Derivation of exponential (in a general algebra)

76 Views Asked by At

There is a curious formula relating the derivation of an exponential in some, possibly noncommutative (associative) algebra, namely $$ D(e^K) = \int_0^1 e^{(1-s)K}D(K)e^{sK} ds $$ where $D$ is a derivation (it verifies the usual Leibniz rule) and $K$ is a nilpotent element (there exists some $n\in \mathbb{N}$ such that $K^n = 0$).

The first idea would be to write $$ D(e^K) = D\left( \displaystyle\sum_{l\geq 0} \frac{K^l}{l!} \right) = \displaystyle\sum_{l\geq0} \frac{1}{l!} D(K^l) $$ Now we apply the Leibniz rule: $$ \begin{array}{rcl} D(K^l) &=& D(K)K^{l-1} + K D(K^{l-1}) \\ &=& D(K)K^{l-1} + K D(K) K^{l-2} + K^2D(K^{l-2}) \\ &=& \dots \\ &=& \displaystyle\sum_{j=0}^l K^{j-1} D(K) K^{l-j} \end{array} $$ Then we arrived to $$ D(e^K) = \sum_{l\geq 0} \sum_{j=0}^l \dfrac{1}{l!} K^{j-1} D(K)K^{l-j} $$ The latter sum is over all integer points inside, say, the first octant of the plane. The horizontal direction being the $l$-direction and the vertical one is the $j$-direction.
Using Fubini's theorem, we transform the sum of the vertical slices of the first octant into the sum of its horizontal slices: $$ \sum_{l\geq 0} \sum_{j=0}^l \dfrac{1}{l!} K^{j-1} D(K)K^{l-j} = \sum_{j \geq 0} \sum_{l \geq j} \dfrac{1}{l!} K^{j-1} D(K)K^{l-j} $$ Then set $m= l-j$ to get $$ \sum_{j \geq 0} \sum_{m \geq 0} \dfrac{1}{(m+j)!} K^{j-1} D(K)K^{m} $$ I guess the coefficient $(m+j)!$ should be related with some integration over the $1$-simplex $\{(1-s,s): s\in [0,1]\}$ but I couldn't find yet.

Any hint to finish the proof of the formula? Or maybe a clever approach ?

1

There are 1 best solutions below

7
On

The formula is usually called Duhamel's formula in the math literature and it is also valid for general $K$ (without the requirement of being nilpotent). Here is an elementary derivation.

Consider

\begin{align} \frac{d}{dt}e^{(A+B)t}e^{-At} &=e^{(A+B)t}(A+B)e^{-At}-e^{(A+B)t}Ae^{-At} \\ &=e^{(A+B)t}Be^{-At}, \end{align}

now integrate both sides from zero to one:

$$ e^{A+B}e^{-A}- 1 =\int_{0}^{1}dt \, e^{(A+B)t}Be^{-At} $$ from which we obtain

$$ e^{A+B}-e^{A}=\int_{0}^{1}dt \, e^{(A+B)t}Be^{(1-t)A}. $$

Now set $B=\epsilon D(A)$ and, assuming that $A$ is differentiable, we obtain

\begin{align} De^{A} & =\lim_{\epsilon\to0}\frac{e^{(A+\epsilon DA)}-e^{A }}{\epsilon} \\ &=\int_{0}^{1}dt\, e^{At}D(A)e^{(1-t)A} . \end{align}