In 'Functional Analysis'by Rudin, Lemma $4.21$

Let $M$ be a closed subspace of a topological vector space $X$.

(a) If $X$ is locally convex and $dim(M) < \infty$, then $M$ is complemented in $X$.

(b) If $dim(X/M) < \infty$, then $M$ is complemented in $X$.

(The following proof is copied from the book)

Proof: (a) Let $\{ e_1,...,e_n \}$ be a basis for $M$. Every $x \in M$ has then a unique representation $$x = \alpha_1(x)e_1 + ...+ \alpha_n(x)e_n.$$

Each $\alpha_i$ is continuous linear functional on $M$ which extends to a member of $X^*$, by the Hahn-Banach Theorem. Let $N$ be the intersection of the null spaces of these extensions. Then $X = M \oplus N$.

(b) Let $\pi: X \rightarrow X / M$ be the quotient map, let $\{ e_1,...,e_n \}$ be a basis for $X/M$, pick $x_i \in X$ so that $\pi(x_i)=e_i$ for $1 \leq i \leq n$, and let $N$ be the vector space spanned by $\{ x_1,...,x_n \}$. Then $X=M \oplus N$.


(1) How to show that $X \subseteq M + N$ for both cases?

(2) How to show that $M \cap N = \{ 0 \}$ for (b)?

(3) Why we need $X$ to be locally convex for (a)?

  • $\begingroup$ Why can you extend the functional alpha_i to all of X? X is not a banach space here so one needs to be careful as the statement demands a sublinear function p which is greater than the functionals. $\endgroup$ – random123 Oct 30 '15 at 5:10
  • $\begingroup$ @Idonknow: You need local convexity to produce a sublinear function $p$ that dominates $\alpha_i$. For instance, the minkowski functional of a balanced, absorbing convex neighbourhood of the origin will work. $\endgroup$ – Prahlad Vaidyanathan Oct 30 '15 at 10:15

(1) We can extend $\alpha_i$ from $M$ to the whole of $X$ (here we need local convexity, as Hahn-Banach does not hold in the general case), let's call these extensions $x_i^* \in X^*$. Write $N := \bigcap_{i=1}^n \ker x_i^*$. We will show that $X = M + N$. Let $x \in X$, and define $m := \sum_{i=1}^n x_i^*(x)e_i \in M$. We have for $1 \le j \le n$, that \begin{align*} x_j^*(m) &= \sum_{i=1}^n x_i^*(x)x_j^*(e_i)\\ &= \sum_{i=1}^n x_i^*(x)\alpha_j(e_i) \end{align*} Now note, that as the $e_i$ form a basis of $M$, the coefficients in $$ e_i = 0\cdot e_1 + \cdots + 1 \cdot e_i + \cdots + 0 \cdot e_n \stackrel!= \sum_{j=1}^n \alpha_j(e_i)e_j $$ are unique, hence $\alpha_j(e_i) = \delta_{ij}$, so we may continue \begin{align*} x_j^*(m) &= \sum_{i=1}^n x_i^*(x)\delta_{ij}\\ &= x_j^*(x) \end{align*} That is $x_j^*(x-m) =0$ for all $j$, hence $x-m \in N$, or $$ x \in m + N \subseteq M + N $$ as $x$ was arbitrary, we have $X = M+N$.

(2) We will start showing $X = M+N$, so let $x \in X$, write $$ \pi(x) = \sum_{i=1}^n \lambda_i e_i $$ for some $\lambda_i \in \mathbf K$. Define $n := \sum_{i=1}^n \lambda_i x_i$, then we have $$ \pi(n) = \sum_{i=1}^n \lambda_i \pi(x_i) = \sum_{i=1}^n \lambda_i e_i = \pi(x) $$ that is $\pi(x-n) = 0$ or $x-n \in M$. Hence $x \in n+M \subseteq M+N$.
To show $M \cap N = \{0\}$, let $x \in M \cap N$. As $x \in N$, we may write $x = \sum_{i=1}^n \lambda_i x_i$ for some $\lambda_i \in \mathbf K$. As $x \in M$, we have $\pi(x) = 0$. Now $$ 0 = \pi(x) = \sum_{i=1}^n \lambda_i \pi(x_i) = \sum_{i=1}^n \lambda_i e_i $$ As the $e_i$ are linearly independent, it follows that $\lambda_i = 0$, hence $x = \sum_i\lambda_i x_i = 0$.

  • $\begingroup$ in case $b)$ have $N$ same dimension that $X/M$ ? $\endgroup$ – user89940 Oct 24 '18 at 0:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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