Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

The Lie derivative of a function $V:\mathbb{R}^n\to\mathbb{R}$ along the trajectories of the dynamical system $\dot{x}=f(x)$, with $x\in\mathbb{R}^n$ and $f:\mathbb{R}^n\to\mathbb{R}^n$, is defined as [H.K.Khalil, Nonlinear Systems, Prentice Hall editions, 3rd edition, ISBN: 0-13-067389-7, London 2002, p.114,509-510]: $$ \dot{V}(x)=\frac{\partial V}{\partial x}\cdot f(x) $$ Using this formula and assuming that $V(x)=\frac{1}{2}x'Px$ (with $P$ symmetric) and $f(x)=Ax$ we have: $$ \dot{V}(x)=\frac{\partial V}{\partial x}\cdot f(x)=x'PAx $$ However, this is wrong. The correct answer follows this derivation: $$ \dot{V}(x)=\frac{d}{dt}V(x)=\frac{d}{dt}(x'Px)=\frac{1}{2}(\dot{x}'Px+x'P\dot{x}) $$ we now substitute $\dot{x}=Ax$ to get: $$ \dot{V}(x)=\frac{1}{2}x'(A'P+PA)x $$

So I have a series of questions:

  1. What is wrong with the first approach?
  2. Why don't we denote $\frac{\partial V}{\partial x}$ simply as $\frac{dV}{dx}$ since $V$ is a function of a single variable - namely $x$.
  3. What is the utility of the first formula? an example?
share|cite|improve this question
up vote 4 down vote accepted
  1. There is nothing wrong, since the two answers are equal (note that $x'A'Px$ is a scalar, so it equals its own transpose $(x'A'Px)'=x'P'A''x''=x'PAx$).

  2. The variable $x$ is a vector $(x_1,\dots,x_n)$, and $\partial V/\partial x$ is the vector of partial derivatives (i.e., the gradient of $V$), which I guess is the reason for the curved d's. But this is just one of many notations for the gradient. Pick your favourite. ;-)

  3. Which formula? The definition of $\dot V$? Well, for example, $\dot V=0$ iff $V$ is a constant of motion. And you might want to read about Liapunov functions too.

share|cite|improve this answer
Thanks a lot for the answer. If I may ask one more thing... The Lyapunov equation is known as $A'P+PA=-Q$ where $Q$ is a positive (semi)definite matrix. Why don't we use $PA=-Q$ ? (I understand that $A'P+PA\neq 2PA$ although $x'(A'P+PA)x=2x'PAx$). – Pantelis Sopasakis Nov 3 '11 at 8:53
@Pantelis: The notion of "positive definite" really only makes sense for symmetric matrices (or Hermitian matrices if you deal with complex numbers), since it's really about the quadratic form $x'Qx$ rather than just the matrix $Q$ itself. So if you want $Q$ to be symmetric, you can't take just $PA=-Q$, since $PA$ need not be symmetric. – Hans Lundmark Nov 3 '11 at 9:37

Indeed they are both correct, even if $P$ is not symmetric. If $x$ is a $n\times 1$ column vector (which, using your algebra, it is), then the gradient of $V$ i.e. $\frac{\partial V}{\partial x}$ is the derivative of a quadratic form,

$$\frac{\partial V }{\partial x}=\frac{1}{2}\frac{\partial xPx}{\partial x}=\frac{1}{2}(P+P^{T})x$$

In the special case that $P$ is symmetric, obviously you get,

$$\frac{\partial V }{\partial x}=Px$$

Observe that $Px$ (in matrix notation) is a $n\times 1$ column vector. If $P$ is not symmetric, $(P+P^{T})$ is indeed so. Now, the Lie derivative of the function $V$ with respect to the vector field $f$, is given by,

$$\dot V(x)=\frac{\partial V }{\partial x} \cdot f(x)=\frac{1}{2}[(P+P^{T})x]^{T} Ax = \frac{1}{2}x^{T}(P+P^{T})Ax$$

Apparently we have tranposed $\frac{\partial V}{\partial x}$ in the above equation because we used matrix notation and the dotting between vectors is equivalent to matrix multiplication, with one vector transposed. Notice for the transposition that $P+P^{T}$ is a symmetric matrix hence $P+P^{T}=(P+P^{T})^{T}$. Now, if $P$ is symmetric the Lie derivative is,

$$\dot V(x)=x^{T}PAx$$

Using your second derivation, the Lie derivative is,

$$\dot V(x)= \frac{1}{2}x^{T}(A^{T}P+PA)x$$

which also works for non-symmetric matrices. You want to prove that they are equal. Take the difference of the two results, i.e.

$$\Delta V(x)= \frac{1}{2}x^{T}(P+P^{T})Ax - \frac{1}{2}x^{T}(A^{T}P+PA)x = \frac{1}{2}x^{T}[PA+P^{T}A-A^{T}P-PA]x =$$ $$ =\frac{1}{2}x^{T}[P^{T}A-A^{T}P ]x$$

For $x=0 \rightarrow \Delta V=0$. Taking the derivative w.r.t $x$ (notice that $\Delta V$ is a quadratic form),

$$\frac{\partial \Delta V(x)}{\partial x}=\frac{1}{2}([P^{T}A-A^{T}P]+[P^{T}A-A^{T}P]|^{T})x$$

some algebra,

$$P^{T}A-A^{T}P+[P^{T}A-A^{T}P]|^{T}=P^{T}A-A^{T}P+[(P^{T}A)^{T}-(A^{T}P)^{T}]=$$ $$=P^{T}A-A^{T}P+[A^{T}T-P^{T}A]=0$$

Thus $\Delta V(x)= const=0$.

Also, regarding the dels in the derivative $\frac{\partial V(x)}{\partial x}$, Hans is also correct, but in the general case the Lie derivative is also defined for two vector fields, which in this case $\frac{\partial V(x)}{\partial x}$ is the Jacobian matrix of the vector field $V$.

share|cite|improve this answer

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.