Eigenspaces

Suppose a linear operator stretches some vectors and shrinks others. It leaves a whole line of vectors fixed except for scaling. A natural question arises. For a fixed scalar λ\lambda, what is the entire collection of vectors scaled by exactly λ\lambda? This collection is never a random, shapeless set. It is always a subspace, and it deserves a name: the eigenspace of λ\lambda. Eigenspaces matter for an important reason. They organize the scattered individual eigenvectors from Eigenvalues and Eigenvectors. These become coherent geometric objects — lines, planes, or higher-dimensional flats. On these objects, an operator acts by pure scaling. Understanding an operator often reduces to understanding how it splits the ambient space into such directions. Eigenspaces are the precise bookkeeping device for that splitting. The big picture is this: an eigenspace is literally a familiar object in disguise. It is the Kernel and null space of the shifted operator T−λIT - \lambda I. This identification is not a coincidence or a convenient trick. It is the reason eigenspaces are automatically subspaces. It is also the reason they are TT-invariant. And it is the computational engine behind every eigenvector calculation you have ever done via row reduction. This article gives the precise definition of an eigenspace. It proves the eigenspace is a subspace and is TT-invariant. It works through two asymmetric numerical examples. It also shows exactly what breaks if you drop the hypothesis that λ\lambda is an eigenvalue. It sets up the language needed for Geometric vs Algebraic Multiplicity and the eventual Diagonalizability Criterion.

Key takeaways
  • The eigenspace of TT for eigenvalue λ\lambda is E(λ,T)=null⁡(T−λI)E(\lambda, T) = \operatorname{null}(T - \lambda I).
  • It is always a subspace, containing 00 along with every eigenvector for λ\lambda.
  • Eigenspaces for distinct eigenvalues intersect only at 00, and in fact form a direct sum.
  • Every eigenspace is TT-invariant: TT maps E(λ,T)E(\lambda,T) into itself.
  • The dimension of E(λ,T)E(\lambda,T) is the geometric multiplicity of λ\lambda, a key ingredient in diagonalizability.
  • Key idea: an eigenspace is nothing but the null space of a shifted operator, so all null-space machinery transfers immediately to eigenvector-finding.

Where this lives today

Eigenspaces are the computational backbone of diagonalization: once you know the eigenspaces span the whole space, their union of bases forms an eigenbasis, letting you write TT as a diagonal matrix in a new coordinate system — this is the subject of Diagonalizability Criterion. In quantum mechanics, the eigenspace of an observable operator for eigenvalue λ\lambda is exactly the subspace of quantum states for which a measurement is guaranteed to yield the value λ\lambda; degenerate eigenspaces (dimension greater than 11) correspond to physically distinct states sharing the same measured energy or momentum. In statistics and data science, Principal Component Analysis diagonalizes a covariance matrix; the eigenspace for the largest eigenvalue is the direction of maximal variance in the data, and eigenspaces for smaller eigenvalues capture progressively less-important directions of spread. In graph theory and network science, the eigenspace of a graph's adjacency or Laplacian matrix for eigenvalue 00 encodes connected components; spectral clustering algorithms partition data by examining low-dimensional eigenspaces of a similarity graph's Laplacian. In engineering, the eigenspaces of a system's state matrix determine the modes of a dynamical system — each eigenspace corresponds to a mode of oscillation or decay with a characteristic rate given by its eigenvalue, crucial in vibration analysis and control theory. In cryptography and coding theory, eigenspace decompositions of linear operators over finite fields underlie algorithms for discrete logarithms and for analyzing linear feedback shift registers.

Ask the tutor: If two different eigenvalues happen to have eigenspaces of the same dimension, does that tell us anything about whether the operator is diagonalizable?

DefinitionEigenspace

Let VV be a vector space over a field F\mathbb{F}, let T:V→VT : V \to V be a linear operator, and let λ∈F\lambda \in \mathbb{F} be an eigenvalue of TT (meaning there exists a nonzero v∈Vv \in V with Tv=λvTv = \lambda v). The eigenspace of TT corresponding to λ\lambda is the set

E(λ,T)  =  { v∈V:Tv=λv }  =  null⁡(T−λI),E(\lambda, T) \;=\; \{\, v \in V : Tv = \lambda v \,\} \;=\; \operatorname{null}(T - \lambda I),

where I:V→VI : V \to V is the identity operator, so (T−λI)(v)=Tv−λv(T-\lambda I)(v) = Tv - \lambda v. Equivalently, if VV is finite-dimensional with dim⁡V=n\dim V = n and AA is the matrix of TT in some basis, then

E(λ,T)  =  { x∈Fn:(A−λIn)x=0 },E(\lambda, T) \;=\; \{\, x \in \mathbb{F}^n : (A - \lambda I_n)x = 0 \,\},

the solution set of the homogeneous linear system (A−λI)x=0(A-\lambda I)x=0. By convention E(λ,T)E(\lambda, T) contains the zero vector even though 00 is not itself called an eigenvector; the nonzero elements of E(λ,T)E(\lambda,T) are precisely the eigenvectors of TT associated to λ\lambda.

Intuition

Think of TT as a machine that takes in vectors and spits out transformed vectors. Most input directions get rotated, sheared, or otherwise distorted. But for an eigenvalue λ\lambda, there is a special direction (or family of directions) where TT acts exactly like multiplication by the scalar λ\lambda — no rotation, no shearing, purely a rescale. The eigenspace E(λ,T)E(\lambda,T) collects all such directions into one subspace. The key trick that makes eigenspaces easy to compute is rewriting the eigenvalue equation. The condition Tv=λvTv = \lambda v looks like it mixes TT and scalar multiplication awkwardly, but subtracting gives (T−λI)v=0(T - \lambda I)v = 0. This is now a statement that vv lies in a null space — an object you already know is always a subspace, closed under addition and scalar multiplication, for free, with no eigenvalue-specific argument needed. The entire subtlety of eigenspaces evaporates once you see them as null spaces of a shifted operator. Geometrically, in R2\mathbb{R}^2 or R3\mathbb{R}^3, an eigenspace is typically a line through the origin (dimension 11) or, in degenerate cases, a plane through the origin (dimension 22) or the entire space (dimension nn, only when TT is itself a scalar multiple of the identity). The picture below shows a 2×22\times 2 matrix with two distinct eigenvalues, each giving a one-dimensional eigenspace — two special lines along which the map is a pure stretch, while every other vector gets rotated toward one of these lines under repeated application of TT.

Figure
Two eigenspaces (lines) of a 2×22\times2 operator with eigenvalues λ1=3, λ2=−1\lambda_1=3,\ \lambda_2=-1
x y E(3,T): span(1,1) E(−1,T): span(1,−2) v=(1,1) Tv=3v=(3,3) w=(1,−2) Tw=−w=(−1,2) Origin at (210,150); each colored line is one eigenspace.

Central results

TheoremEigenspace Structure Theorem

Let VV be a vector space over a field F\mathbb{F}, let T:V→VT:V\to V be linear, and let λ∈F\lambda \in \mathbb{F}.

(a) E(λ,T)=null⁡(T−λI)E(\lambda, T) = \operatorname{null}(T - \lambda I) is a subspace of VV (whether or not λ\lambda is an eigenvalue; if λ\lambda is not an eigenvalue, this subspace is just {0}\{0\}).

(b) E(λ,T)E(\lambda, T) is TT-invariant: if v∈E(λ,T)v \in E(\lambda, T) then Tv∈E(λ,T)Tv \in E(\lambda, T).

(c) If λ1,…,λk\lambda_1, \dots, \lambda_k are distinct eigenvalues of TT, then the sum

E(λ1,T)+E(λ2,T)+⋯+E(λk,T)E(\lambda_1, T) + E(\lambda_2, T) + \cdots + E(\lambda_k, T)

is a direct sum; equivalently, if v1+v2+⋯+vk=0v_1 + v_2 + \cdots + v_k = 0 with vi∈E(λi,T)v_i \in E(\lambda_i, T) for each ii, then v1=v2=⋯=vk=0v_1 = v_2 = \cdots = v_k = 0.

(d) In particular, nonzero eigenvectors corresponding to distinct eigenvalues are linearly independent.

Proof

(a) E(λ,T)E(\lambda,T) is a subspace. We use the technique of directly verifying the Kernel and null space subspace criteria for the linear map S:=T−λIS := T - \lambda I. First, SS is linear: for u,w∈Vu,w \in V and scalars a,b∈Fa,b \in \mathbb{F},

S(au+bw)=T(au+bw)−λ(au+bw)=a Tu+b Tw−aλu−bλw=a Su+b Sw,S(au+bw) = T(au+bw) - \lambda(au+bw) = a\,Tu + b\,Tw - a\lambda u - b\lambda w = a\,Su + b\,Sw,

using linearity of TT in the second equality. So SS is a linear operator on VV, and E(λ,T)=null⁡(S)E(\lambda,T) = \operatorname{null}(S). Now verify the null space is a subspace directly (this is the standard argument, repeated here so the proof is self-contained): Contains 00: S(0)=0S(0) = 0 since SS is linear, so 0∈null⁡(S)=E(λ,T)0 \in \operatorname{null}(S) = E(\lambda,T). Closed under addition: if v1,v2∈E(λ,T)v_1, v_2 \in E(\lambda,T), i.e. Sv1=0Sv_1 = 0 and Sv2=0Sv_2=0, then by linearity

S(v1+v2)=Sv1+Sv2=0+0=0,S(v_1+v_2) = Sv_1 + Sv_2 = 0+0 = 0,

so v1+v2∈E(λ,T)v_1+v_2 \in E(\lambda,T). Closed under scalar multiplication: if v∈E(λ,T)v \in E(\lambda,T) and c∈Fc \in \mathbb{F}, then

S(cv)=c Sv=c⋅0=0,S(cv) = c\,Sv = c\cdot 0 = 0,

so cv∈E(λ,T)cv \in E(\lambda,T). Hence E(λ,T)E(\lambda,T) is a subspace of VV. If λ\lambda is not an eigenvalue of TT, then by definition no nonzero vector satisfies Tv=λvTv=\lambda v, so null⁡(S)={0}\operatorname{null}(S) = \{0\}, still a (trivial) subspace.

(b) TT-invariance. Let v∈E(λ,T)v \in E(\lambda, T), so Tv=λvTv = \lambda v. We must show Tv∈E(λ,T)Tv \in E(\lambda,T), i.e. that T(Tv)=λ(Tv)T(Tv) = \lambda(Tv). Compute, using linearity of TT and that scalars commute with TT:

T(Tv)=T(λv)=λ Tv.T(Tv) = T(\lambda v) = \lambda\, T v.

This is exactly the statement T(Tv)=λ(Tv)T(Tv) = \lambda(Tv), so Tv∈E(λ,T)Tv \in E(\lambda,T). Thus TT maps E(λ,T)E(\lambda,T) into itself, proving TT-invariance.

(c) Directness of the sum, distinct eigenvalues. We argue by induction on kk, the technique being a standard "distinct eigenvalues" induction. Base case k=1k=1: the statement "v1=0v_1 = 0 whenever v1∈E(λ1,T)v_1 \in E(\lambda_1,T) and v1=0v_1=0" is trivial. Inductive step: suppose the claim holds for any k−1k-1 distinct eigenvalues (induction hypothesis), and suppose

v1+v2+⋯+vk=0,vi∈E(λi,T),v_1 + v_2 + \cdots + v_k = 0, \qquad v_i \in E(\lambda_i, T),

with λ1,…,λk\lambda_1,\dots,\lambda_k distinct. Apply TT to both sides. Since each viv_i satisfies Tvi=λiviTv_i = \lambda_i v_i (part (b) is not even needed here, just the definition of E(λi,T)E(\lambda_i,T)), we get

λ1v1+λ2v2+⋯+λkvk=T(0)=0.(∗)\lambda_1 v_1 + \lambda_2 v_2 + \cdots + \lambda_k v_k = T(0) = 0. \tag{$\ast$}

Also multiply the original relation by λk\lambda_k:

λkv1+λkv2+⋯+λkvk=0.(∗∗)\lambda_k v_1 + \lambda_k v_2 + \cdots + \lambda_k v_k = 0. \tag{$\ast\ast$}

Subtract (∗∗)(\ast\ast) from (∗)(\ast):

(λ1−λk)v1+(λ2−λk)v2+⋯+(λk−1−λk)vk−1+0⋅vk=0.(\lambda_1 - \lambda_k)v_1 + (\lambda_2-\lambda_k)v_2 + \cdots + (\lambda_{k-1}-\lambda_k)v_{k-1} + 0\cdot v_k = 0.

This is a relation among vectors v1,…,vk−1v_1,\dots,v_{k-1} lying in E(λ1,T),…,E(λk−1,T)E(\lambda_1,T),\dots,E(\lambda_{k-1},T) respectively (note the coefficient (λi−λk)vi(\lambda_i - \lambda_k)v_i still lies in E(λi,T)E(\lambda_i,T) by part (a), since eigenspaces are closed under scalar multiplication). By the induction hypothesis applied to the k−1k-1 distinct eigenvalues λ1,…,λk−1\lambda_1,\dots,\lambda_{k-1},

(λi−λk)vi=0for each i=1,…,k−1.(\lambda_i - \lambda_k)v_i = 0 \quad\text{for each } i=1,\dots,k-1.

Since λ1,…,λk\lambda_1,\dots,\lambda_k are distinct, λi−λk≠0\lambda_i - \lambda_k \neq 0 for i<ki < k, so vi=0v_i = 0 for i=1,…,k−1i=1,\dots,k-1. Substituting back into the original relation v1+⋯+vk=0v_1+\cdots+v_k=0 gives vk=0v_k = 0 as well. Thus v1=⋯=vk=0v_1=\cdots=v_k=0, completing the induction and proving the sum E(λ1,T)+⋯+E(λk,T)E(\lambda_1,T)+\cdots+E(\lambda_k,T) is direct.

(d) Independence of eigenvectors. Suppose v1,…,vkv_1,\dots,v_k are nonzero eigenvectors with Tvi=λiviTv_i = \lambda_i v_i and λ1,…,λk\lambda_1,\dots,\lambda_k distinct. If c1v1+⋯+ckvk=0c_1v_1+\cdots+c_kv_k=0 for scalars cic_i, then setting wi:=civi∈E(λi,T)w_i := c_iv_i \in E(\lambda_i,T), part (c) gives wi=0w_i = 0 for all ii, i.e. civi=0c_iv_i=0. Since vi≠0v_i \neq 0, this forces ci=0c_i=0 for every ii. Hence v1,…,vkv_1,\dots,v_k are linearly independent. ■\blacksquare

∎

Worked examples

Example 1 (a genuinely asymmetric 3×33\times 3 matrix). Let T:R3→R3T : \mathbb{R}^3 \to \mathbb{R}^3 be given by the matrix

A=(41−125−2112).A = \begin{pmatrix} 4 & 1 & -1 \\ 2 & 5 & -2 \\ 1 & 1 & 2 \end{pmatrix}.

We first find the eigenvalues via the characteristic polynomial det⁡(A−λI)=0\det(A-\lambda I)=0. Expanding along the first row:

det⁡(A−λI)=det⁡(4−λ1−125−λ−2112−λ).\det(A-\lambda I) = \det\begin{pmatrix} 4-\lambda & 1 & -1 \\ 2 & 5-\lambda & -2 \\ 1 & 1 & 2-\lambda \end{pmatrix}.

Cofactor expansion along the first row:

(4−λ)[(5−λ)(2−λ)−(−2)(1)]−1[2(2−λ)−(−2)(1)]+(−1)[2(1)−(5−λ)(1)].(4-\lambda)\big[(5-\lambda)(2-\lambda) - (-2)(1)\big] - 1\big[2(2-\lambda) - (-2)(1)\big] + (-1)\big[2(1) - (5-\lambda)(1)\big].

Compute each bracket:

(5−λ)(2−λ)+2=10−7λ+λ2+2=λ2−7λ+12,(5-\lambda)(2-\lambda)+2 = 10 -7\lambda+\lambda^2+2 = \lambda^2 - 7\lambda + 12,
2(2−λ)+2=4−2λ+2=6−2λ,2(2-\lambda)+2 = 4-2\lambda+2 = 6-2\lambda,
2−(5−λ)=λ−3.2-(5-\lambda) = \lambda - 3.

So

det⁡(A−λI)=(4−λ)(λ2−7λ+12)−(6−2λ)−(λ−3).\det(A-\lambda I) = (4-\lambda)(\lambda^2-7\lambda+12) - (6-2\lambda) - (\lambda-3).

Expand (4−λ)(λ2−7λ+12)=4λ2−28λ+48−λ3+7λ2−12λ=−λ3+11λ2−40λ+48(4-\lambda)(\lambda^2-7\lambda+12) = 4\lambda^2-28\lambda+48 -\lambda^3+7\lambda^2-12\lambda = -\lambda^3+11\lambda^2-40\lambda+48. Then

det⁡(A−λI)=−λ3+11λ2−40λ+48−6+2λ−λ+3=−λ3+11λ2−39λ+45.\det(A-\lambda I) = -\lambda^3+11\lambda^2-40\lambda+48 - 6+2\lambda -\lambda+3 = -\lambda^3+11\lambda^2-39\lambda+45.

Set this to zero: λ3−11λ2+39λ−45=0\lambda^3 - 11\lambda^2+39\lambda-45=0. Try λ=3\lambda=3: 27−99+117−45=027-99+117-45=0. ✓ So (λ−3)(\lambda-3) divides. Polynomial division:

λ3−11λ2+39λ−45=(λ−3)(λ2−8λ+15)=(λ−3)(λ−3)(λ−5).\lambda^3-11\lambda^2+39\lambda-45 = (\lambda-3)(\lambda^2-8\lambda+15) = (\lambda-3)(\lambda-3)(\lambda-5).

So eigenvalues are λ=3\lambda=3 (algebraic multiplicity 22) and λ=5\lambda = 5 (algebraic multiplicity 11). Eigenspace for λ=5\lambda=5: solve (A−5I)x=0(A-5I)x=0:

A−5I=(−11−120−211−3).A - 5I = \begin{pmatrix} -1 & 1 & -1 \\ 2 & 0 & -2 \\ 1 & 1 & -3 \end{pmatrix}.

Row reduce. R2→R2+2R1R_2 \to R_2 + 2R_1: (2,0,−2)+2(−1,1,−1)=(0,2,−4)(2,0,-2)+2(-1,1,-1) = (0,2,-4). R3→R3+R1R_3 \to R_3 + R_1: (1,1,−3)+(−1,1,−1)=(0,2,−4)(1,1,-3)+(-1,1,-1)=(0,2,-4).

(−11−102−402−4)→R3→R3−R2(−11−102−4000).\begin{pmatrix} -1 & 1 & -1 \\ 0 & 2 & -4 \\ 0 & 2 & -4 \end{pmatrix} \xrightarrow{R_3 \to R_3 - R_2} \begin{pmatrix} -1 & 1 & -1 \\ 0 & 2 & -4 \\ 0 & 0 & 0 \end{pmatrix}.

From row 2: 2y−4z=0⇒y=2z2y - 4z = 0 \Rightarrow y = 2z. From row 1: −x+y−z=0⇒x=y−z=2z−z=z-x+y-z=0 \Rightarrow x = y-z = 2z-z = z. So with z=tz=t free, (x,y,z)=(t,2t,t)=t(1,2,1)(x,y,z) = (t,2t,t) = t(1,2,1).

E(5,T)=span⁡{(1,2,1)},dim⁡E(5,T)=1.E(5,T) = \operatorname{span}\{(1,2,1)\}, \qquad \dim E(5,T) = 1.

Check: A(1,2,1)T=(4+2−1, 2+10−2, 1+2+2)T=(5,10,5)T=5(1,2,1)TA(1,2,1)^T = (4+2-1,\ 2+10-2,\ 1+2+2)^T = (5,10,5)^T = 5(1,2,1)^T. ✓ Eigenspace for λ=3\lambda=3: solve (A−3I)x=0(A-3I)x=0:

A−3I=(11−122−211−1).A-3I = \begin{pmatrix} 1 & 1 & -1 \\ 2 & 2 & -2 \\ 1 & 1 & -1 \end{pmatrix}.

All three rows are scalar multiples of (1,1,−1)(1,1,-1), so row reduction gives rank 11:

(11−1000000).\begin{pmatrix} 1 & 1 & -1 \\ 0 & 0 & 0 \\ 0 & 0 & 0 \end{pmatrix}.

The single equation is x+y−z=0x+y-z=0, i.e. z=x+yz = x+y, with x,yx,y free. So

E(3,T)={(x,y,x+y):x,y∈R}=span⁡{(1,0,1), (0,1,1)},dim⁡E(3,T)=2.E(3,T) = \{(x,y,x+y): x,y \in \mathbb{R}\} = \operatorname{span}\{(1,0,1),\,(0,1,1)\}, \qquad \dim E(3,T) = 2.

Check: A(1,0,1)T=(4−1,2−2,1+2)T=(3,0,3)T=3(1,0,1)TA(1,0,1)^T = (4-1, 2-2, 1+2)^T = (3,0,3)^T = 3(1,0,1)^T. ✓ And A(0,1,1)T=(1−1,5−2,1+2)T=(0,3,3)T=3(0,1,1)TA(0,1,1)^T = (1-1,5-2,1+2)^T=(0,3,3)^T = 3(0,1,1)^T. ✓ Here the geometric multiplicity of λ=3\lambda=3 (namely 22) equals its algebraic multiplicity (also 22), while for λ=5\lambda=5 both multiplicities equal 11. Since dim⁡E(3,T)+dim⁡E(5,T)=2+1=3=dim⁡R3\dim E(3,T)+\dim E(5,T) = 2+1=3=\dim \mathbb{R}^3, this operator is in fact diagonalizable — a preview of the Diagonalizability Criterion.

Example 2 (a 2×22\times 2 matrix with a deficient eigenspace, to see the contrast). Let

B=(51−13).B = \begin{pmatrix} 5 & 1 \\ -1 & 3 \end{pmatrix}.

Characteristic polynomial: det⁡(B−λI)=(5−λ)(3−λ)−(1)(−1)=15−8λ+λ2+1=λ2−8λ+16=(λ−4)2\det(B-\lambda I) = (5-\lambda)(3-\lambda) - (1)(-1) = 15 -8\lambda+\lambda^2+1 = \lambda^2-8\lambda+16 = (\lambda-4)^2. So λ=4\lambda = 4 is the only eigenvalue, with algebraic multiplicity 22. Eigenspace for λ=4\lambda=4: solve (B−4I)x=0(B-4I)x=0:

B−4I=(11−1−1).B-4I = \begin{pmatrix} 1 & 1 \\ -1 & -1 \end{pmatrix}.

Row 2 is −1-1 times row 1, so rank is 11: the single equation is x+y=0x+y=0, i.e. y=−xy=-x. So

E(4,B)=span⁡{(1,−1)},dim⁡E(4,B)=1.E(4,B) = \operatorname{span}\{(1,-1)\}, \qquad \dim E(4,B) = 1.

Check: B(1,−1)T=(5−1,−1−3)T=(4,−4)T=4(1,−1)TB(1,-1)^T = (5-1,-1-3)^T = (4,-4)^T = 4(1,-1)^T. ✓ Here the algebraic multiplicity of λ=4\lambda=4 is 22, but the geometric multiplicity dim⁡E(4,B)=1\dim E(4,B) = 1 is strictly smaller. There is no second independent eigenvector to pair with (1,−1)(1,-1): R2\mathbb{R}^2 cannot be spanned by eigenvectors of BB, so BB is not diagonalizable. This numerical gap between algebraic and geometric multiplicity is exactly the phenomenon studied in Geometric vs Algebraic Multiplicity.

CounterexampleWhy λ\lambda must actually be an eigenvalue for E(λ,T)E(\lambda,T) to be interesting

The Eigenspace Structure Theorem's part (a) is true for any scalar λ\lambda — E(λ,T)E(\lambda,T) is always a subspace, trivially {0}\{0\} if λ\lambda isn't an eigenvalue. But if you carelessly assume E(λ,T)E(\lambda,T) is nonzero-dimensional, or try to use it in a direct-sum decomposition, without first checking λ\lambda is actually an eigenvalue, you get nonsense. Concretely, take T:R2→R2T:\mathbb{R}^2\to\mathbb{R}^2 given by the rotation matrix

R=(0−110),R = \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix},

which rotates every vector by 90∘90^\circ. Try λ=2\lambda = 2. Then

R−2I=(−2−11−2),det⁡(R−2I)=4+1=5≠0,R - 2I = \begin{pmatrix} -2 & -1 \\ 1 & -2 \end{pmatrix}, \qquad \det(R-2I) = 4+1 = 5 \neq 0,

so R−2IR-2I is invertible and null⁡(R−2I)={0}\operatorname{null}(R-2I) = \{0\}. Indeed E(2,R)={0}E(2,R) = \{0\}: the "eigenspace" collapses to the trivial subspace because 22 is not an eigenvalue of RR (over R\mathbb{R}, RR has no real eigenvalues at all, since its characteristic polynomial is λ2+1\lambda^2+1, which has no real roots). The hypothesis being tested here is: "λ\lambda is an eigenvalue of TT" in the definition. Dropping the requirement that E(λ,T)E(\lambda,T) be nontrivial (i.e., forgetting to check λ\lambda is an actual eigenvalue before treating E(λ,T)E(\lambda,T) as "the eigenspace worth discussing") leads someone to falsely search for eigenvectors that provably do not exist over R\mathbb{R}. The subspace null⁡(T−λI)\operatorname{null}(T-\lambda I) is always well-defined, but it is only called the eigenspace in a meaningful sense — containing nonzero vectors — precisely when λ\lambda satisfies det⁡(T−λI)=0\det(T-\lambda I)=0.

Scope, edge cases & common traps

SituationWhat happens to E(λ,T)E(\lambda,T)
λ\lambda is an eigenvalueE(λ,T)≠{0}E(\lambda,T) \neq \{0\}; contains all eigenvectors for λ\lambda plus 00
λ\lambda is not an eigenvalueE(λ,T)={0}E(\lambda,T) = \{0\} (trivial subspace, no eigenvectors)
T=cIT = cI (scalar operator)E(c,T)=VE(c,T) = V, the whole space
Distinct eigenvalues λ1,…,λk\lambda_1,\dots,\lambda_kSum of eigenspaces is direct; nonzero eigenvectors are independent
dim⁡E(λ,T)<\dim E(\lambda,T) < algebraic multiplicity of λ\lambdaOperator fails to be diagonalizable (see Example 2)
Infinite-dimensional VV (e.g. differential operators)Definition E(λ,T)=null⁡(T−λI)E(\lambda,T)=\operatorname{null}(T-\lambda I) still applies verbatim

The mathematician behind it

Origins · 1 minDavid Hilbert and how this was discovered
DH
David Hilbert
1862–1943 · Göttingen, foundational work on spectral theory of operators

Hilbert did not invent the elementary notion of an eigenspace for matrices — that traces back to Cauchy's and Jacobi's 19th-century work on principal axes of quadratic forms — but Hilbert's development of spectral theory for operators on infinite-dimensional spaces (what are now called Hilbert spaces) gave the modern, operator-theoretic language in which "eigenspace as null space of a shifted operator" became the standard, unifying definition used across finite and infinite dimensions alike. His school at Göttingen turned eigenvalue problems from a computational tool in mechanics and geometry into a central organizing concept of analysis and algebra.

History
- 1829: Cauchy studies the principal axes of quadric surfaces, an early precursor to eigenvector computations.
- 1846: Jacobi develops systematic methods for diagonalizing symmetric matrices via what we now call eigenvectors.
- 1904–1910: Hilbert publishes his foundational papers on integral equations, introducing spectral theory for operators and the term "eigenwert" (eigenvalue).
- 1925: Heisenberg's matrix mechanics formulation of quantum theory makes eigenspaces the mathematical home of physical observables' possible measured values.
- 1930s: von Neumann rigorizes the spectral theorem for self-adjoint operators on Hilbert spaces, cementing eigenspaces as central objects in functional analysis.
Is the zero vector considered an eigenvector?

No. By convention, an eigenvector must be nonzero — otherwise every scalar λ\lambda would trivially satisfy T0=λ0=0T0=\lambda 0=0, making the notion of eigenvalue meaningless. However, the zero vector does belong to the eigenspace E(λ,T)E(\lambda,T), since eigenspaces must be subspaces and every subspace contains 00. So the eigenspace is the eigenvectors for λ\lambda together with the zero vector — this is precisely why we define E(λ,T)E(\lambda,T) as a null space rather than simply as "the set of eigenvectors," which would not be closed under addition or scalar multiplication (adding two eigenvectors could produce 00, which isn't allowed to be called an eigenvector itself).

How do I actually compute an eigenspace in practice?

First find the eigenvalues by solving det⁡(A−λI)=0\det(A-\lambda I)=0 for the characteristic polynomial. For each eigenvalue λ\lambda, form the matrix A−λIA-\lambda I and row-reduce it to find null⁡(A−λI)\operatorname{null}(A-\lambda I), exactly as you would for any homogeneous system in Kernel and null space. The free variables in the reduced row echelon form parametrize a basis for E(λ,T)E(\lambda,T). This is illustrated step-by-step in both worked examples above, where row reduction directly produces spanning vectors for each eigenspace.

Can an eigenspace have dimension greater than 1?

Yes, and this happens exactly when there is more than one linearly independent eigenvector for the same eigenvalue. In Example 1, E(3,T)E(3,T) has dimension 22, spanned by (1,0,1)(1,0,1) and (0,1,1)(0,1,1) — every nonzero linear combination of these is also an eigenvector for λ=3\lambda=3. Such higher-dimensional eigenspaces are common when the operator has extra symmetry, and their dimension is called the geometric multiplicity of the eigenvalue, a concept developed further in Geometric vs Algebraic Multiplicity.

What is the relationship between eigenspaces and diagonalizability?

An operator TT on a finite-dimensional space VV is diagonalizable if and only if the eigenspaces of TT together span all of VV, equivalently if and only if dim⁡V\dim V equals the sum of the dimensions of all the distinct eigenspaces. Since the Eigenspace Structure Theorem guarantees this sum is always direct, diagonalizability reduces to a dimension-counting question: do the eigenspaces, glued together via direct sum, fill up the whole space, or fall short (as in Example 2)? This equivalence is the content of the Diagonalizability Criterion.

Does the theory change for complex vector spaces versus real vector spaces?

The definition and the Eigenspace Structure Theorem hold verbatim over any field, including C\mathbb{C} and R\mathbb{R}. What changes is existence: over an algebraically closed field like C\mathbb{C}, the characteristic polynomial always factors completely, so every operator has at least one eigenvalue and hence at least one nontrivial eigenspace. Over R\mathbb{R}, as the counterexample with the rotation matrix shows, an operator can have no real eigenvalues at all, meaning every E(λ,T)E(\lambda,T) for real λ\lambda is trivial.

Is E(λ,T)E(\lambda,T) the same as the kernel of TT?

Only when λ=0\lambda = 0. By definition E(0,T)=null⁡(T−0⋅I)=null⁡(T)E(0,T) = \operatorname{null}(T-0\cdot I) = \operatorname{null}(T), so the eigenspace for eigenvalue 00 is exactly the ordinary kernel of TT. For any other λ\lambda, E(λ,T)E(\lambda,T) is the kernel of the shifted operator T−λIT-\lambda I, not of TT itself. This is a useful special case to remember: an operator has 00 as an eigenvalue precisely when it is not injective, i.e., precisely when it has nontrivial kernel.

Why do eigenspaces for distinct eigenvalues never overlap except at zero?

This is a consequence of the direct-sum part of the Eigenspace Structure Theorem, proved by an induction argument using distinctness of the λi\lambda_i's. Intuitively, if a nonzero vector vv were in both E(λ1,T)E(\lambda_1,T) and E(λ2,T)E(\lambda_2,T) with λ1≠λ2\lambda_1\neq\lambda_2, then Tv=λ1vTv=\lambda_1 v and Tv=λ2vTv=\lambda_2v simultaneously, forcing λ1v=λ2v\lambda_1v=\lambda_2v, i.e. (λ1−λ2)v=0(\lambda_1-\lambda_2)v=0. Since λ1≠λ2\lambda_1\neq\lambda_2, this forces v=0v=0, contradicting v≠0v\neq 0. This is exactly the k=2k=2 case of the general theorem.

Before this

Where this leads