Visual walkthrough — Observability matrix — rank test

This page is the flagship picture-tour of the parent result. If a word here feels unfamiliar, it gets built here — from line one.

---

Step 1 — The box we can't see inside

WHAT. We have a machine whose insides are described by a list of numbers we call the state, written $x$. If there are $n$ hidden numbers, we say $x$ lives in ${\mathbb{R}}^n$ — think of it as an arrow (a vector) pointing somewhere in an $n$-dimensional room. We cannot look at $x$ directly. All we get is a reading on a sensor, a smaller list of numbers $y$.

WHY. Everything about "can we figure out the start?" hinges on this gap: many hidden $x$, few visible $y$. We must ask whether the few readings pin down the many unknowns.

PICTURE. The box has hidden gears (the components of $x$); a single external needle shows $y$.

Here $p$ is how many sensor numbers we read; often $p=1$ (a single dial).

---

Step 2 — Throw away everything the input does

WHAT. The full system is $\dot{x}=Ax+Bu,y=Cx+Du$. The dot on $\dot{x}$ means "rate of change of $x$ in time". The $u$ is the input — knobs we ourselves turn. We set $u=0$.

WHY. Because we chose $u$, its entire effect on $y$ is already known to us — we can subtract it off perfectly. What is left is the machine's free response: how it drifts on its own from wherever it started. Observability is only about that free part, so $B,D,u$ simply vanish from the question. Only $A$ and $C$ remain.

PICTURE. The known input-contribution is peeled away like a sticker, leaving the clean free-response curve.

$$y(t)=\underset{\text{dial wiring}}{\underbrace{C}}\underset{\text{how }x\text{ drifts in time}}{\underbrace{e^{At}}}\underset{\text{the prize}}{\underbrace{x(0)}}$$

---

Step 3 — One instant is not enough

WHAT. At the single instant $t=0$ the reading is $y(0)=Cx(0)$.

WHY it fails alone. $C$ has only $p$ rows (say $1$). One equation $y(0)=Cx(0)$ fixes only the shadow of $x(0)$ along one direction. All arrows $x(0)$ that cast the same shadow give the identical reading — we can't tell them apart. We need more equations.

PICTURE. A whole line of candidate arrows all project to the same $y(0)$ — the dial can't distinguish them.

$$y(0)=\underset{\text{1 row}}{\underbrace{C}}x(0)\Rightarrow \text{only 1 direction of the }n\text{ pinned down.}$$

---

Step 4 — Watch how fast the needle moves (differentiate)

WHAT. Take time-derivatives of $y$ at $t=0$. The derivative $\dot{y}$ means "how fast the reading changes". Because $\frac{d}{dt}{e}^{At}=A{e}^{At}$, each derivative pulls down one more factor of $A$:

$$\begin{array}{rlrl}y(0)& =Cx(0)& & \text{(the reading)}\\ \dot{y}(0)& =CAx(0)& & \text{(its speed)}\\ \ddot{y}(0)& =C{A}^{2}x(0)& & \text{(its acceleration)}\\ & ⋮& & \\ y^{(k)}(0)& =C{A}^{k}x(0)& & (k\text{-th derivative})\end{array}$$

WHY. Each derivative is a brand-new equation in the same unknown $x(0)$, but with a new coefficient row $C{A}^k$. Every genuinely new row points at a new direction of $x(0)$ we can now see. Position gives you $C$; velocity gives you $CA$; and so on — the needle's motion leaks out the hidden gears.

PICTURE. Each derivative order lights up a new arrow direction being captured.

Term-by-term in $y^{(k)}(0)=C{A}^{k}x(0)$:

• $y^{(k)}(0)$ — the $k$-th derivative of the reading, a number we can measure.
• $C$ — dial wiring (fixed).
• $A^k$ — the linkage applied $k$ times; this is what rotates us into fresh directions.
• $x(0)$ — still the one prize we're solving for.

---

Step 5 — Why we stop at $A^{n-1}$ (Cayley–Hamilton)

WHAT. We do not keep differentiating forever. We stop at the $(n-1)$-th derivative, using rows $C,CA,\dots ,C{A}^{n-1}$.

WHY. The Cayley–Hamilton theorem says every $n\times n$ matrix satisfies its own characteristic equation, which rearranges to:

$$A^n=c_{n-1}{A}^{n-1}+\cdots +c_{1}A+c_{0}I.$$

So $A^n$ is just a recipe mixing lower powers. Multiply by $C$: the row $C{A}^n$ is a mix of rows we already have. It contributes zero new directions. Every derivative past $n-1$ is redundant.

PICTURE. The ladder of rows grows, then hits a ceiling — new rungs after $C{A}^{n-1}$ lie flat on the old ones.

---

Step 6 — Solve for the prize: the rank test

WHAT. Stack the measured derivatives into one tall equation:

$$\underset{\text{known column}}{\underbrace{\left[\begin{array}{c}y(0)\\ \dot{y}(0)\\ ⋮\\ y^{(n-1)}(0)\end{array}\right]}}=\underset{\mathcal{O}}{\underbrace{\left[\begin{array}{c}C\\ CA\\ ⋮\\ C{A}^{n-1}\end{array}\right]}}x(0).$$

Solving for $x(0)$ uniquely is possible iff $\mathcal{O}$ hits every direction of the $n$-dimensional room — i.e. its columns are independent, i.e. its rank equals $n$.

WHY. Rank $=n$ means $\mathcal{O}$ flattens no nonzero arrow to zero: its null space is only $\{0\}$. Two different starts can never give identical readings. Fewer than $n$ means some arrow slips through invisibly.

PICTURE. Full rank = the map $\mathcal{O}$ is injective; every distinct $x(0)$ lands on a distinct reading-stack.

$$\boxed{(A,C)\text{observable}⟺\mathrm{rank}(\mathcal{O})=n}$$

This is the same yes/no gate a Kalman Filter needs before its state estimate can converge, and the PBH test gives an eigenvalue-flavoured version of the very same check.

---

Step 7 — The degenerate case: a gear that never moves the dial

WHAT. Suppose $\mathrm{rank}(\mathcal{O})=r<n$. Then there is a nonzero arrow $x_0$ with $\mathcal{O}{x}_0=0$.

WHY it's fatal. If $\mathcal{O}{x}_0=0$ then every row kills it: $C{x}_0=0,CA{x}_0=0,\dots$. Feed it forward and $y(t)=C{e}^{At}{x}_0\equiv 0$ for all time. This arrow is an unobservable state — an internal motion that never budges the needle.

PICTURE. Two decoupled identical modes ($A=2I$, single dial): the difference-direction slides into the null space and vanishes from the reading.

---

The one-picture summary

The whole tour on one canvas: input peeled off → free response → differentiate to harvest directions → ladder $C,CA,\dots$ capped at $n-1$ by Cayley–Hamilton → count independent rungs. Rank $=n$ ⇒ every gear seen (observable); rank $<n$ ⇒ a hidden gear (a null-space arrow that reads $y\equiv 0$).

Recall Feynman: the whole walkthrough in plain words

You've got a sealed box of spinning gears and just one dial on the outside. First you ignore anything you did to the box (the knobs) — you know that effect, subtract it. Now watch the dial drift on its own. Reading it once tells you a little: one shadow of the gears. But watch how fast it moves, and how fast that changes — each new "speed of the speed" is a fresh clue pointing at gears you couldn't see before. You keep collecting clues, but there's a mathematical guarantee (Cayley–Hamilton) that after $n-1$ speeds there are no genuinely new clues left — everything after just repeats. Stack all those clue-rows into one tall table $\mathcal{O}$. Count the truly independent rows. If that count equals the number of gears $n$, the dial secretly reveals every gear — the box is observable, and you can rewind to its exact starting spin. If the count falls short, some combination of gears cancels perfectly and never nudges the dial — that hidden motion is forever invisible, and no amount of watching will recover it.

---

Connections

• Observability matrix — rank test (parent result)
• Cayley–Hamilton theorem (why the ladder stops at $A^{n-1}$)
• Controllability matrix — rank test (the mirror-image derivation)
• Rank and null space of a matrix (what "rank $=n$" and "kernel $=\{0\}$" mean)
• Kalman Filter (needs observability to converge)
• Kalman decomposition (isolates the unobservable arrows of Step 7)
• State-space representation (the $\dot{x}=Ax+Bu$ language)
• PBH test (eigenvalue-based twin of the rank test)