The Ideals, the Centre and the Simplicity of $M_n(\mathbb{C})$
Introduction
$M_n(\mathbb{C})$ has exactly two two-sided ideals, $0$ and itself, and exactly one centre, the scalars; those two facts together are what keeps the Peirce decomposition a decomposition of the vector space and never of the algebra.
The article proves the absence of proper two-sided ideals, computes the centre, describes the left and the right ideals as the subspaces of the domain and of the codomain, and draws the consequences of simplicity: every automorphism is inner, every derivation is inner and the derivations form a space of dimension $n^{2}-1$, and there is exactly one irreducible module. It then contrasts the algebra with the triangular algebra $UT_n$, which has a radical and several ideals, and it closes at $n=2$. It assumes The Element: Rank, Trace, Determinant, Characteristic Polynomial for the rank and The Minimal Polynomial and the Jordan Form for the classification. It defers the Peirce corners to The Peirce Decomposition and the Corner Recursion of $M_n(\mathbb{C})$ and the radical of the triangular case to The Antisymmetric Part inside $UT_n$: the Radical as the Commutator.
The Two-Sided Ideals
A nonzero two-sided ideal is the whole algebra. Let $I\ne0$ be a two-sided ideal and choose $A\in I$ with $A_{kl}\ne0$. For all $i,j$,
$$ E_{ik}AE_{lj}=A_{kl}E_{ij}\in I . $$
The scalar $A_{kl}$ is nonzero, so $E_{ij}\in I$ for every pair, and the invertibles span $M_n(\mathbb{C})$. Hence $I=M_n(\mathbb{C})$.
The same argument shows that any nonzero element generates the algebra. The two-sided ideal generated by a single nonzero $A$ is the whole algebra, so a nonzero element is never contained in a proper two-sided ideal. Equivalently, the only element whose conjugates and products stay small is zero.
The algebra is a product of matrix algebras in only the trivial way. The Wedderburn decomposition of $M_n(\mathbb{C})$ as a product of simple factors has one factor, the algebra itself. This is why the corner decomposition of The Peirce Decomposition and the Corner Recursion of $M_n(\mathbb{C})$ is a block-diagonal subalgebra and not a sum of ideals: a sum of ideals would require the projections onto the summands to be central, and the only central elements are the scalars.
The Centre
The centre is $\mathbb{C}\mathrm{I}$, of dimension one. If $A$ commutes with every $E_{kl}$, then comparing the entries of $AE_{kl}=E_{kl}A$ gives $A_{ik}=0$ for $i\ne k$ and $A_{ii}=A_{kk}$ for all $i,k$. So $A=\lambda\mathrm{I}$, and conversely every scalar matrix is central. A direct computation at $n=3$ confirms the dimension: the space of matrices commuting with all nine units has dimension $1$.
A central simple algebra is simple with centre the base field, and $M_n(\mathbb{C})$ is the split model of that class. The two properties together say that the algebra cannot be decomposed by ideals and that its scalars are the only elements that behave like scalars. The definition is not idle: the biquaternion algebra $\mathbb{B}$ is central simple over $\mathbb{C}$ too, and the matrix algebra is its split representative of degree two.
The centre is not the fixed set of an involution. Every element splits as $\tfrac12(A+A^{\mathsf T})+\tfrac12(A-A^{\mathsf T})$, and the two halves are large. The centre is much smaller, and the difference between a decomposition by an involution and a decomposition by central idempotents is the difference between two halves of the space and two ideals of the algebra.
The Left and the Right Ideals
A right ideal is the set of all maps into a fixed subspace. For a subspace $W\subseteq\mathbb{C}^{n}$ put
$$ R_W=\{A:\operatorname{im}A\subseteq W\}. $$
Then $\operatorname{im}(AB)\subseteq\operatorname{im}A\subseteq W$, so $R_W$ is closed under right multiplication, and every right ideal is of this form: given a right ideal $R$, take $W=\sum_{A\in R}\operatorname{im}A$, and then $R\subseteq R_W$ while $A\in R_W$ forces $A=AP_0$ for a projection $P_0$ onto $W$, hence $A\in R$. Since a matrix in $R_W$ has each of its $n$ columns in $W$,
$$ \dim R_W=n\dim W . $$
A left ideal is the set of all maps that kill a fixed subspace. For $U\subseteq\mathbb{C}^{n}$ put
$$ L_U=\{A:\ker A\supseteq U\}. $$
Then $\ker(BA)\supseteq\ker A$, so $L_U$ is closed under left multiplication, and every left ideal is of this form. Its dimension is
$$ \dim L_U=n(n-\dim U), $$
the count of the free entries once the first $\dim U$ columns are set to zero in a basis adapted to $U$. At $n=3$ a line $U$ gives $6$ and a plane gives $3$; at $n=4$ the values are $12$ and $8$.
The minimal ideals have dimension $n$, and they are the blocks of the Peirce array. A minimal right ideal is $R_W$ with $\dim W=1$, namely $P_kM_n$: the set of matrices whose only nonzero rows lie in the $k$-th one, of dimension $n$. A minimal left ideal is $L_U$ with $\dim U=n-1$, namely $M_nP_k$, which decomposes as
$$ M_nP_k=\bigoplus_{l=1}^{n}A_{lk}, $$
a sum of $n$ Peirce lines, one per value of the first index. So the minimal left ideals are the columns of the array $A_{lk}=P_lM_nP_k$ and the minimal right ideals are its rows, and both are of dimension $n$. Their number is $n$ in each case.
Both families are Grassmannians. The right ideals of dimension $nk$ are parametrised by $W\in\mathrm{Gr}(k,n)$, so they form a Grassmannian of dimension $k(n-k)$; the left ideals of dimension $nk$ are parametrised by $U$ of dimension $n-k$ and form the same Grassmannian. The two families are exchanged by transposition, and the space of a family is the space of a flag.
Simplicity and Its Consequences
Every automorphism is inner. An automorphism $\varphi$ of $M_n(\mathbb{C})$ sends the minimal left ideal $L_V$ to another minimal left ideal, hence sends a line of the domain to a line and induces a projective transformation of $\mathbb{P}^{n-1}$. Lifting that transformation to a linear $S$ gives $\varphi(A)=SAS^{-1}$, and $S$ is unique up to a central scalar. This is the Skolem–Noether theorem in the matrix case, and it is the reason the Peirce gauges of The Peirce Basis and the Corners exhaust the freedom of a frame.
Every derivation is inner, and the derivations form a space of dimension $n^{2}-1$. A derivation is a linear $D$ with $D(AB)=D(A)B+AD(B)$. Every $D_A(X)=[A,X]$ is a derivation, and conversely every derivation is of that form: the map $A\mapsto D_A$ has kernel the centre, so
$$ \dim\operatorname{Der}M_n(\mathbb{C})=n^{2}-1 . $$
The dimension was verified by solving the derivation equations at $n=2,3,4$, where the values are $3,8,15$. The inner derivations are the Lie algebra $\mathfrak{sl}_n$ acting on the algebra by the commutator.
There is exactly one irreducible module. A module is a vector space $V$ with an action of the algebra; taking a cyclic vector $v$ gives a surjection $M_n(\mathbb{C})\to V$, $A\mapsto Av$, whose kernel is a left ideal. By simplicity of the module the kernel is maximal, hence $V\cong\mathbb{C}^{n}$, and every irreducible module is the standard one. The algebra is therefore a full matrix algebra on one representation, which is the structural reason that its left ideals are indexed by subspaces of a single space.
The Contrast with $UT_n$
The triangular algebra is not simple, and its radical carries the difference. Let $UT_n$ be the algebra of upper triangular matrices, of complex dimension $n(n+1)/2$, with radical $N$ the strictly upper triangular part, of dimension $n(n-1)/2$. Then $N$ is a nonzero two-sided ideal, nilpotent of index $n$, and
$$ UT_n/N\cong\mathbb{C}^{n}, $$
the diagonal, which is commutative. So $UT_n$ is solvable and has a nontrivial radical, while $M_n(\mathbb{C})$ has radical zero.
$UT_n$ has several two-sided ideals and $M_n(\mathbb{C})$ has two. Inside $N$ the ideals are the sums of root spaces over the order ideals of the poset of pairs $(i,j)$ with $i $$
0,\quad\mathbb{C}E_{13},\quad\mathbb{C}E_{12}\oplus\mathbb{C}E_{13},\quad\mathbb{C}E_{23}\oplus\mathbb{C}E_{13},\quad N,
$$ five in all, each verified as an ideal by closure under multiplication by the six basis elements on both sides. The proper ideals of $UT_3$ are these five and the algebra is the sixth; at $n=2$ the count is three, $0$, $\mathbb{C}E_{12}$ and $UT_2$. The two algebras share the carrier space of the symmetric matrices and not the product. Both $\operatorname{Sym}_n$ and $UT_n$ have dimension $n(n+1)/2$ over $\mathbb{C}$, and the split $M_n(\mathbb{C})=\operatorname{Sym}_n\oplus\operatorname{Skew}_n$ reads the first as half of the algebra. The triangular algebra is a different multiplication on the same space, and the difference is exactly the presence of a radical. This is the contrast developed in $UT_n$ as a Linear Model of $\operatorname{Sym}_n$: Same Space, Wrong Product. At $n=2$ the ideal structure is the smallest nontrivial one. The algebra has two two-sided ideals, the centre is the line $\mathbb{C}\mathrm{I}$, the minimal left ideals are the two columns of dimension $2$, and the derivations form a space of dimension $3$. The count $3$ is the dimension of $\mathfrak{sl}_2$, the Lie algebra that acts by commutator, and $3=n^{2}-1$ is the first value of the general dimension. The two columns are the two halves of the Peirce decomposition. With the idempotents $\tilde\Pi_\pm$ of The Peirce Decomposition of the Biquaternion Algebra, the left ideals are $M_2\tilde\Pi_+$ and $M_2\tilde\Pi_-$, each of dimension $2$ and each the sum of two Peirce lines. Their intersection is $0$ and their sum is the algebra, but the sum is not direct as algebras: the mixed lines belong to neither, so the decomposition into columns is a decomposition of the space only. $M_n(\mathbb{C})$ has no proper nonzero two-sided ideal, and its centre is $\mathbb{C}\mathrm{I}$; it is central simple. The right ideals are the sets $R_W$ of matrices with image in a subspace $W$, of dimension $n\dim W$; the left ideals are the sets $L_U$ of matrices that kill a subspace $U$, of dimension $n(n-\dim U)$. The minimal ones have dimension $n$ and are the rows and the columns of the Peirce array. Simplicity forces every automorphism to be inner, every derivation to be inner with $\dim\operatorname{Der}=n^{2}-1$, and the irreducible module to be unique. The triangular algebra $UT_n$ fails all three tests: it has the radical $N$ with $UT_n/N\cong\mathbb{C}^{n}$, its two-sided ideals are the order ideals of the root poset, five inside $N$ at $n=3$, and its dimension coincides with that of $\operatorname{Sym}_n$ while its product does not.The Case $n=2$
Summary
Summary of Notation
Symbol
Meaning
$I$
a two-sided ideal of the algebra
$E_{ij}$
the matrix unit
$R_W=\{A:\operatorname{im}A\subseteq W\}$
the right ideal attached to a subspace $W$
$L_U=\{A:\ker A\supseteq U\}$
the left ideal attached to a subspace $U$
$\operatorname{Gr}(k,n)$
the Grassmannian of $k$-planes, of dimension $k(n-k)$
$\operatorname{Der}M_n(\mathbb{C})$
the space of derivations, of dimension $n^{2}-1$
$UT_n$
the upper triangular algebra, of dimension $n(n+1)/2$
$N$
the strictly upper triangular radical, of dimension $n(n-1)/2$
Further Reading