Does $\sum\limits_{i=1}^{n} f(\lambda_i) = \text{tr}(f(A))$ hold for any function $f$?

64 Views Asked by At

We know that given a $n$-by-$n$ matrix $A$, and

  • its eigen values $\{\lambda_i\}_1^n$
  • its trace $\text{tr}(A)$.

then the following holds:

$$ \sum\limits_{i=1}^{n} \lambda_i = \text{tr}(A) $$

then I read from this paper (page 2, equations on the upper left, staring with $\log \det B$ ), which implies the following

given any function $T_k(x)$ defined by Chebyshev polynomial, the following also holds:

$$ \sum\limits_{i=1}^{n} T_k(\lambda_i) = \text{tr}(T_k(A)) \quad (1)$$

where for matrix input, $T_k(A)=\{T_k(A_{ij})\}_{ij}$ (equivalent to applying $T_j$ on $A$ element-wisely.)

My question is:

If we replace $T_j$ with any function $f$ in $(1)$, does $(1)$ still hold?