0.
Preliminaries: Vector Spaces, Inner Products, Matrices
1. Field ℝ
o Contains elements 0,1 with usual arithmetic; closed under +,×;
+,× associative and commutative; distributivity; each nonzero
has a multiplicative inverse; an ordered complete field.
2. Real Vector Space ℝⁿ
o Elements are n-tuples x=(x₁,…,xₙ).
o Vector addition and scalar multiplication satisfy the eight axioms
(commutativity, associativity, zero vector, additive inverses,
distributivity, etc.).
3. Standard Inner Product
o For x,y∈ℝⁿ, define
n
⟨ x , y ⟩:=∑ x i y i .
i=1
o Axioms:
1. Symmetry: ⟨x,y⟩=⟨y,x⟩.
2. Linearity in the first slot: ⟨αx+βz,y⟩=α⟨x,y⟩+β⟨z,y⟩.
3. Positive‐definiteness: ⟨x,x⟩>0 for x≠0; ⟨x,x⟩=0⇔x=0.
4. Matrices and Transpose
o A∈ℝ^{n×n} acts on x∈ℝⁿ by the usual rule.
o The transpose Aᵀ satisfies (Aᵀ){ij}=A{ji}.
o Symmetric: A=Aᵀ means A_{ij}=A_{ji} for all i,j.
5. Orthogonality and Orthonormal Sets
o Vectors u,v are orthogonal iff ⟨u,v⟩=0.
o A set {q₁,…,qₖ} is orthonormal if ⟨qᵢ,qⱼ⟩=δ_{ij} (Kronecker
delta).
6. Eigenvalues and Eigenvectors
o λ∈ℂ and v∈ℂⁿ{0} satisfy Av=λv. Such (λ,v) is an eigenpair.
o If v can be chosen in ℝⁿ, we call it a real eigenvector.
1. All Eigenvalues Are Real
Let A∈ℝ^{n×n} be symmetric. Suppose (λ,v) is an eigenpair with v∈ℂⁿ,
v≠0. We will show λ∈ℝ.
1. Form the standard Hermitian inner product on ℂⁿ:
n
⟨ x , y ⟩:=∑ x i y i .
i=1
2. Compute
⟨ Av , v ⟩=∑ ( Av )i v i=∑ ∑
i i ( j )
A ij v j v i =v ¿ A v=v ¿ ( λv )=λ v ¿ v=λ ⟨ v , v ⟩ .
3. On the other hand, since A is real and symmetric, Aᵀ=A implies A is
Hermitian (A*=A), so
⟨ Av , v ⟩=⟨ v , Av ⟩=⟨ Av , v ⟩ .
Hence the scalar ⟨Av,v⟩ equals its own complex conjugate, so it is real.
4. Thus
λ ⟨ v , v ⟩=⟨ Av , v ⟩ ∈ R .
But ⟨v,v⟩>0 (since v≠0). Therefore λ is real.
2. Eigenvectors for Distinct Eigenvalues Are Orthogonal
Let λ₁≠λ₂ be two real eigenvalues with real eigenvectors v₁,v₂∈ℝⁿ. Then
1. ⟨Av₁,v₂⟩ = ⟨λ₁v₁, v₂⟩ = λ₁⟨v₁,v₂⟩.
2. ⟨v₁,Av₂⟩ = ⟨v₁,λ₂v₂⟩ = λ₂⟨v₁,v₂⟩.
3. But by symmetry of A, ⟨Av₁,v₂⟩ = ⟨v₁,Av₂⟩. Hence
λ ₁ ⟨ v ₁, v ₂ ⟩=λ ₂ ⟨ v ₁ , v ₂ ⟩ ⇒ ( λ ₁− λ ₂ ) ⟨ v ₁ , v ₂ ⟩=0 .
Since λ₁≠λ₂, we get ⟨v₁,v₂⟩=0.
3. Existence of n Real Eigenvalues and an Orthonormal
Basis
We now prove there are n real eigenvalues (counting algebraic multiplicity)
and an orthonormal basis of eigenvectors. We give two proofs.
3.1 Algebraic Proof via the Characteristic Polynomial
1. The characteristic polynomial of A is
p A ( λ ) :=det ( A− λI ) ∈ R [ λ ] ,
a real‐coefficient polynomial of degree n.
2. By the Fundamental Theorem of Algebra, p_A(λ) factors over ℂ as
p A ( λ )=( λ−λ 1) ⋯ ( λ−λ n )
for some λ₁,…,λₙ∈ℂ (counted with multiplicity).
3. From §1, each λᵢ must be real. Hence all n roots are in ℝ.
4. For each (real) root λᵢ, there is a nonzero eigenvector vᵢ with (A−λᵢI)vᵢ=0,
so the geometric multiplicity of λᵢ is ≥1 and ≤ its algebraic multiplicity.
5. Within each eigenspace E_{λᵢ} = ker(A−λᵢI), choose a basis and then
apply the Gram–Schmidt process to convert it into an orthonormal basis. By
§2, vectors from distinct eigenspaces are already orthogonal. Altogether, we
obtain n orthonormal eigenvectors {q₁,…,qₙ}.
6. Form Q=[q₁ … qₙ]∈ℝ^{n×n}. Then QTQ=I and
AQ=[ Aq ₁ … Aq ₙ ] =[ λ ₁ q ₁ … λ ₙq ₙ ] =Q diag ( λ ₁ ,… , λ ₙ ) .
Hence
T T
A=Q diag ( λ1 ,… , λ n ) Q =QΛ Q
3.2 Analytic Proof via the Rayleigh Quotient and Induction
This proof avoids the characteristic polynomial altogether.
1. Define the Rayleigh quotient R: S^{n−1}→ℝ by
⟨ Ax , x ⟩
R ( x )= , x≠0.
⟨ x, x ⟩
Restricted to the unit sphere S^{n−1}={x:‖x‖=1}, R is a continuous
function on a compact set, so it attains a maximum μ and a minimum ν,
both real.
2. Let q be any unit‐vector achieving the maximum μ. We show Aq=μq:
For any small vector h orthogonal to q (⟨q,h⟩=0), consider f(t)=R(q + t
h). Because R has an extremum at q on the sphere, the directional
derivative in any tangent direction must vanish. One computes (by
expanding numerator and denominator to first order) that
$$ Aq,q,q,h
o Aq,h,q,q= 0
Aq,h= μ,q,h= 0. $$
Since this holds for all h⊥q, we deduce Aq is a linear combination of q
alone, hence Aq=μ q. So (μ,q) is an eigenpair.
3. Now set V₁ = span{q}. Its orthogonal complement V₁^⊥ has dimension
n−1 and is invariant under A: if x⊥q then
⟨ Ax , q ⟩=⟨ x , Aq ⟩=⟨ x , μq ⟩=μ ⟨ x , q ⟩=0 ,
so A x∈V₁^⊥.
4. By induction on dimension, there is an orthonormal basis of V₁^⊥
consisting of eigenvectors of A|_{V₁^⊥}. Adjoining q gives an orthonormal
basis of ℝⁿ of eigenvectors.
4. Putting It All Together: Spectral Decomposition
Eigenvalues λ₁,…,λₙ are real (possibly with repetition).
Eigenvectors q₁,…,qₙ can be chosen orthonormal.
Assemble Q=[q₁ … qₙ], Λ=diag(λ₁,…,λₙ). Then
T T
Q Q=I , A=QΛ Q .
Equivalently,
n
A=∑ λi q i q Ti ,
i=1
where each projector Pᵢ = qᵢ qᵢ^T is rank-1 and PᵢPⱼ=0 for i≠j, ∑ᵢPᵢ=I.
5. Illustrative 2×2 Example (All Steps Written Out)
Let
A= (21 12) .
1. Verify symmetry: Aᵀ=A.
2. Characteristic polynomial:
det ( A−λI ) =det (2−λ
1
1
2−λ ) 2 2
=( 2−λ ) −1=λ −4 λ +3 .
3. Solve λ²−4λ+3=0 ⇒ λ=1 or λ=3.
4. Eigenvectors:
o For λ=3: solve (A−3I)v=0:
(−11 −11 )( vv )=0 ⇒ v =v .
1
2
1 2
Pick v^{(1)}=(1,1)ᵀ.
o For λ=1: solve (A−I)v=0:
(11 11)( vv )=0 ⇒ v =−v .
1
2
1 2
Pick v^{(2)}=(1,−1)ᵀ.
5. Normalize:
q 1=
1 1
√2 1 ()
,q 2=
1 1
√ 2 −1
. ( )
6. Assemble Q = [q₁ q₂], Λ=diag(3,1). Check QᵀQ=I and A=QΛQᵀ:
Q ΛQ =
T
(
1 1 1 3 0 1 1
)( )(
2 1 −1 0 1 1 −1
=
2 1
1 2
=A. )( )
Final Conclusion
Every real symmetric matrix A admits an orthonormal basis of ℝⁿ
consisting of its eigenvectors, and thus is diagonalizable by an orthogonal
change of basis. In symbols:
$$ A = Q\,Λ\,Q^T,\quad Q^T Q = I,\quad Λ=\mathrm{diag}(\lambda_1,\
dots,\lambda_n)\inℝ^{n×n}. $$
This is the Spectral Theorem in the real symmetric case, proved in full
detail from field, vector space, and inner product axioms with no omitted
steps.
