Foundations — Observability matrix — rank test

This page assumes you have seen nothing. We build every letter, matrix, and operation in Observability matrix — rank test from the ground up, one brick at a time. Read top to bottom; each item uses only what came before it.

0. The picture we keep returning to

Think of a sealed box. Inside are spinning gears you cannot see. On the outside is a single needle on a dial. That is the entire story.

• The gears' positions = the state (hidden).
• The needle's reading = the output (visible).
• The physics that makes gears turn other gears = the matrix $A$.
• The wiring that connects gears to the needle = the matrix $C$.

Every symbol below is one of these four things, or a tool for measuring them.

1. A number line, a vector, and ${\mathbb{R}}^n$

The picture: a vector with $2$ slots is an arrow in a flat plane; the first number is how far right, the second how far up. A vector with $3$ slots is an arrow in the room you sit in. For $n$ slots we can't draw it, but the idea is identical: $n$ independent knobs, each set to some number.

Why the topic needs it: the hidden state $x$ is a vector in ${\mathbb{R}}^n$. The number $n$ is how many hidden gears there are. Everything — the size of matrices, when to stop taking derivatives — is counted in $n$.

2. A matrix as a machine that eats a vector

The picture: don't picture the grid. Picture a machine: you feed a vector in one end, and a (possibly different) vector comes out the other. The grid of numbers is just the machine's settings.

Why "eats a vector"? $A$ takes the current state $x$ and tells you the rate of change of the state — how the gears push each other. $C$ takes the state $x$ and produces the meter reading. Both are machines that turn one vector into another.

If matrices are new, spend real time in State-space representation and Rank and null space of a matrix before continuing.

3. The four LTI matrices: $A,B,C,D$

The parent note opens with $\dot{x}=Ax+Bu,y=Cx+Du$. Here is every symbol in that line.

The picture: below, the arrows show information flow. $A$ loops the state back on itself; $C$ taps the state out to the needle.

Why only $A$ and $C$ matter for observability: you know $u$ (you pushed the knobs) and you know $B,D$ (they're the wiring you built). So the parts of $y$ caused by $u$ can be calculated and subtracted away. What's left — called the free response — depends only on $A$ (how the state evolves) and $C$ (how it reaches the sensor). That is why the whole rank test uses just $(A,C)$.

4. Transpose $(\cdot {)}^{\top }$

The picture: stand the grid on its side. A wide row lying down becomes a tall column standing up.

Why the topic needs it: the "duality shortcut" says observability of $(A,C)$ equals controllability of $(A^{\top },C^{\top })$. To even read that sentence you must know what the little $\top$ does. See Controllability matrix — rank test.

5. Stacking matrices (block matrices)

The picture: bricks laid one on top of another. The observability matrix $\mathcal{O}$ is a tower of such bricks: $C$, then $CA$, then $C{A}^2$, and so on. Each brick is one more "view" of the state, obtained from one more derivative.

Why the topic needs it: $\mathcal{O}$ is this tower. Understanding it as stacked rows — not a mysterious single symbol — is the key to the rank test.

6. The derivative and $\dot{y}(0)$ — why we differentiate

The picture: watch the needle. Its position is $y(0)$. But you also see it moving — that speed is $\dot{y}(0)$. And whether that motion is speeding up — that is $\ddot{y}(0)$. Each new derivative is a new, independent piece of information you extract just by watching the same needle more carefully.

Why a derivative and not, say, an integral or a Fourier transform? Because the derivative is exactly the tool that asks "how is this quantity changing right now?" — and the model $\dot{x}=Ax$ tells us the change is $Ax$. So differentiating the output multiplies in another copy of $A$, which is precisely how we reach $CA$, $C{A}^2$, deeper into the state. No other operation produces this clean ladder of $C{A}^k$.

7. The matrix exponential $e^{At}$

The picture: a time-machine matrix. Feed it $x(0)$ and it returns $x(t)$ — where every gear will be after $t$ seconds of coupled spinning.

Why the topic needs it: the free response is $y(t)=C{e}^{At}x(0)$. This is the equation observability is built on. You don't need to compute $e^{At}$ — you only need to know it exists, that it starts at the identity ($e^{A\cdot 0}=I$), and that differentiating it brings down a factor of $A$: $\frac{d}{dt}{e}^{At}=A{e}^{At}$. That last fact is the engine behind the whole $C{A}^k$ ladder. See Kalman Filter for where this reconstruction is used in practice.

8. The identity matrix $I$ and matrix powers $A^k$

The picture: $I$ is a mirror that hands your vector back unchanged. $A^2$ is running the state one time-step of physics, then another.

Why the topic needs it: the rows of $\mathcal{O}$ are $C,CA,C{A}^2,\dots ,C{A}^{n-1}$ — literally $C$ times successive powers of $A$. And $A^0=I$ is why the very first row is just $C$.

9. Rank and null space — the heart of the test

The picture: shine a light through the matrix. Rank counts how many independent shadows it can cast. Any state that casts no shadow at all lives in the null space — it is invisible.

Why the topic needs it — this is the punchline:

• We want to solve $\left[\begin{array}{c}y(0)\\ \dot{y}(0)\\ ⋮\end{array}\right]=\mathcal{O}x(0)$ for the unknown $x(0)$.
• This has a unique answer for every state iff $\mathcal{O}$ has rank $n$ (full column rank), meaning no two states produce the same output record.
• If rank $<n$, some nonzero $x_0\in \ker (\mathcal{O})$ makes $y(t)\equiv 0$ — an unobservable state, a gear that never nudges the needle.

Deep-dive on both ideas: Rank and null space of a matrix.

10. Cayley–Hamilton — why the ladder stops at $n-1$

The picture: the ladder $C,CA,C{A}^2,\dots$ eventually stops giving new rungs. After $C{A}^{n-1}$, every next rung $C{A}^n,C{A}^{n+1},\dots$ is just a mix of ones you already have. So the tower $\mathcal{O}$ is complete at $n$ blocks.

Why the topic needs it: it tells us exactly where to stop. Without it, we wouldn't know that $n$ derivatives are enough. Full story in Cayley–Hamilton theorem.

Prerequisite map

Equipment checklist

Test yourself — cover the right side and answer aloud.

Connections

• State-space representation (where $A,B,C,D$ come from)
• Rank and null space of a matrix (the core linear-algebra tool)
• Cayley–Hamilton theorem (why we stop at $A^{n-1}$)
• Controllability matrix — rank test (the transpose-dual)
• Kalman Filter (uses this reconstruction in practice)
• Observability matrix — rank test (the parent — go there next)