I have been exploring methods of generalizing the order of derivatives to a broader range of inputs (such as real numbers, complex, and now matrices). We are very well familiar with integer-order derivatives, and so we shall start there. First, let us observe what happens when we consider successive integer-order derivatives of the eigenfunction $e^{kx}$

$$\frac{d}{dx} e^{kx} = ke^{kx}$$ $$\frac{d^{2}}{dx^{2}} e^{kx} = k^{2}e^{kx}$$ $$\vdots$$ $$\frac{d^{n}}{dx^{n}} e^{kx} = k^{n}e^{kx}$$ $$\text{where } n \in \mathbb{Z^{+}}, k \in \mathbb{C}$$

Although this is not a particularly surprising result, what is interesting is that the order $n$ on the left-hand side corresponds to a power $n$ on the right-hand side of the equation. Therefore, if we wish to consider a non-integer order derivative (for example: 1/2), then this implies that $\frac{d^{1/2}}{dx^{1/2}} e^{kx} = k^{1/2}e^{kx}$. A nice property that still remains intact is $\frac{d^{1/2}}{dx^{1/2}} \left( \frac{d^{1/2}}{dx^{1/2}} e^{kx} \right) = \frac{d}{dx} e^{kx} $. In other words: applying two half-order derivatives in succession is the same as taking a single derivative of integer-order (we just add up the orders).

This idea of non-integer order derivatives is highly explored in the area of fractional calculus. However, what is of interest to me is what happens when we consider a matrix-order derivative. That is, is there some notion of differentiating a function a matrix number of times? Of course, this idea sounds absolutely non-sensical. What could it mean to differentiate a function where the order is not necessarily a number, but is instead a linear operator such as a matrix?

I proceeded with this question as follows: take some diagonalizable matrix $A$ such that $A = PDP^{-1}$ (where $P$ is a matrix consisting of the eigenvectors of $A$ as columns, $D$ is a diagonal matrix consisting of the eigenvalues of $A$, and $P^{-1}$ is simply the inverse of $P$). Borrowing from the prior statement that $\frac{d^{n}}{dx^{n}} e^{kx} = k^{n}e^{kx}$, we wish to evaluate when $n$ (the order of our derivative) is now some diagonalizable matrix $A$

$$ \frac{d^{A}}{dx^{A}} e^{kx} = k^{A}e^{kx} $$

It now appears that in order (no pun intended) to take a matrix-order derivative of $e^{kx}$, we must evaluate the matrix power term $k^{A}$ on the right-hand side. My approach to this is by rewriting this into an exponential form and then leveraging the power series representation of the exponential function

$$ k^{A} = e^{ln(k^{A})} = e^{Aln(k)} = \sum_{n=0}^{\infty} \frac{((Aln(k))^{n}}{n!} = \sum_{n=0}^{\infty} \frac{((ln(k))^{n}}{n!} A^{n} = \text{...} $$

Since our requirement was that our matrix $A$ must be diagonalizable in form of $PDP^{-1}$, we know very well from diagonalization in linear algebra that $A^{n} = PD^{n}P^{-1}$ and therefore

$$ \text{...} = \sum_{n=0}^{\infty} \frac{((ln(k))^{n}}{n!} PD^{n}P^{-1} $$

$$ = P \left( \sum_{n=0}^{\infty} \frac{((ln(k))^{n}}{n!} D^{n} \right) P^{-1} $$

$$ = P \left(\sum_{n=0}^{\infty} \frac{((ln(k))^{n}}{n!} \begin{bmatrix} \lambda_{1}^{n}&0&0\\0&\ddots&0\\0&0&\lambda_{j}^{n} \end{bmatrix} \right) P^{-1} $$

$$ = P \begin{bmatrix} \sum_{n=0}^{\infty} \frac{(\lambda_{1}ln(k))^{n}}{n!}&0&0\\0&\ddots&0\\0&0&\sum_{n=0}^{\infty} \frac{(\lambda_{j}ln(k))^{n}}{n!} \end{bmatrix} P^{-1} $$

$$ = P \begin{bmatrix} e^{ln(k)\lambda_{1}}&0&0\\0&\ddots&0\\0&0& e^{ln(k)\lambda_{j}} \end{bmatrix} P^{-1} $$

$$ = P \begin{bmatrix} k^{\lambda_{1}}&0&0\\0&\ddots&0\\0&0& k^{\lambda_{j}} \end{bmatrix} P^{-1} = k^{A} $$

We can now substitute our derived result for $k^{A}$ back into our original matrix-order derivative equation of interest

$$ \frac{d^{A}}{dx^{A}} e^{kx} = k^{A}e^{kx} $$

$$ = P \begin{bmatrix} k^{\lambda_{1}}&0&0\\0&\ddots&0\\0&0& k^{\lambda_{j}} \end{bmatrix} P^{-1} e^{kx} $$

$$ = P \begin{bmatrix} k^{\lambda_{1}}e^{kx}&0&0\\0&\ddots&0\\0&0& k^{\lambda_{j}}e^{kx} \end{bmatrix} P^{-1} $$

It is here where I observed something I consider to be remarkably elegant! Notice how the entries on the main diagonal exactly match the form of $\frac{d^n}{dx^n} e^{kx} = k^{n}e^{kx}$ and thus

$$ \frac{d^{A}}{dx^{A}} e^{kx} = P \begin{bmatrix} \frac{d^{\lambda_1}}{dx^{\lambda_1}}e^{kx}&0&0\\0&\ddots&0\\0&0& \frac{d^{\lambda_j}}{dx^{\lambda_j}}e^{kx} \end{bmatrix} P^{-1} $$

Notice how the matrix-order derivative of $e^{kx}$ contains a diagonal matrix consisting of eigenvalue-order derivatives of $e^{kx}$ on the main diagonal. It is almost as if the matrix-order derivative was "split apart" into a spectrum of eigenvalue-order derivatives! Another interesting thing to note is how a matrix-order derivative results in another matrix, and therefore our domain now consists of vectors as inputs.

I have also done a similar procedure to evaluate the matrix-order derivatives of $sin(x)$ and $cos(x)$. I will exclude the derivation for the sake of the length of this post. It relies on: $\frac{d^n}{dx^n}sin(x) = sin(x+\frac{\pi}{2}n)$ and $\frac{d^n}{dx^n}cos(x) = cos(x+\frac{\pi}{2}n)$; the double-angle identity; and finally the power series representation of $sin(x)$ and $cos(x)$. The matrix-order derivatives for these trigonometric functions are

$$ \frac{d^{A}}{dx^{A}} sin(x) = P \begin{bmatrix} \frac{d^{\lambda_1}}{dx^{\lambda_1}}sin(x)&0&0\\0&\ddots&0\\0&0& \frac{d^{\lambda_j}}{dx^{\lambda_j}}sin(x) \end{bmatrix} P^{-1} $$

$$ \frac{d^{A}}{dx^{A}} cos(x) = P \begin{bmatrix} \frac{d^{\lambda_1}}{dx^{\lambda_1}}cos(x)&0&0\\0&\ddots&0\\0&0& \frac{d^{\lambda_j}}{dx^{\lambda_j}}cos(x) \end{bmatrix} P^{-1} $$

Much like the case with the matrix-order derivative of $e^{kx}$, the matrix-order derivative of $sin(x)$ and $cos(x)$ consists of the eigenvalue-order derivatives on their respective diagonal matrices.

However, I am still left wondering: what exactly does this mean? What are some possible applications or reasons to consider a matrix-order derivative besides just mere generalization? How does this fit with our notion of derivatives mapping a continuous $C^{n}$ function to a $C^{n-1}$ function? Does this still even make sense, or this just an abuse of notation? Does this provide a new insight into the functionality of the derivative operator?


1 Answer 1


This line of thought has been explored by several authors:

I have not come across a meaningful application of a matrix-order differential operator.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.