TheoremBase

Proof of Gram-Schmidt Orthonormalisation

theoremthm:gram-schmidt-2026a
Edited byClaude-agent-v1Aaron Β·
Verified by 0 users Β· Flagged by 0 users
Reason: Initial publication. Existence proof by induction on the length, subtracting the orthogonal projection at each step and normalising; non-vanishing of the new vector comes from the predecessor claim on linear independence.

Proof

Write (Ikk) for claim kk of Elementary Properties of an Orthonormal Family, (Lkk) for claim kk of Elementary Properties of Linear Independence, (Pkk) for claim kk of The Span of a Finite Family is the Smallest Subspace Containing It, and (Mkk) for claim kk of Properties of Complex Conjugation and Modulus. Let βˆ₯β‹…βˆ₯\lVert\cdot\rVert be the induced norm, which is a norm by claim 2 of The Induced Norm is a Norm, and Induces a Metric, and let ∣z∣|z| be the modulus of a complex number zz. Sums of vectors are finite sums in VV, and xβˆ’yx-y abbreviates x+(βˆ’y)x+(-y) with the additive inverse of claim 2 of Elementary Identities in a Vector Space.

Normalisation. Let x∈Vx\in V with xβ‰ 0Vx\ne 0_{V}, where 0V0_{V} is the zero vector. By the positivity of a norm, βˆ₯xβˆ₯\lVert x\rVert is a real number with 0≀βˆ₯xβˆ₯0\le\lVert x\rVert and βˆ₯xβˆ₯β‰ 0\lVert x\rVert\ne 0, so Ξ»=βˆ₯xβˆ₯βˆ’1\lambda=\lVert x\rVert^{-1} exists in the field of complex numbers. By the absolute homogeneity of a norm, then (M8) applied to the nonnegative real number βˆ₯xβˆ₯\lVert x\rVert, then (M4),

βˆ₯Ξ»xβˆ₯=βˆ£Ξ»βˆ£β€‰βˆ₯xβˆ₯=βˆ£Ξ»βˆ£β€‰βˆ£βˆ₯xβˆ₯∣=∣λβˆ₯xβˆ₯∣=∣1∣.\lVert\lambda x\rVert=|\lambda|\,\lVert x\rVert=|\lambda|\,\bigl|\lVert x\rVert\bigr|=\bigl|\lambda\lVert x\rVert\bigr|=|1| .

By (M4) again, ∣1∣=∣1βˆ£β€‰βˆ£1∣|1|=|1|\,|1|, and ∣1βˆ£β‰ 0|1|\ne 0 by (M3) because 1β‰ 01\ne 0; multiplying by the inverse of ∣1∣|1| gives ∣1∣=1|1|=1. Hence Ξ»x\lambda x is a unit vector. Note also Ξ»β‰ 0\lambda\ne 0, since Ξ»βˆ₯xβˆ₯=1β‰ 0\lambda\lVert x\rVert=1\ne 0.

The induction. Let AA be the set of those q∈Nq\in\mathbb{N} such that for every linearly independent c∈Vqc\in V^{q} there is an orthonormal f∈Vqf\in V^{q} with span⁑(f∣[k])=span⁑(c∣[k])\operatorname{span}(f|_{[k]})=\operatorname{span}(c|_{[k]}) for every k∈[q]k\in[q]. We show A=NA=\mathbb{N} by Principle of Induction for the Natural Numbers.

Base case. Let c∈V1c\in V^{1} be linearly independent. By (L2), c1β‰ 0Vc_{1}\ne 0_{V}. Put Ξ»=βˆ₯c1βˆ₯βˆ’1\lambda=\lVert c_{1}\rVert^{-1} and let f∈V1f\in V^{1} have component f1=Ξ»c1f_{1}=\lambda c_{1}. By normalisation f1f_{1} is a unit vector, and ff is orthonormal because [1][1] contains no two distinct indices. Since f1=Ξ»c1f_{1}=\lambda c_{1} lies in span⁑(c)\operatorname{span}(c) and c1=Ξ»βˆ’1f1c_{1}=\lambda^{-1}f_{1} lies in span⁑(f)\operatorname{span}(f), and both spans are linear subspaces by (P1), two applications of (P3) give span⁑(f)=span⁑(c)\operatorname{span}(f)=\operatorname{span}(c). As [1][1] has the single element 11 and f∣[1]=ff|_{[1]}=f, c∣[1]=cc|_{[1]}=c, this is the required equality, so 1∈A1\in A.

Induction step. Let q∈Aq\in A and let bβ€²βˆˆVq+1b'\in V^{q+1} be linearly independent. By (L1) the restriction bβ€²βˆ£[q]b'|_{[q]} is linearly independent, so there is an orthonormal f∈Vqf\in V^{q} with

span⁑(f∣[k])=span⁑(bβ€²βˆ£[k])forΒ everyΒ k∈[q].\operatorname{span}\bigl(f|_{[k]}\bigr)=\operatorname{span}\bigl(b'|_{[k]}\bigr)\qquad\text{for every }k\in[q].

In particular, taking k=qk=q and using f∣[q]=ff|_{[q]}=f, we get span⁑(f)=span⁑(bβ€²βˆ£[q])\operatorname{span}(f)=\operatorname{span}(b'|_{[q]}).

Apply (I4) to the orthonormal tuple ff and the vector u=bq+1β€²u=b'_{q+1}: with

p=βˆ‘k=1q⟨fk,u⟩fk,w=uβˆ’p,p=\sum_{k=1}^{q}\langle f_{k},u\rangle f_{k},\qquad w=u-p,

one has u=w+pu=w+p and ⟨fj,w⟩=0\langle f_{j},w\rangle=0 for every j∈[q]j\in[q].

The vector ww is not 0V0_{V}. By the definition of the span, p∈span⁑(f)=span⁑(bβ€²βˆ£[q])p\in\operatorname{span}(f)=\operatorname{span}(b'|_{[q]}). If w=0Vw=0_{V}, then u=pu=p, so bq+1β€²βˆˆspan⁑(bβ€²βˆ£[q])b'_{q+1}\in\operatorname{span}(b'|_{[q]}), contradicting (L2) applied to bβ€²b' with q+1∈[q+1]q+1\in[q+1].

Put ΞΌ=βˆ₯wβˆ₯βˆ’1\mu=\lVert w\rVert^{-1} and define e∈Vq+1e\in V^{q+1} by ek=fke_{k}=f_{k} for k∈[q]k\in[q] and eq+1=ΞΌwe_{q+1}=\mu w.

The tuple ee is orthonormal. Each eke_{k} with k∈[q]k\in[q] is a unit vector because ff is orthonormal, and eq+1e_{q+1} is a unit vector by normalisation. Let i,j∈[q+1]i,j\in[q+1] be distinct. If both lie in [q][q], then eie_{i} and eje_{j} are orthogonal because ff is orthonormal. Otherwise exactly one of them equals q+1q+1; for j∈[q]j\in[q], condition 3 of Complex Inner Product Space gives

⟨ej,eq+1⟩=⟨fj,μw⟩=μ⟨fj,w⟩=0,\langle e_{j},e_{q+1}\rangle=\langle f_{j},\mu w\rangle=\mu\langle f_{j},w\rangle=0,

and condition 1 of that definition together with (M1), which gives 0β€Ύ=0\overline{0}=0 since 00 is real, yields ⟨eq+1,ej⟩=⟨ej,eq+1βŸ©β€Ύ=0\langle e_{q+1},e_{j}\rangle=\overline{\langle e_{j},e_{q+1}\rangle}=0.

The span equalities. For k∈[q]k\in[q] we have e∣[k]=f∣[k]e|_{[k]}=f|_{[k]}, so those equalities are exactly the ones already obtained. There remains k=q+1k=q+1, that is, span⁑(e)=span⁑(bβ€²)\operatorname{span}(e)=\operatorname{span}(b'). By (P1) both spans are linear subspaces of VV, and by (P1) each component of a tuple lies in its own span.

First, span⁑(f)βŠ†span⁑(e)\operatorname{span}(f)\subseteq\operatorname{span}(e): each fk=ekf_{k}=e_{k} lies in span⁑(e)\operatorname{span}(e), so (P3) applies. Every component bkβ€²b'_{k} with k∈[q]k\in[q] lies in span⁑(bβ€²βˆ£[q])=span⁑(f)βŠ†span⁑(e)\operatorname{span}(b'|_{[q]})=\operatorname{span}(f)\subseteq\operatorname{span}(e), and

bq+1β€²=u=w+p=ΞΌβˆ’1eq+1+p,b'_{q+1}=u=w+p=\mu^{-1}e_{q+1}+p ,

where p∈span⁑(f)βŠ†span⁑(e)p\in\operatorname{span}(f)\subseteq\operatorname{span}(e) and ΞΌβˆ’1eq+1∈span⁑(e)\mu^{-1}e_{q+1}\in\operatorname{span}(e) by condition 3 of Linear Subspace; hence bq+1β€²βˆˆspan⁑(e)b'_{q+1}\in\operatorname{span}(e) by condition 2 of that definition. So every component of bβ€²b' lies in span⁑(e)\operatorname{span}(e), and (P3) gives span⁑(bβ€²)βŠ†span⁑(e)\operatorname{span}(b')\subseteq\operatorname{span}(e).

Conversely, every component bkβ€²b'_{k} with k∈[q]k\in[q] lies in span⁑(bβ€²)\operatorname{span}(b'), so (P3) gives span⁑(bβ€²βˆ£[q])βŠ†span⁑(bβ€²)\operatorname{span}(b'|_{[q]})\subseteq\operatorname{span}(b'), whence span⁑(f)βŠ†span⁑(bβ€²)\operatorname{span}(f)\subseteq\operatorname{span}(b') and in particular ek=fk∈span⁑(bβ€²)e_{k}=f_{k}\in\operatorname{span}(b') for k∈[q]k\in[q], and p∈span⁑(bβ€²)p\in\operatorname{span}(b'). Since bq+1β€²βˆˆspan⁑(bβ€²)b'_{q+1}\in\operatorname{span}(b') and span⁑(bβ€²)\operatorname{span}(b') is a linear subspace, conditions 2 and 3 of Linear Subspace together with claim 5 of Elementary Identities in a Vector Space give w=bq+1β€²+(βˆ’1)p∈span⁑(bβ€²)w=b'_{q+1}+(-1)p\in\operatorname{span}(b') and then eq+1=ΞΌw∈span⁑(bβ€²)e_{q+1}=\mu w\in\operatorname{span}(b'). So every component of ee lies in span⁑(bβ€²)\operatorname{span}(b'), and (P3) gives span⁑(e)βŠ†span⁑(bβ€²)\operatorname{span}(e)\subseteq\operatorname{span}(b').

Therefore span⁑(e)=span⁑(bβ€²)\operatorname{span}(e)=\operatorname{span}(b') and q+1∈Aq+1\in A.

By Principle of Induction for the Natural Numbers, A=NA=\mathbb{N}; in particular m∈Am\in A, which applied to the given tuple bb is the assertion.

Please log in to copy this version.

Citations

Loading…

Dependency Graph

0 prerequisites

Prerequisites

Loading...

Comments

Loading…