ProofComplete

Hilbert Spaces - Key Proof

We present a detailed proof of the Projection Theorem, one of the most fundamental results in Hilbert space theory, which establishes the existence and uniqueness of best approximations in closed subspaces.

TheoremProjection Theorem

Let HH be a Hilbert space and MHM \subset H a closed subspace. Then for every xHx \in H, there exists a unique mMm \in M such that xm=infyMxy=d(x,M)\|x - m\| = \inf_{y \in M} \|x - y\| = d(x, M) Moreover, mm is characterized by the condition (xm)M(x - m) \perp M.

Proof

Existence: Let d=infyMxyd = \inf_{y \in M} \|x - y\|. Choose a sequence (mn)(m_n) in MM such that xmnd\|x - m_n\| \to d as nn \to \infty.

We show (mn)(m_n) is Cauchy using the parallelogram law: mnmk2=(xmk)(xmn)2\|m_n - m_k\|^2 = \|(x - m_k) - (x - m_n)\|^2 =2xmk2+2xmn22xmkmn2= 2\|x - m_k\|^2 + 2\|x - m_n\|^2 - \|2x - m_k - m_n\|^2 =2xmk2+2xmn24xmk+mn22= 2\|x - m_k\|^2 + 2\|x - m_n\|^2 - 4\left\|x - \frac{m_k + m_n}{2}\right\|^2

Since MM is a subspace, (mk+mn)/2M(m_k + m_n)/2 \in M, so xmk+mn22d2\left\|x - \frac{m_k + m_n}{2}\right\|^2 \geq d^2

As n,kn, k \to \infty, we have xmn2,xmk2d2\|x - m_n\|^2, \|x - m_k\|^2 \to d^2, therefore mnmk22d2+2d24d2=0\|m_n - m_k\|^2 \leq 2d^2 + 2d^2 - 4d^2 = 0

So (mn)(m_n) is Cauchy. Since HH is complete and MM is closed, mnmMm_n \to m \in M. By continuity of the norm, xm=d\|x - m\| = d.

Orthogonality Characterization: For any yMy \in M and λK\lambda \in \mathbb{K}, we have m+λyMm + \lambda y \in M, so xm2xmλy2=xm22Re(λxm,y)+λ2y2\|x - m\|^2 \leq \|x - m - \lambda y\|^2 = \|x - m\|^2 - 2\text{Re}(\lambda \langle x - m, y \rangle) + |\lambda|^2 \|y\|^2

This gives 2Re(λxm,y)λ2y22\text{Re}(\lambda \langle x - m, y \rangle) \leq |\lambda|^2 \|y\|^2 for all λ\lambda.

Choosing λ=txm,y\lambda = t \langle x - m, y \rangle for small t>0t > 0 yields 2txm,y2t2xm,y2y22t |\langle x - m, y \rangle|^2 \leq t^2 |\langle x - m, y \rangle|^2 \|y\|^2. Dividing by tt and letting t0t \to 0 gives xm,y=0\langle x - m, y \rangle = 0 for all yMy \in M.

Uniqueness: If m1,m2Mm_1, m_2 \in M both satisfy xm1=xm2=d\|x - m_1\| = \|x - m_2\| = d, then by the parallelogram law: m1m22=2xm12+2xm224xm1+m222\|m_1 - m_2\|^2 = 2\|x - m_1\|^2 + 2\|x - m_2\|^2 - 4\left\|x - \frac{m_1 + m_2}{2}\right\|^2 2d2+2d24d2=0\leq 2d^2 + 2d^2 - 4d^2 = 0

Therefore m1=m2m_1 = m_2.

ExampleComputing Projections
  1. Finite-Dimensional Subspaces: If M=span{e1,,en}M = \text{span}\{e_1, \ldots, e_n\} is orthonormal, then PM(x)=i=1nx,eieiP_M(x) = \sum_{i=1}^n \langle x, e_i \rangle e_i

  2. Fourier Series: The projection of fL2[0,2π]f \in L^2[0, 2\pi] onto trigonometric polynomials of degree N\leq N is the partial Fourier sum

Remark

The Projection Theorem requires the subspace MM to be closed. For non-closed subspaces, the infimum may not be attained. This is why completeness of Hilbert spaces is essential.

This theorem is the foundation for least squares approximation, Fourier analysis, and variational methods in PDEs.