I am trying to understand the balanced truncation algorithm and have some trouble distinguishing between controllability matrix and controllability Gramian. If my understanding is correct, a linear time invariant system $\dot x(t) = Ax(t) + Bu$ is controllable if the controllability matrix

$$P = \begin{bmatrix} B & AB & A^2B & \dots & A^{n-1}B\end{bmatrix}$$

has full row rank. Then we have the controllability Gramian

$$\textit{P} = \int^{t_1}_{t_0} e^{At}BB^Te^{A^Tt} \, \mathrm d t$$

If the controllability matrix $P$ does not satisfy the full row rank requirement and the system is not controllable, is the Gramian still positive definite, nonsingular, hermitian and all that? In other words, can I proceed to the Cholesky decomposition with a non-controllable system's controllability Gramian?

Any literature stating this would also be much appreciated. Thank you.


1 Answer 1


In general, the controllability Gramian is given as the unique solution to the continuous time Lyapunov matrix equation $$ AP + PA^T = BB^T.$$ (In the context of model reduction, $A$ is typically an asymptotically stable matrix, so this equation has a unique solution). The solution can be written as $$ P = \int_0^\infty e^{At} BB^T e^{A^T t} dt $$ from which it follows that $P$ is (always) symmetric positive semi-definite and that null space of $P$ is identical to the null space of the controllability matrix $K$ given by $$ K = \begin{bmatrix} B & AB & A^2 B & \dotsc & A^{n-1} B \end{bmatrix}, \quad \text{K for Krylov} $$ In particular, $P$ is symmetric positive definite if and only if $K$ has full rank. In principle, you can test if the system is controllable by attempting a Cholesky factorization of $P$. This will work nicely for small toy systems.

However, in practice, even when the system is controllable, the eigenvalues of $P$ decay so rapidly, that the matrix is singular to machine precision. In particular, a Cholesky factorization will fail and you will be left with the impression that the system is uncontrollable. Typically, an exceedingly good approximation $P$ of very low rank exists and can frequently be found without computing and truncating the full eigen-decomposition of $P$. This is the so-called low rank phenomenon for Lyapunov matrix equations. It is foundation upon which modern iterative solvers of Lyapunov equations are built and it is closely related to the success of balanced truncation.

As for literature you must obtain a copy of A. C. Antoulas Approximation of Large Scale Dynamical Systems, see


for the details. K. Zhou's book "Robust and Optimal Control" may be of use to you as well, but I think it best to start with Antoulas's book.

EDIT: If we wish to determine if the system is controllable, then we need to investigate the controllability matrix $K$ in an numerically reliable manner. I am not sure of the details for the general case where $B$ has $k>1$ columns, but if $B = b \in \mathbb{R}^n$, then the Arnoldi algorithm can be applied. Given $(A,b)$ this algorithm attempts to build an orthonormal basis for the Krylov subspace $$K(A,b) = \text{span} \{ A^j b \: : \: j = 1,2,\dotsc \}. $$ Now, in exact arithmetic, if the algorithm completes $k$ iterations, then it produces a factorization of the form $$ A V_k = V_{k+1} \overline{H}_k,$$ where $V_k \in \mathbb{R}^{n \times k}$ is an orthonormal matrix with $$K_k(A,b) := \text{span} \{ A^j b \: : \: j = 1,2,\dotsc,k-1 \} = \text{Ran} V_k$$ and $\overline{H}_k \in \mathbb{R}^{(k+1) \times k}$ is an quasi upper Hessenberg matrix. The last subdiagonal entry $h_{k+1,k}$ is critical, in the sense that it very precisely measures the size of a perturbation $\Delta A$ of $A$ such that $$ (A + \Delta A)V_k = V_k H_k $$ where $H_k$ is the upper Hessenberg matrix which consists of all but the last row of $\overline{H}_k$. If you find that $h_{k+1,k}$ is tiny relative to the norm of the matrix $A$, then $K$ is has numerical rank $k$ and your system is close to an uncontrollable system. If on the other hand, you can complete $n-1$ steps of the Arnoldi algorithm and none of these subdiagonal entries are insignificant, then your system is controllable and far from being uncontrollable.

It is possible to say substantially more about these matters. For a good introduction to the Arnoldi method, I refer to Yousef Saad's book "Iterative methods for linear systems". The first edition of Saad's book is freely available at his website, see


  • $\begingroup$ This is very insightful, thank you. Just want to make sure I got it right: technically the gramian is not always nonsingular, but in practice, we can assume so and proceed to the Cholesky with the modern numerical approaches and all? $\endgroup$
    – milez
    Jun 3, 2016 at 15:22
  • 1
    $\begingroup$ @Milez I have added some material in response to your question. The Cholesky factorization of $P$ is likely to fail even when the system is controllable. This is because $P$ is frequently singular to machine precision, eventhough $P$ is positive definite. Instead a more elaborate test must be devised. I have given some of the details. $\endgroup$ Jun 3, 2016 at 21:17
  • $\begingroup$ Did you mean to say that it is unlikely to fail even if the system is uncontrollable? In other words, controllability is not a requirement for balancing? $\endgroup$
    – milez
    Jun 4, 2016 at 10:35
  • 1
    $\begingroup$ Cholesky's factorization is frequently promoted as a way to test if a matrix $P$ is SPD. But you have to be mindful of the difference between exact and finite precision arithmetic. If Cholesky runs to completion in exact arithmetic, then the matrix is SPD. If it fails, then the matrix is not SPD. However, in finite precision arithmetic, Cholesky's method can fail even though the matrix is SPD. So while the Graminan $P$ is SPD iff the system is controllable, the Gramian $P$ may fail Cholesky's test because of rounding error. $\endgroup$ Jun 4, 2016 at 11:01
  • $\begingroup$ @CarlChristian: You state that $\boldsymbol{K}$ is used for Krylov. Did Krylov derive the controllability matrix? A reference would be nice! $\endgroup$ Feb 28, 2019 at 13:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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