# How to prove that eigenvectors from different eigenvalues are linearly independent [duplicate]

How can I prove that if I have $n$ eigenvectors from different eigenvalues, they are all linearly independent?

• This is equivalent to showing that a set of eigenspaces for distinct eigenvalues always form a direct sum of subspaces (inside the containing space). That is a question that has been asked many times on this site. I will therefore close this question as duplicate of one of them (which is marginally more recent than this one, but that seems hardly an issue after more than a decade). May 18, 2022 at 10:09

I'll do it with two vectors. I'll leave it to you do it in general.

Suppose $\mathbf{v}_1$ and $\mathbf{v}_2$ correspond to distinct eigenvalues $\lambda_1$ and $\lambda_2$, respectively.

Take a linear combination that is equal to $0$, $\alpha_1\mathbf{v}_1+\alpha_2\mathbf{v}_2 = \mathbf{0}$. We need to show that $\alpha_1=\alpha_2=0$.

Applying $T$ to both sides, we get $$\mathbf{0} = T(\mathbf{0}) = T(\alpha_1\mathbf{v}_1+\alpha_2\mathbf{v}_2) = \alpha_1\lambda_1\mathbf{v}_1 + \alpha_2\lambda_2\mathbf{v}_2.$$ Now, instead, multiply the original equation by $\lambda_1$: $$\mathbf{0} = \lambda_1\alpha_1\mathbf{v}_1 + \lambda_1\alpha_2\mathbf{v}_2.$$ Now take the two equations, \begin{align*} \mathbf{0} &= \alpha_1\lambda_1\mathbf{v}_1 + \alpha_2\lambda_2\mathbf{v}_2\\ \mathbf{0} &= \alpha_1\lambda_1\mathbf{v}_1 + \alpha_2\lambda_1\mathbf{v}_2 \end{align*} and taking the difference, we get: $$\mathbf{0} = 0\mathbf{v}_1 + \alpha_2(\lambda_2-\lambda_1)\mathbf{v}_2 = \alpha_2(\lambda_2-\lambda_1)\mathbf{v}_2.$$

Since $\lambda_2-\lambda_1\neq 0$, and since $\mathbf{v}_2\neq\mathbf{0}$ (because $\mathbf{v}_2$ is an eigenvector), then $\alpha_2=0$. Using this on the original linear combination $\mathbf{0} = \alpha_1\mathbf{v}_1 + \alpha_2\mathbf{v}_2$, we conclude that $\alpha_1=0$ as well (since $\mathbf{v}_1\neq\mathbf{0}$).

So $\mathbf{v}_1$ and $\mathbf{v}_2$ are linearly independent.

Now try using induction on $n$ for the general case.

• I believe that you wrote $\lambda_2$ instead of $\lambda_1$ in the row before "Now take" Mar 14, 2014 at 8:08
• Is there any intuition behind this? Any pictorial way of thinking? May 7, 2016 at 8:42
• I don't understand how this generalises to N eigenvectors. Jan 29, 2020 at 23:57
• @fielder: You don’t “generalize”; you use induction. As the answer states. Jan 30, 2020 at 3:47
• @NathanielRuiz: $v_1$ is assumed to be an eigenvector of $T$ corresponding to the eigenvalue $\lambda_1$; $v_2$ is assumed to be an eigenvector of $T$ corresponding to the eigenvalue $\lambda_2$. Do you not know the definition of "eigenvalue" and "eigenvector"? And since the $\alpha_i$ are scalars, $T(\alpha_1v_1+\alpha_2v_2) = \alpha_1T(v_1) + \alpha_2T(v_2)$. Jan 25, 2021 at 19:05

Alternative:

Let $j$ be the maximal $j$ such that $v_1,\dots,v_j$ are independent. Then there exists $c_i$, $1\leq i\leq j$ so that $\sum_{i=1}^j c_iv_i=v_{j+1}$. But by applying $T$ we also have that

$$\sum_{i=1}^j c_i\lambda_iv_i=\lambda_{j+1}v_{j+1}=\lambda_{j+1}\sum_{i=1}^j c_i v_i$$ Hence $$\sum_{i=1}^j \left(\lambda_i-\lambda_{j+1}\right) c_iv_i=0$$ which is a contradiction since $\lambda_i\neq \lambda_{j+1}$ for $1\leq i\leq j$.

Hope that helps,

• P.S. The argument uses the well-ordering principle on the naturals (by looking at the least $j$ such that $v_1,\ldots,v_{j+1}$ is dependent). Well-ordering for the naturals is equivalent to induction. Mar 28, 2011 at 1:20
• @Arturo: Thanks! Should be fixed now. (Didn't realize I was using well ordering in that way) Mar 28, 2011 at 1:30
• No problem; I'll delete the other two comments since it's been fixed. In any case, many people prefer arguments along these lines to an explicit induction, even if they are logically equivalent (you can think of the argument you give as a proof by contradiction of the inductive step, with the base being taken for granted [or as trivial, since $v_1$ is nonzero]; cast that way, it may be clearer why the two arguments are closely connected). Mar 28, 2011 at 1:34
• @Eric Naslund This proofs seems very similar to the one given in Axler...
– user38268
Aug 31, 2011 at 7:29
• @Eric Naslund Sheldon Axler's Linear Algebra Done Right.
– user38268
Sep 3, 2011 at 21:52

Hey I think there's a slick way to do this without induction. Suppose that $T$ is a linear transformation of a vector space $V$ and that $v_1,\ldots,v_n \in V$ are eigenvectors of $T$ with corresponding eigenvalues $\lambda_1,\ldots,\lambda_n \in F$ ($F$ the field of scalars). We want to show that, if $\sum_{i=1}^n c_i v_i = 0$, where the coefficients $c_i$ are in $F$, then necessarily each $c_i$ is zero.

For simplicity, I will just explain why $c_1 = 0$. Consider the polynomial $p_1(x) \in F[x]$ given as $p_1(x) = (x-\lambda_2) \cdots (x-\lambda_n)$. Note that the $x-\lambda_1$ term is "missing" here. Now, since each $v_i$ is an eigenvector of $T$, we have \begin{align*} p_1(T) v_i = p_1(\lambda_i) v_i && \text{ where} && p_1(\lambda_i) = \begin{cases} 0 & \text{ if } i \neq 1 \\ p_1(\lambda_1) \neq 0 & \text{ if } i = 1 \end{cases}. \end{align*}

Thus, applying $p_1(T)$ to the sum $\sum_{i=1}^n c_i v_i = 0$, we get $$p_1(\lambda_1) c_1 v_1 = 0$$ which implies $c_1 = 0$, since $p_1(\lambda_1) \neq 0$ and $v_1 \neq 0$.

• You have a typo in your answer and I can't edit it. $p_1(x) is missing a space after the dollar sign. Apr 30, 2019 at 8:01 • Could not understand what is x in p_1(x) as x-\lambda suggests that it should be scalar but p_1(T) suggests as if it is a transformation matrix. Please clarify. Mar 28, 2020 at 12:33 • @user3779172 Given a polynomial$p(x)$and a matrix$T$, one writes$p(T)$for the matrix obtained by replacing all the$x$s by$T$s. Note$x^0$becomes$T^0$which is to be interpreted as the identity matrix. For example, if$p(x) = 5x^2+3$and$T = \begin{bmatrix}1 & 1 \\ 0 & 1 \end{bmatrix}$, then$p(T) = 5\begin{bmatrix}1 & 1 \\ 0 & 1 \end{bmatrix}^2 +3 \begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix}$. If you know some abstract algebra, then you can fit this into a more general story about polynomial rings and evalation homomorphisms. planetmath.org/evaluationhomomorphism Mar 28, 2020 at 18:30 For eigenvectors$\vec{v^1},\vec{v^2},\dots,\vec{v^n}$with different eigenvalues$\lambda_1\neq\lambda_2\neq \dots \neq\lambda_n$of a$ n\times n$matrix$A$. Given the$ n\times n$matrix$P$of the eigenvectors (with eigenvectors as the columns). $$P=\Big[\vec{v^1},\vec{v^2},\dots,\vec{v^n}\Big]$$ Given the$ n\times n$matrix$\Lambda$of the eigenvalues on the diagonal (zeros elsewhere): $$\Lambda = \begin{bmatrix} \lambda_1 & 0 & \dots & 0 \\ 0 & \lambda_2 & \dots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & \dots & \lambda_n \end{bmatrix}$$ Let$\vec{c}=(c_1,c_2,\dots,c_n)^T$We need to show that only$c_1=c_2=...=c_n=0$can satisfy the following: $$c_1\vec{v^1}+c_2\vec{v^2}+...= \vec{0^{}}$$ Applying the matrix to this equation gives: $$c_1\lambda_1\vec{v^1}+c_2\lambda_2\vec{v^2}+...+c_n\lambda_n\vec{v^n}= \vec{0^{}}$$ We can write this equation in the form of vectors and matrices: $$P\Lambda \vec{c^{}}=\vec{0^{}}$$ But with since$A$can be diagonalised to$\Lambda$, we know$P\Lambda=AP$$$\implies AP\vec{c^{}}=\vec{0^{}}$$ since$AP\neq 0$, we have$\vec{c}=0$. • Let me add a comment to the last. In order to have$c=0$the only solution,$AP$must be invertible but since$A$is already invertible because$\lambda_i \neq \lambda_j \forall i\neq j$, the only demand is$P$to be invertible$\Rightarrow$independent eigenvectors. Sep 13, 2016 at 16:36 • why is$P \ne A$? Jan 30, 2020 at 0:06 • I think we cannot be sure of that AP is invertible as we don't know if P has linearly independent columns Mar 28, 2020 at 12:23 Suppose $$v_k$$ are eigenvectors of $$A$$, that is, $$\sum_k c_k v_k=0$$ with $$v_k\neq0$$, and $$Av_k=\lambda_k v_k$$. We only include in this sum eigenvectors corresponding to distinct eigenvalues, so that $$\lambda_i\neq\lambda_j$$ for $$i\neq j$$. We want to prove that this implies $$c_k=0$$. Define the operators $$A^{(i)}\equiv \prod_{k\neq i}(A-\lambda_k I).$$ In particular, note that $$A^{(i)}v_k=\delta_{ik} d_i$$, where $$d_i\equiv \prod_{k\neq i}(\lambda_i-\lambda_k)\neq0$$, and $$\delta_{ik}$$ is the standard Dirac delta function. Then $$0 = A^{(i)}(0) = A^{(i)}\left(\sum_k c_k v_k\right) =d_i c_i v_i.$$ It follows that $$c_i=0$$ for all $$i$$ (because, again, $$d_i\neq0$$ and $$v_i\neq0$$), and thus the vectors $$\{v_i\}_i$$ are linearly independent. • This is a good argument that does not explicitly use induction. The punchline could have been delivered a bit better though; I would have mentioned explicitly the assumption$v_i\neq0$(because$v_i$is an eigenvector), and started after the display by "Then$0=A^{(i)}(0)=A^{(i)}(\sum\ldots)$" so that it is clear that from$0=d_ic_iv_i$one can indeed deduce$c_i=0$. It may be worth noting that, restricted to the sum of eigenspaces in question, the operator$A^{i)}/d_i$is the projection on the eigenspace for$\lambda_i\$, which projections can only exist if the sum is direct. Jul 3, 2021 at 7:03
• @MarcvanLeeuwen thank you for the suggestions. I hopefully made the answer clearer
– glS
Jul 15, 2021 at 11:22

Let $$A$$ be a $$n \times n$$ matrix with pairwaise different eigenvalues $$\lambda_1, \ldots, \lambda_n$$ and their eigenvectors $$v_1,\ldots,v_n$$. Put $$P=[v_1\;\cdots\;v_n],\qquad \Lambda=\left[\begin{array}{ccc} \lambda_1 & & \\ & \ddots & \\ & & \lambda_n\end{array}\right],\qquad V=\left[\begin{array}{cccc} 1 & \lambda_1 & \cdots & \lambda_1^{n-1} \\ \vdots & \vdots & \ddots &\vdots \\ 1 & \lambda_n & \cdots & \lambda_n^{n-1}\end{array}\right]$$ $$C=\left[\begin{array}{c} c_1 \\ \vdots \\ c_n\end{array}\right],\qquad C'=\left[\begin{array}{ccc} c_1 & & \\ & \ddots & \\ & & c_n\end{array}\right]$$

Suppose $$PC=c_1v_1+\ldots+c_nv_n=O_{n,1}$$. Our aim is to show that $$PC'=O_{n,n}$$ (equivalently $$c_iv_i=O_{n,1}$$ for $$i=1,\ldots,n$$, which implies $$c_i=0$$, because $$v_i\neq O_{n,1}$$ for all $$i$$).

It is clear that $$AP=[Av_1\;\cdots\;Av_n] = [\lambda_1v_1\;\cdots\;\lambda_nv_n] = P\Lambda$$. Then $$O_{n,1} = AO_{n,1} = A(PC) = (AP)C = (P\Lambda)C = P(\Lambda C),$$ so $$P(\Lambda C)=O_{n,1}$$ and inductively $$P(\Lambda^kC)=O_{n,1}$$ for $$k=0,1,2,\ldots$$.

Matrix $$V$$ is Vandermonde matrix and it is invertible because $$\det V = \prod_{1\le i < j \le n}(\lambda_j-\lambda_i)\neq0$$.

We have $$PC' = PC'(VV^{-1}) = P(C'V)V^{-1} = P\left[\begin{array}{cccc} c_1 & \lambda_1c_1 & \cdots & \lambda_1^{n-1}c_1 \\ \vdots & \vdots & \ddots &\vdots \\ c_n & \lambda_nc_n & \cdots & \lambda_n^{n-1}c_n \end{array}\right]V^{-1} \\ = P[C\;\;\Lambda C\;\; \cdots\;\; \Lambda^{n-1}C]V^{-1} = [PC\;\;P\Lambda C\;\; \cdots\;\; P\Lambda^{n-1}C]V^{-1} = O_{n,n}V^{-1} = O_{n,n},$$ completing the proof.

I would like to add a bit of intuition.

Suppose v1, v2, w are eigenvectors of M with different eigenvalue. and

w = ⍺1v1 + ɑ2v2

w as composition of v1 and v2

Think of M as a transformation, for w to come out in the same direction, each of its components needs to be scaled proportionally (by the same amount). That is, v1 and v2 need to have identical eigenvalue, which contradicts with the assumption.

Consider the matrix $$A$$ with two distinct eigen values $$\lambda_1$$ and $$\lambda_2$$. First note that the eigen vectors cannot be same , i.e., If $$Ax = \lambda_1x$$ and $$Ax = \lambda_2x$$ for some non-zero vector $$x$$. Well , then it means $$(\lambda_1- \lambda_2)x=\bf{0}$$. Since $$\lambda_1$$ and $$\lambda_2$$ are scalars , this can only happen iff $$x = \bf{0}$$ (which is trivial) or when $$\lambda_1 =\lambda_2$$ .

Thus now we can safely assume that given two eigenvalue , eigenvector pair say $$(\lambda_1, {\bf x_1})$$ and $$(\lambda_2 , {\bf x_2})$$ there cannot exist another pair $$(\lambda_3 , {\bf x_3})$$ such that $${\bf x_3} = k{\bf x_1}$$ or $${\bf x_3} = k{\bf x_2}$$ for any scalar $$k$$. Now let $${\bf x_3} = k_1{\bf x_1}+k_2{\bf x_2}$$ for some scalars $$k_1,k_2$$

Now, $$A{\bf x_3}=\lambda_3{\bf x_3} \\$$ $${\bf x_3} = k_1{\bf x_1} + k_2{\bf x_2} \:\:\: \dots(1)\\$$ $$\Rightarrow A{\bf x_3}=\lambda_3k_1{\bf x_1} + \lambda_3k_2{\bf x_2}\\$$ but, $$\:\: {\bf x_1}=\frac{1}{\lambda_1}Ax_1$$ and $${\bf x_2}=\frac{1}{\lambda_2}Ax_2$$. Substituting in above equation we get, $$A{\bf x_3}=\frac{\lambda_3k_1}{\lambda_1}A{\bf x_1} + \frac{\lambda_3k_2}{\lambda_2}A{\bf x_2} \\$$

$$\Rightarrow {\bf x_3}=\frac{\lambda_3k_1}{\lambda_1}{\bf x_1} + \frac{\lambda_3k_2}{\lambda_2}{\bf x_2} \:\:\: \dots (2)$$

From equation $$(1)$$ and $$(2)$$ if we compare coefficients we get $$\lambda_3 = \lambda_1$$ and $$\lambda_3 = \lambda_2$$ , which implies $$\lambda_1 = \lambda_2 = \lambda_3$$ but according to our assumption they were all distinct!!! (Contradiction)

NOTE: This argument generalizes in the exact same fashion for any number of eigenvectors. Also, it is clear that if $${\bf x_3}$$ cannot be a linear combination of two vectors then it cannot be a linear combination of any $$n >2$$ vectors (Try to prove!!!)