Why are all compact Lie groups matrix Lie groups? I've seen it said that this is a consequence of the Peter-Weyl theorem (here), but I don't know how to do this and have been unsuccessful in finding a reference.
 A: The key difference between compact Lie groups and general compact (Hausdorff) topological groups is a Noetherian property:
Lemma. Let $G$ be a compact Lie group and
$$
... <G_2 < G_1< G=G_0
$$
be a chain of proper compact (necessarily Lie) subgroups.
Then this sequence is necessarily finite.
Proof. Dimension in this sequence can drop only finitely many times, hence, WLOG, for each $i$, $G_i< G_{i-1}$ is a codimension 0 subgroup; hence $G_i$ is open in $G_{i-1}$; hence (by compactness) is of finite index in $G_{i-1}$. However, the index $G_n<G$ (if finite) is at most the number of connected components of $G$, which has to be finite by compactness of $G$ (and since $G$ is a manifold!). Thus, the sequence eventually terminates. qed
With this in mind:
Let $\lambda: G\to L^2(G)$ be the left-regular representation; it is a faithful representation. By the P-W theorem, this representation splits as a direct sum of irreducible finite-dimensional factors $V_\alpha, \alpha\in A$. Take $\beta_1\in A$ such that $\lambda_1: G\to GL(V_{\beta_1})$ (the projection of $\lambda$) is a nontrivial representation and let $G_1<G$ be the kernel of  $\lambda_1$. Clearly, $G_1$ is a closed subgroup of $G$. If $G_1=\{1\}$, we are done. Otherwise, there exists $\beta_2\in B$ such that $G\to GL(V_{\beta_2})$ is nontrivial on $G_1$. Let $G_2< G_1$ denote the kernel of
$$
G\to GL(V_{\beta_1}\oplus V_{\beta_2}).
$$
Continue inductively. According to Lemma, this sequence of subgroups of $G$ eventually terminates and we obtain a faithful representation
$$
G\to GL(\oplus_{i=1}^n V_{\beta_i}). 
$$
As for smoothness of this representation, see this link given by freakish.
Remark. I proved the lemma using the theorem  that closed subgroups of a Lie group are Lie subgroups.  In fact, one can avoid appealing to this theorem and to the lemma in full generality: In our setting, each $G_i< G_{i-1}$ is the kernel of a continuous matrix representation of $G_{i-1}$ where we can assume (inductively) that $G_{i-1}$ is a compact Lie group. There is an elementary argument (given by José Carlos Santos in his answer here) that such a representation is necessarily smooth, hence, its kernel is a Lie subgroup.
