad - \textrm{i}\gamma {\widehat{I}}_2^{\pm }(\varphi ^+(t_n),\varphi ^-(t_n),t_n) + {\mathscr {O}} \! \left( \tau ^3\right) . \end{aligned}$$
(3.10)
This approximation suggests a numerical method with local error of order
\({\mathscr {O}} \! \left( \tau ^3\right) \) which, however, would not be stable. The reason for this instability is the term
\({{\mp }} \textrm{i}{\mathscr {D}}\varphi ^{\pm }(t_n)\) which appears in
\(\varTheta ^{\pm }\), cf. (
3.7), and thus also in
\(\zeta '_{\pm }\). A bound for the norm of
\({\mathscr {D}}\) that is independent of
\(\varepsilon \) can only be established when interpreting
\({\mathscr {D}}\) as mapping from
\(H^{2}\) to
\(L_2 \), cf. (
2.9). Hence, the
\(L_2\)-norm of
\(\varTheta ^{\pm }\) and thus
\(\zeta '_{\pm }\),
\({\widehat{I}}_1^{\pm }\) and
\({\widehat{I}}_2^{\pm }\) can only be bounded using the
\(H^2\)-norm of
\(\varphi ^{\pm }(t_n)\), which would not be sufficient for stability. This is why we replace
\({\mathscr {D}}\) in
\(\varTheta ^{\pm }\) by a filtered version
$$\begin{aligned} {\widetilde{{\mathscr {D}}}}(\tau ) = \frac{\sin (\tau {\mathscr {D}})}{\tau }. \end{aligned}$$
(3.11)
as, e.g., in [
6,
8]. It is not difficult to show that for every
\(\tau >0\),
\({\widetilde{{\mathscr {D}}}}(\tau )\) is a bounded operator from
\(L_2\) to
\(L_2\) with
\(\bigl \Vert {\widetilde{{\mathscr {D}}}}(\tau ) \bigr \Vert \le \frac{1}{\tau }\), and that
$$\begin{aligned} \bigl \Vert ({\mathscr {D}}- {\widetilde{{\mathscr {D}}}}(\tau )) v \bigr \Vert _{L_2 } \le \frac{\tau }{2} \left\Vert v \right\Vert _{H^{4}}. \end{aligned}$$
for all
\(v\in (H^{4}({\mathbb {R}}^3))^4\). Since Theorem
2.3 yields that
\( \varphi ^{\pm }\in C\big ([0,T_2],(H^{4}({\mathbb {R}}^3))^4\big ) \) under assumptions (I) and (II), it follows that replacing
\({\mathscr {D}}\) by
\({\widetilde{{\mathscr {D}}}}(\tau )\) in
\(\varTheta ^{\pm }\) and hence also in
\(\zeta '_{\pm }\) causes an error of
\({\mathscr {O}} \! \left( \tau \right) \). But in
\({\widehat{I}}_1^{\pm }\) and
\({\widehat{I}}_2^{\pm }\), the terms including
\(\varTheta ^{\pm }\) or
\(\zeta '_{\pm }\) are multiplied by a factor
\(\tau ^2\). Thus, substituting
\({\widetilde{{\mathscr {D}}}}(\tau )\) for
\({\mathscr {D}}\) in the right-hand side of (
3.10) causes only an additional error of
\({\mathscr {O}} \! \left( \tau ^3\right) \) and hence does not affect the overall approximation error. All in all, this yields the numerical method
$$\begin{aligned} \varphi ^{\pm }_{n+1}&= \varPhi _{\tau }^{\pm }(\varphi ^+_{n},\varphi ^-_{n},t_n), \qquad n\in {\mathbb {N}}_0, \end{aligned}$$
(3.12)
with the numerical flow
$$\begin{aligned} \varPhi _{\tau }^{\pm }(\varphi ^+_{n},\varphi ^-_{n},t_n)&= e^{{{\mp }} \textrm{i}\tau {\mathscr {D}}} \varphi ^{\pm }_{n} -\textrm{i}{\widetilde{I}}_1^{\pm }(\varphi ^+_{n},\varphi ^-_{n},t_n) -\textrm{i}\gamma {\widetilde{I}}_2^{\pm }(\varphi ^+_{n},\varphi ^-_{n},t_n). \end{aligned}$$
(3.13)
\({\widetilde{I}}_1^{\pm }\) and
\({\widetilde{I}}_2^{\pm }\) correspond to
\({\widehat{I}}_1^{\pm }\) and
\({\widehat{I}}_2^{\pm }\), respectively, but with
\({\mathscr {D}}\) replaced by
\({\widetilde{{\mathscr {D}}}}(\tau )\) in
\(\varTheta ^{\pm }\) and
\(\zeta '_{\pm }\), i.e.
$$\begin{aligned} {\widetilde{I}}_1^{\pm }(\varphi ^+_{n},\varphi ^-_{n},t_n)&= \tau p_1 ( {{\mp }} \textrm{i}\tau {\mathscr {D}}) \, \varPi ^{\pm }\left[ W(t_n) \varphi ^{\pm }_{n} \right] \\&\quad + \tau ^2 p_2( {{\mp }} \textrm{i}\tau {\mathscr {D}}) \, \varPi ^{\pm }\left[ W(t_n) {\widetilde{\varTheta }}^{\pm }+ \partial _t W(t_n) \varphi ^{\pm }_{n} \right] , \\ {\widetilde{I}}_2^{\pm }(\varphi ^+_{n},\varphi ^-_{n},t_n)&= \tau p_1 \left( {{\mp }} \textrm{i}\tau {\mathscr {D}}\right) \varPi ^{\pm }\left[ \zeta _{\pm } \right] + \tau ^2 p_2 \left( {{\mp }} \textrm{i}\tau {\mathscr {D}}\right) \varPi ^{\pm }\left[ \widetilde{\zeta '_{\pm }} \right] , \end{aligned}$$
with
$$\begin{aligned} \zeta _{\pm }= \zeta _{\pm }\left( \varphi ^+_{n},\varphi ^-_{n},t_n\right)&= \left( \left|\varphi ^+_{n} \right|^2 + \left|\varphi ^-_{n} \right|^2 \right) \varphi ^{\pm }_{n}, \\ {\widetilde{\varTheta }}^{\pm }= {\widetilde{\varTheta }}^{\pm }(\varphi ^+_{n},\varphi ^-_{n},t_n)&= {{\mp }} \textrm{i}{\widetilde{{\mathscr {D}}}}(\tau ) \varphi ^{\pm }_{n} - \textrm{i}\varPi ^{\pm }\left[ W(t_n) \varphi ^{\pm }_{n} + \gamma \zeta _{\pm } \right] , \\ \widetilde{\zeta '_{\pm }}= \widetilde{\zeta '_{\pm }}(\varphi ^+_{n},\varphi ^-_{n},t_n)&= \left( \left|\varphi ^+_{n} \right|^2 + \left|\varphi ^-_{n} \right|^2 \right) {\widetilde{\varTheta }}^{\pm }+ 2 {\text {Re}} \Big ( ({\widetilde{\varTheta }}^+)^* \varphi ^+_{n} + ({\widetilde{\varTheta }}^-)^* \varphi ^-_{n} \Big ) \varphi ^{\pm }_{n}. \end{aligned}$$
For an efficient implementation, the two integrals
\({\widetilde{I}}_1^{\pm }\) and
\(\gamma {\widetilde{I}}_2^{\pm }\) can be combined to
$$\begin{aligned} {\widetilde{I}}^{\pm }(\varphi ^+_{n},\varphi ^-_{n},t_n)&= {\widetilde{I}}_1^{\pm }(\varphi ^+_{n},\varphi ^-_{n},t_n) + \gamma {\widetilde{I}}_2^{\pm }(\varphi ^+_{n},\varphi ^-_{n},t_n) \\&= \tau p_1 \left( {{\mp }} \textrm{i}\tau {\mathscr {D}}\right) \varPi ^{\pm }\left[ W(t_n) \varphi ^{\pm }_{n} + \gamma \zeta _{\pm } \right] \\&\quad + \tau ^2 p_2 \left( {{\mp }} \textrm{i}\tau {\mathscr {D}}\right) \varPi ^{\pm }\left[ W(t_n) {\widetilde{\varTheta }}^{\pm }+ \partial _t W(t_n) \varphi ^{\pm }_{n} + \gamma \widetilde{\zeta '_{\pm }} \right] . \end{aligned}$$
Under assumptions (I)–(IV) the local error in
\(L_2\) is bounded by
\(C\tau ^3\) by construction. With well-known techniques, it can be shown that under assumptions (I)–(IV) there are constants
\(\tau _0>0\) and
C such that for all step sizes
\(\tau \in (0,\tau _0]\) the bound
$$\begin{aligned} \left\Vert \varphi ^{\pm }_{n} - \varphi ^{\pm }(t_n) \right\Vert _{L_2 } \le C \tau ^2, \qquad n = 1,2,...,\lfloor T/\tau \rfloor \end{aligned}$$
for the global error holds. We omit the proof, because our focus is not on the benchmark method. The step size restriction
\( \tau \le \tau _0\) is required to obtain uniform boundedness of the numerical approximations in
\(H^2({\mathbb {R}}^3)\), which is required for stability; for the EEMR this issue is discussed in the proof of Theorem
3.7.