The Elementary Symmetric Functions and $\det(\mathrm{I}+tA)$
Introduction
One polynomial carries the whole spectrum of $A$: the generating polynomial $\det(\mathrm{I}+tA)$ has the elementary symmetric functions of $A$ as its coefficients, and those coefficients are the complete set of conjugation-invariant polynomial functions on the algebra.
The article defines $e_k(A)$ as the sum of the principal $k\times k$ minors, proves that $\det(\mathrm{I}+tA)=\sum_k e_k(A)t^{k}$, derives Newton's identities that recover the $e_k$ from the power sums $\operatorname{tr}(A^{m})$, and shows that the $e_k$ generate the ring of invariant polynomials with no relation among them. The last two sections read the derivative at the origin, which is the trace and which forces $\det e^{A}=e^{\operatorname{tr}A}$, and the case $n=2$. It assumes The Element: Rank, Trace, Determinant, Characteristic Polynomial for the four invariants and for the characteristic polynomial. It defers the structure of the invariant ring as a symmetric algebra to The Trace Form and Its Inertia on the Two Halves, and the exponential itself to The Exponential of a Matrix and Its Image.
The Elementary Symmetric Functions
The $k$-th elementary symmetric function of a matrix is the sum of its principal $k\times k$ minors. For $A\in M_n(\mathbb{C})$ and $0\le k\le n$, write $K$ for a subset of $\{1,\dots,n\}$ of size $k$ and $A_K$ for the submatrix on the rows and columns of $K$. Then
$$ e_k(A)=\sum_{|K|=k}\det A_K,\qquad e_0=1,\qquad e_1=\operatorname{tr}A,\qquad e_n=\det A . $$
There are $\binom{n}{k}$ summands, so $e_k$ is a homogeneous polynomial of degree $k$ in the entries. The definition uses principal submatrices only; the non-principal minors enter the rank but not the spectrum.
The $e_k$ are the elementary symmetric functions of the eigenvalues. If $A$ is triangular with diagonal $\lambda_1,\dots,\lambda_n$, then $A_K$ is triangular with diagonal the corresponding $\lambda$'s, so $\det A_K$ is the product of those $\lambda$'s and $e_k$ is the sum of all products of $k$ distinct eigenvalues. Both sides are polynomial in the entries and the triangular matrices are dense, so the identity holds for all $A$.
At $n=2$ and $n=3$ the list is short.
$$ e_1=\operatorname{tr}A,\qquad e_2=\tfrac12\big((\operatorname{tr}A)^{2}-\operatorname{tr}(A^{2})\big),\qquad e_3=\det A $$
at $n=3$, with $e_2$ also equal to the sum of the three principal $2\times2$ minors. The identity $e_2=\tfrac12((\operatorname{tr}A)^{2}-\operatorname{tr}A^{2})$ is the case $k=2$ of Newton's identities and holds at every $n$.
The Generating Polynomial
The polynomial $\det(\mathrm{I}+tA)$ has the $e_k$ as coefficients. For every $A$ and every scalar $t$,
$$ \det(\mathrm{I}+tA)=\sum_{k=0}^{n}e_k(A)\,t^{k}. $$
Proof. Expand the determinant by the Leibniz rule over the $n$ columns. The column $j$ of $\mathrm{I}+tA$ is $e_j+tA_{\cdot j}$, and a term of the expansion chooses, for each $j$, either the unit column $e_j$ or the column $tA_{\cdot j}$. The columns where $tA$ is chosen form a set $K$; the remaining columns are unit vectors, so the term is nonzero only if those unit vectors occupy the rows of the complement of $K$, and the value is then $t^{|K|}\det A_K$. Summing over $K$ gives the formula.
The same polynomial read at $t=-1/t'$ gives the characteristic polynomial. Since $\chi_A(t)=t^{n}\det(\mathrm{I}-A/t)$,
$$ \chi_A(t)=\sum_{k=0}^{n}(-1)^{k}e_k(A)\,t^{n-k}, $$
which is the coefficient statement of The Element: Rank, Trace, Determinant, Characteristic Polynomial, now with the $e_k$ named and computable from the minors.
The polynomial is multiplicative across a product in either order. For square $A,B$ of the same size,
$$ \det(\mathrm{I}+tAB)=\det(\mathrm{I}+tBA). $$
Proof. For invertible $A$, $\mathrm{I}+tAB=A(\mathrm{I}+tBA)A^{-1}$. Both sides are polynomial in the entries of $A$, and the invertible matrices are dense, so the identity extends to all $A$. The consequence is that $AB$ and $BA$ have the same characteristic polynomial and hence the same $e_k$ for every $k$, although they need not be conjugate.
A table of the first cases.
| $n$ | $\det(\mathrm{I}+tA)$ |
|---|---|
| $2$ | $1+t\operatorname{tr}A+t^{2}\det A$ |
| $3$ | $1+t\operatorname{tr}A+t^{2}e_2+t^{3}\det A$ |
| $4$ | $1+t\operatorname{tr}A+t^{2}e_2+t^{3}e_3+t^{4}\det A$ |
The middle coefficients at $n\ge4$ are new data, and they are the reason that the trace and the determinant do not determine the spectrum beyond $n=2$.
Newton's Identities and the Power Sums
The power sums determine the $e_k$, and Newton's identities give the recursion. Put $p_i=\operatorname{tr}(A^{i})=\sum_k\lambda_k^{i}$. Then for $1\le k\le n$,
$$ k\,e_k=\sum_{i=1}^{k}(-1)^{i-1}e_{k-i}\,p_i . $$
For $k=1$ this reads $e_1=p_1$, for $k=2$ it reads $2e_2=e_1p_1-p_2$, and for $k=3$ it reads $3e_3=e_2p_1-e_1p_2+p_3$. The identity is the statement that the generating functions are related by
$$ \sum_{k\ge0}e_kt^{k}=\exp\Big(\sum_{i\ge1}(-1)^{i-1}\frac{p_i}{i}t^{i}\Big), $$
which is the expansion of $\prod_k(1+t\lambda_k)$ in terms of $\log(1+t\lambda_k)$.
The recursion shows that the $n$ power sums $p_1,\dots,p_n$ carry the same information as the $n$ coefficients. Each $e_k$ is a polynomial in $p_1,\dots,p_k$ with rational coefficients, and conversely $p_k$ is a polynomial in $e_1,\dots,e_k$. The two lists are two bases of the same ring, and neither is smaller.
The trace is the derivative of the determinant along the identity, and the other $e_k$ are the higher derivatives. Expanding at $t=0$,
$$ \det(\mathrm{I}+tA)=1+t\operatorname{tr}A+\tfrac12t^{2}\big((\operatorname{tr}A)^{2}-\operatorname{tr}A^{2}\big)+\cdots, $$
so the first derivative is $\operatorname{tr}$ and the second is $2e_2$. This reading is the one used for the exponential in §The Derivative at the Origin.
The Invariant Ring
A polynomial function on the algebra is conjugation invariant exactly when it is a polynomial in the $e_k$. Let $\mathbb{C}[M_n]$ be the polynomial ring in the $n^{2}$ entries. Then
$$ \mathbb{C}[M_n]^{GL_n}=\mathbb{C}[e_1,\dots,e_n], $$
the subring of invariants is a polynomial ring in $n$ generators, of degrees $1,2,\dots,n$.
The $n$ generators are algebraically independent, and their degrees sum to $n(n+1)/2$. Independence holds because the Jacobian of $(e_1,\dots,e_n)$ with respect to the diagonal entries is the Vandermonde determinant of $\lambda_1,\dots,\lambda_n$, which is nonzero on the diagonal matrices with distinct eigenvalues; a relation among the $e_k$ would therefore vanish on an open set of diagonal matrices and be zero. The degree sum is
$$ 1+2+\cdots+n=\frac{n(n+1)}{2}=3,6,10,15,21\quad\text{for }n=2,3,4,5,6 . $$
This number is also the dimension of the space of symmetric matrices, the carrier of the Jordan algebra of The Jordan Algebra $\operatorname{Sym}_n$, and the coincidence is not accidental: on a symmetric matrix the $e_k$ are read from symmetric data.
The rank is not an invariant of this ring. The rank is not a polynomial function: it is the largest $r$ with a nonzero minor, and the vanishing of all $(r+1)\times(r+1)$ minors is a closed condition, not an equation. So the polynomial invariants see the spectrum and not the size of the image, which is the algebraic form of the failure recorded in The Element: Rank, Trace, Determinant, Characteristic Polynomial.
The Derivative at the Origin
The logarithmic derivative of the determinant is the trace. For a differentiable family $A(t)$ with $A(0)=\mathrm{I}$,
$$ \frac{d}{dt}\Big|_{t=0}\det A(t)=\operatorname{tr}\dot A(0), $$
which is Jacobi's formula $\frac{d}{dt}\det A=\det A\cdot\operatorname{tr}(A^{-1}\dot A)$ evaluated at the identity. Applied to $A(t)=\mathrm{I}+tA$ it recovers the coefficient $e_1=\operatorname{tr}A$.
The formula forces $\det e^{A}=e^{\operatorname{tr}A}$. Let $f(t)=\det e^{tA}$. Jacobi's formula with $A(t)=e^{tA}$ and $\dot A=A(t)A$ gives
$$ f'(t)=f(t)\operatorname{tr}A,\qquad f(0)=1, $$
so $f(t)=e^{t\operatorname{tr}A}$ and $t=1$ is the statement. This is the only place where the determinant and the trace are tied by a formula rather than by a coefficient, and it is used in The Exponential of a Matrix and Its Image to place the exponential inside $SL_n(\mathbb{C})$.
The same computation at the level of the $e_k$ gives the derivative of each coefficient. Differentiating $\det(\mathrm{I}+tA)=\sum_ke_kt^{k}$ in $t$ and setting $t=0$ gives $e_1$, and differentiating $k$ times gives $k!e_k$. So the whole polynomial is the Taylor expansion of the determinant along the line $\mathrm{I}+tA$, and the $e_k$ are its Taylor coefficients divided by $k!$.
The Case $n=2$
At $n=2$ the polynomial has three terms and the pair $(e_1,e_2)$ is the whole spectrum. For $A\in M_2(\mathbb{C})$,
$$ \det(\mathrm{I}+tA)=1+t\operatorname{tr}A+t^{2}\det A,\qquad e_2=\det A=\tfrac12\big((\operatorname{tr}A)^{2}-\operatorname{tr}(A^{2})\big). $$
Newton's identity at $k=2$ is the second formula, and it is the reason that a $2\times2$ matrix satisfies its own characteristic polynomial with no intermediate coefficient.
The invariant ring at $n=2$ has two generators of degrees $1$ and $2$, summing to $3$. The conjugation orbits in $M_2(\mathbb{C})$ are therefore described by two numbers, $\operatorname{tr}$ and $\det$, with the discriminant
$$ \Delta=(\operatorname{tr}A)^{2}-4\det A $$
separating the two distinct eigenvalues from the double one. The orbits are the strata of The Minimal Polynomial and the Jordan Form: two distinct eigenvalues, one eigenvalue with a diagonalisable matrix, and one eigenvalue with a nontrivial Jordan block, the last being the single orbit of $A\ne\lambda\mathrm{I}$ with $\Delta=0$.
The biquaternion norm is not the determinant. At $n=2$ the algebra is the biquaternion algebra $\mathbb{B}$, and the article Introduction to the Algebra of Biquaternions attaches to each element a real norm $N$. That norm is real-valued and indefinite, while $\det$ is complex-valued and multiplicative; the two agree in neither inertia nor domain. The determinant of the matrix of $q$ is the complex number $\det q$, and the statement $N(q)=\det q$ is false for a biquaternion. What survives is the trace: $\operatorname{tr}$ is complex-linear on $\mathbb{B}$, and its real part is the real part of the biquaternion trace.
Summary
The elementary symmetric functions $e_k(A)$, the sum of the principal $k\times k$ minors, are the coefficients of $\det(\mathrm{I}+tA)=\sum_ke_k(A)t^{k}$, and they are the elementary symmetric functions of the eigenvalues. Newton's identities recover them from the power sums $\operatorname{tr}(A^{m})$ by a recursion with rational coefficients, so the two lists carry the same information. The $e_k$ are algebraically independent and generate the ring of conjugation-invariant polynomials, with degrees $1,\dots,n$ summing to $n(n+1)/2$. The derivative of $\det(\mathrm{I}+tA)$ at the origin is $\operatorname{tr}A$, and Jacobi's formula turns that into $\det e^{A}=e^{\operatorname{tr}A}$. At $n=2$ the polynomial is $1+t\operatorname{tr}A+t^{2}\det A$, and the biquaternion norm $N$ is not the determinant.
Summary of Notation
| Symbol | Meaning |
|---|---|
| $e_k(A)$ | the sum of the principal $k\times k$ minors of $A$ |
| $A_K$ | the principal submatrix on the index set $K$ |
| $\det(\mathrm{I}+tA)$ | the generating polynomial of the $e_k$ |
| $p_i=\operatorname{tr}(A^{i})$ | the $i$-th power sum of the eigenvalues |
| $\mathbb{C}[M_n]^{GL_n}$ | the ring of conjugation-invariant polynomials |
| $\Delta=(\operatorname{tr}A)^{2}-4\det A$ | the discriminant at $n=2$ |
| $N$ | the indefinite biquaternion norm, not the determinant |
Further Reading
- Roger A. Horn and Charles R. Johnson, Matrix Analysis (Cambridge, 2013), for the expansion of $\det(\mathrm{I}+tA)$ and Newton's identities.
- Israel M. Gelfand, Lectures on Linear Algebra (Dover, 1989), for the elementary symmetric functions of the spectrum and their generating polynomial.
- David Cox, John Little and Donal O'Shea, Ideals, Varieties, and Algorithms (Springer, 2015), for the invariant ring of the conjugation action and the algebraic independence of the generators.
- The Element: Rank, Trace, Determinant, Characteristic Polynomial, for the four invariants and the coefficient form of $\chi_A$.
- The Frame, the Simple Nilpotents and the Gauge of $M_n(\mathbb{C})$, for the degree count $n(n+1)/2$ against the frame count.
- The Biquaternion Algebra, for the indefinite norm $N$ that is not the determinant.