Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

I'm having trouble with some maths regarding the expression of the matrix quadratic form (i.e. $x^TAx$) and the proof that, where the eigenvalues $\lambda_1,\lambda_2,...,\lambda_n$ are all positive, the quadratic form is positive definite.

My understanding is that the definition of positive-definiteness is when $x^TAx>0$ for all x where at least one element of $x \neq 0$.

My textbook produces the following proof, but I don't understand the last line:

Where $s_i$ = a normalized eigenvector of A, $\lambda_i$ = the corresponding eigenvalue, and $C = [s_1|s_2|...|s_n]$, $C^{-1}AC=D=$ the "diagonalization" of A.

Since the eigenvectors of A are orthonormal, $C^{-1} = C^T$.

Suppose we define a transformation $x=CX$. Then the equation becomes: $x^TAx = (CX)^{T}ACX = X^TDX = \lambda_1X_1^2 + \lambda_2X_2^2+...+\lambda_nX_n^2$.

It follows from this that a quadratic form is positive-definite if and only if all its eigenvalues are positive.

So, in summary, I don't understand why the following derivation true:

$x^TAx = \lambda_1X_1^2 + \lambda_2X_2^2+...+\lambda_nX_n^2$ therefore, for $x^TAx > 0$ for all x where at least one element of $x_i \neq 0$ to be true, $\lambda_1, \lambda_2, ..., \lambda_n > 0$.

Can someone please help with the derivation of the last step?a

Thanks in advance for your help.

share|improve this question
I think your definition of positive definiteness may be the source of your confusion. The correct definition is that $A$ is positive definite if $x^TAx>0$ for all vectors $x$ other than the zero vector. –  Will Orrick Apr 30 '12 at 2:20
@WillOrrick that is what I mean by at least one element of x is not zero. To restate: for all x except x = [0;0]. Thanks for you comment though. –  JonaGik Apr 30 '12 at 2:25
"...therefore, for $x^TAx\gt0$ with at least one element of $x_i\ne0$ to be true, $λ_1$, $λ_2$,...,$λ_n\gt0$." --- I still think you really need to change this to "...for $x^TAx\gt0$ to be true for all $x$ having at least one element $x_i\ne0$..." The "for all $x$" allows you to choose an $x$ that will help you reach the conclusion. Since $x^TAx\gt0$ has to be true for any $x$, you can choose $x$ in such a way that only one of the eigenvalues contributes. Then you can conclude that that eigenvalue must be positive. Repeat for every eigenvalue. –  Will Orrick Apr 30 '12 at 4:10
Thanks for adding that. I think you also need to correct it in the second gray box. This may help clarify the logic of the proof. –  Will Orrick Apr 30 '12 at 5:29
@WillOrrick no problem and fixed in the second gray box too :) –  JonaGik May 1 '12 at 1:08

3 Answers 3

For the "if" part: if all of the $\lambda_j$ are positive then $\lambda_1X_1^2+\ldots+\lambda_nX_n^2$ is positive for all non-zero vectors $X$ and hence for all non-zero vectors $x$.

For the "only if" part: Suppose that $\lambda_1X_1^2+\ldots+\lambda_nX_n^2$ is positive for all non-zero vectors $x$. Then in particular, it is positive for the vector $x$ corresponding to the $X$ with $X_j=1$ and $X_k=0$ for all $k\ne j$. Therefore $\lambda_j$ is positive.

share|improve this answer
Thanks for your response. I don't understand how the latter is an "only if" proof - it seems to only prove that the expression is positive in the particular (not general) case where X is any vector with a single element equal to one and the others equal to zero. Consider the case where $X = [1, 1, 0]$ (which appears to be legal). Now we have $\lambda_1 + \lambda_2 > 0$ which isn't constrained to all $\lambda_i > 0$. Presumably this is a valid counter-example? –  JonaGik Apr 30 '12 at 5:43
The eigenvalues are fixed numbers, depending only on $A$. They are what they are. One thing that we know about them is that $\lambda_1X_1^2+\ldots+\lambda_nX_n^2$ is positive for all $X$ except $X=0$. What conclusions about the values of $\lambda_j$ can we draw from this fact? Both your conclusion, $\lambda_1+\lambda_2\gt0$, and my conclusion, $\lambda_1\gt0$, are valid inferences. Mine just happens to be a stronger conclusion. For every non-zero choice of $X$ we get a conclusion. All of these conclusions are true statements about the $\lambda_j$. –  Will Orrick Apr 30 '12 at 7:23
My understanding is that we want to classify A (as positive-definite, negative-definite, etc) on the basis of its eigenvalues. In that case, we don't know that $\lambda_1X_1^2+\ldots+\lambda_nX_n^2$ is positive. We wish to determine this based on its eigenvalues. What you have given is a case where all the eigenvalues are positive and the matrix is positive definite, but not a general, proven rule for this. –  JonaGik Apr 30 '12 at 9:12
@JonaGik The goal stated in your original post was to prove that a quadratic form is positive-definite if and only if all its eigenvalues are positive. In a previous comment, you questioned the proof of the "only if" part: a quadratic form is positive-definite only if all its eigenvalues are positive. This part is logically equivalent to the statement that if a quadratic form is positive-definite, then all of its eigenvalues are positive. In order to prove this, we take a positive-definite form. By definition of positive-definite, $x^TAx\gt0$ for all non-zero $x$. The manipulations... –  Will Orrick Apr 30 '12 at 12:40
(continued) ...in the proof quoted in your post establish that $\lambda_1X_1^2+\ldots+\lambda_nX_n^2\gt0$ for all non-zero $X$. This is not an assumption; it follows from the definition of positive-definiteness. We may then choose particular values for the $X_j$ in order to draw inferences about the $\lambda_j$. –  Will Orrick Apr 30 '12 at 12:44

I misread the question. @WillOrrick's comment is relevant, if $A$ is positive definite, then you must have $x^T Ax > 0$ for any $x \neq 0$ (which is equivalent to at least one $x_i \neq 0$).

To illustrate why the 'for any' is important, consider the matrix $B = \left( \begin{array}{cc} 1 & 0 \\ 0 & -1 \end{array} \right)$. Then $e_1^T B e_1 > 0$, but $B$ is not positive definite. The 'for any $x \neq 0$' is important.

Since you have $x^TAx = \sum_{i=1}^n \lambda_i X_i^2$ (with $x = CX$), we may choose $x=C e_k$, where $e_k$ is the $k$th basis vector (ie, $x$ is an eigenvector corresponding to $\lambda_k$). It follows that for this choice of $x$, we have $x^TAx = \sum_{i=1}^n \lambda_i X_i = \lambda_k X_k = \lambda_k$. Since $A$ is positive definite, it follows that $\lambda_k>0$. Since $k$ was arbitrary, we have that all eigenvalues are strictly positive.

share|improve this answer
Thanks very much for your response. This seems like its definitely on the right track. For $\sum_{i-1}^n X_i^2 = \underline \lambda \sum_{i-1}^n x_i^2$, I'm having trouble proving that to myself. Do you think you could prove it? Thanks again! –  JonaGik Apr 30 '12 at 3:01
As far as I can tell, your proof assumes that A is a diagonal matrix so that the eigenvectors are [1, 0, ..., 0], [0, 1, 0, ..., 0], ..., [0, ..., 0, 1] and thus $x^TAx = \sum_{i=1}^n \lambda_i X_i = \lambda_k X_k = \lambda_k$ is true. It is possible for more than one $\lambda_iX_i^2$ combination to be non-zero in the general case. Is this right? If so, this doesn't produce a general proof. Also, does this proof prove only that if all eigenvectors are positive, the matrix is positive-definite, not if and only if all eigenvectors are positive the matrix is positive-definite? –  JonaGik Apr 30 '12 at 4:01
Since $C$ is a basis of orthonormal eigenvectors, the matrix $C^T A C $ is diagonal. Note the distinction between $x$ and $X$ in the proof. We have $x^T Ax = (CX)^T A (CX) = X^T (C^T A C) X$. –  copper.hat Apr 30 '12 at 4:15
Your question asked for elaboration of the last step, not an 'iff' proof? The elaboration shows that if $A$ is positive definite, then you must have $\lambda_i > 0$. –  copper.hat Apr 30 '12 at 4:16
Are you saying that $e_k$ is some arbitrary eigenvector of $D = C^TAC$? Because then I can see this working. However, overall, the proof seems to still only prove that if (but not iff) all the eigenvectors are positive, the matrix is positive-definite. Is this right? –  JonaGik Apr 30 '12 at 5:11

The "only if" part (if $A$ is pos.def, then all eigenvalues are positive) can be proved more easily by absurd. Suppose that $x^t A x>0$ for all $x\ne 0$, but we have some $\lambda_i <0 $. The, calling $p_i$ ($\ne 0$) the associated eigenvector, we have $A p_i = \lambda_i p_i$ and multipliyin by $p_i^t$ we get $ p_i^t A p_i = \lambda_i |p_i|^2 <0 $ which contradicts the initial assumption.

share|improve this answer
Unless I'm missing something, this only proves that this is inconsistent when x is the negative eigenvalue's eigenvector - it isn't a proof for the general case. –  JonaGik May 8 '12 at 9:31
No, it's general. Given the assumption ("Suppose that... for all $x$") we find one $x$ that contradicts that assumption. That's enough to prove that the assumption is false. –  leonbloy May 8 '12 at 13:44

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.