In numerical analysis it is common to approximate a solution to a PDE

$$u'(t) = (A+B) u(t), \quad u(0)=u_0$$

which is just given by $e^{(A+B)t}u_0$ by the splitting $e^{B/2 t} e^{At} e^{B/2t}u_0.$ Here, $A,B$ can be assumed to be self-adjoint matrices.

We shall furthermore assume that $A+B$ generates a contraction semigroup, i.e. $\Vert e^{(A+B)t} \Vert \le 1,$i.e. all eigenvalues of $A+B$ are negative.

Since structure preserving schemes are important, I am looking for a criterion such that $$\Vert e^{B/2 t} e^{At} e^{B/2t}\Vert \le C \text{ for all }t \in [0,\infty).$$

There are obviously some conditions like $A,B$ individually having only non-positive eigenvalues etc. that would do the job, but I am looking for a more intelligent criterion that takes into account the fact that this is really supposed to be an approximation of the exponential of $A+B$.

In numerical analysis it is common to approximate a solution to a PDE

$$u'(t) = (A+B) u(t), \quad u(0)=u_0$$

which is just given by $e^{t(A+B)}u_0$ by the splitting $e^{tB/2} e^{tA} e^{tB/2}u_0.$ Here, $A,B$ can be assumed to be self-adjoint matrices.

We shall furthermore assume that $A+B$ generates a contraction semigroup, i.e. $\Vert e^{t(A+B)} \Vert \le 1,$ i.e. all eigenvalues of $A+B$ are negative.

Since structure preserving schemes are important, I am looking for a criterion such that $$\Vert e^{tB/2} e^{tA} e^{tB/2}\Vert \le C \text{ for all }t \in [0,\infty).$$

There are obviously some conditions like $A,B$ individually having only non-positive eigenvalues etc..., that would do the job, but I am looking for a more intelligent criterion that takes into account the fact that this is really supposed to be an approximation of the exponential of $A+B$.