Visual walkthrough — Observability — when KF can estimate state

We only allow ourselves one story: you can watch a signal $y(t)$, and you want to guess the hidden starting condition $x(0)$ that produced it. That's it. Everything below flows from that one wish.

---

Step 1 — The hidden state and the tiny window

WHAT. Picture a box. Inside is a list of numbers that changes over time — the state, written $x(t)$. If the box holds a spinning satellite, $x$ might be "current angle" and "current spin rate." You cannot open the box. You only have a small window that shows you one combined reading $y(t)$.

WHY. Estimation is exactly this situation: sensors never hand you the full internal truth, only a filtered peek. Before any maths, we must agree on what we can and cannot see.

PICTURE. In the figure, the violet cloud is everything living inside the box (the state). The orange window only lets a thin beam out — that beam is $y(t)$. Notice the magenta gear that is fully hidden behind the wall: we will spend the whole page asking "can we still figure out that hidden gear?"

---

Step 2 — How the hidden numbers move, and how the window reads them

WHAT. Two rules govern our box.

1. The state evolves on its own: $\dot{x}=Ax$. The dot means "rate of change per second." The matrix $A$ says how each hidden number nudges the others.
2. The window is a fixed lens: $y=Cx$. The matrix $C$ says which combination of hidden numbers leaks out.

$\underset{\text{how fast the state changes}}{\underbrace{\dot{x}}}=\underset{\text{mixing rule}}{\underbrace{A}}\underset{\text{hidden numbers}}{\underbrace{x}},\underset{\text{what you see}}{\underbrace{y}}=\underset{\text{the window}}{\underbrace{C}}\underset{\text{hidden numbers}}{\underbrace{x}}$

WHY these two matrices and nothing else? We threw away the input term $Bu$ from the full state-space form. Reason: inputs $u$ are things you command, so you already know their effect and can subtract it. Only the unforced motion ($\dot{x}=Ax$) tests whether the window itself is good enough. Observability is a property of the pair $(A,C)$ — mixing rule plus window — never of $B$.

PICTURE. The figure shows $A$ as violet arrows between the hidden gears (who pushes whom) and $C$ as the single orange arrow from the gears to the window. If a gear has no path — arrows or window — reaching the orange beam, trouble is coming.

---

Step 3 — Playing the film forward: the state at time $t$

WHAT. Given the starting list $x(0)$, the rule $\dot{x}=Ax$ has exactly one solution: $x(t)=e^{At}x(0).$ The object $e^{At}$ is the propagator: feed it the start, it fast-forwards the film to time $t$.

WHY the exponential and not something simpler? For a single number, $\dot{x}=ax$ solves to $x(t)=e^{at}x(0)$ — the exponential is the function whose rate of growth equals itself times $a$. Matrices behave the same way once you define $e^{At}=I+At+\frac{1}{2}(At{)}^2+\cdots$. We pick this tool because it is the unique answer to "what function's slope is $A$ times itself?" — which is precisely the equation $\dot{x}=Ax$.

Now plug the window on top. What the sensor actually records is $y(t)=C{e}^{At}x(0).$

PICTURE. The figure shows the same starting dot $x(0)$ (magenta) being carried along a curved violet trajectory by $e^{At}$, and at each instant the orange window samples one point of it to make the $y(t)$ curve on the right.

---

Step 4 — Squeezing more equations out with derivatives

WHAT. One snapshot $y(0)=Cx(0)$ is usually too little. So we watch how $y$ changes — its slope, its curvature, and so on — right at the start $t=0$:

$y(0)=Cx(0),\dot{y}(0)=CAx(0),\ddot{y}(0)=C{A}^{2}x(0),\dots$

Each line is a fresh linear equation in the same unknown $x(0)$.

WHY derivatives? A derivative asks "which way and how fast is the signal moving right now?" Differentiating $y=C{e}^{At}x(0)$ pulls down one factor of $A$ each time (because the slope of $e^{At}$ is $A{e}^{At}$), and at $t=0$ the propagator $e^{A\cdot 0}=I$ vanishes cleanly. So the $k$-th derivative gives the crisp row $C{A}^{k}x(0)$. We use derivatives because they are the natural way to convert the shape of a signal over time into a stack of static equations.

PICTURE. The figure zooms into $t=0$ on the $y$ curve: the orange dot is $y(0)$ (height), the magenta tangent line is $\dot{y}(0)$ (slope), the violet curving arc is $\ddot{y}(0)$ (bending). Each geometric feature = one equation = one more clue about $x(0)$.

---

Step 5 — Why we stop: Cayley–Hamilton caps the pile

WHAT. We could keep differentiating forever: $C{A}^3,C{A}^4,\dots$. But we stop at $C{A}^{n-1}$. The Cayley–Hamilton theorem says every $n\times n$ matrix $A$ 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 of the earlier powers $I,A,\dots ,A^{n-1}$. Multiply by $C$: the row $C{A}^n$ is a blend of rows you already wrote down.

WHY this matters. It promises the test is finite. Without it, "differentiate forever" would never terminate. Cayley–Hamilton guarantees: after $n$ rows, no new direction of information can ever appear. Row $n$, row $n+1$, all the way to infinity — redundant.

PICTURE. The figure stacks the rows $C,CA,\dots$ as horizontal bars. The first $n$ bars are fresh colours; the moment we reach $C{A}^n$ the bar is drawn as a shadow blend of earlier bars — literally the same colour mix — showing it adds nothing new.

---

Step 6 — Stacking into one matrix $\mathcal{O}$

WHAT. Line up all $n$ useful equations into a single tall matrix equation:

$$\underset{\mathcal{O}\text{(observability matrix)}}{\underbrace{\left[\begin{array}{c}C\\ CA\\ C{A}^2\\ ⋮\\ C{A}^{n-1}\end{array}\right]}}x(0)=\left[\begin{array}{c}y(0)\\ \dot{y}(0)\\ \ddot{y}(0)\\ ⋮\\ y^{(n-1)}(0)\end{array}\right].$$

Term by term. Each block $C{A}^k$ is a "clue lens": $C$ is the raw window, and each extra $A$ means "and let the dynamics stir the pot $k$ times before we look." The right-hand column is pure recorded data (values and derivatives of $y$ at $t=0$). The only unknown is $x(0)$.

WHY stack them? A single lens might miss directions of the state. Stacking many lenses (each stirred differently by $A$) is our best possible combined view. If this stack still can't pin down $x(0)$, nothing ever will.

PICTURE. The figure literally stacks the orange-tinted lens rows into the tall violet block labelled $\mathcal{O}$, with the magenta unknown $x(0)$ on the right and the data column below it.

---

Step 7 — The verdict: full column rank means "solvable"

WHAT. We now have $\mathcal{O}x(0)=\text{data}$. When does this have exactly one answer for $x(0)$?

Rank counts how many truly independent directions the rows of $\mathcal{O}$ cover. If $\mathrm{rank}(\mathcal{O})=n$ (as wide as the number of hidden states), the equations pin down every direction — unique solution. If $\mathrm{rank}(\mathcal{O})<n$, there is a leftover direction the lenses never touch.

WHY rank and not "number of rows"? You can pile on a hundred sensors, but if they all report the same combination, they add height and zero new independent directions — no new rank. This is the exact trap the parent note warns about.

PICTURE. The figure shows two cases side by side. Left (observable): $\mathcal{O}$ collapses all of state-space onto a single point → one answer. Right (unobservable): a whole magenta arrow of starting states all map to $y(t)\equiv 0$ → the window can't distinguish any of them.

---

Step 8 — The degenerate cases you must not skip

WHAT. Three edge scenarios decide real GNC designs. Each is its own tiny panel in the figure.

(a) A completely decoupled, unmeasured state. If $A$ has a hidden number that neither pushes anything visible nor is seen by $C$, its column in every $C{A}^k$ block is zero. The lens beam never reaches it. → unobservable (parent's Example 2).

(b) A hidden state that pushes a visible one. Position–velocity with only a position sensor: velocity is invisible directly, but $A_{12}\ne 0$ means velocity drives position, so it leaks into $\dot{y}$. → observable (parent's Example 1). This is the gear-pushing-a-gear win.

(c) Full rank but barely — weak observability. $\mathcal{O}$ technically reaches rank $n$, but one direction contributes almost nothing (its beam is a whisper). Measurement noise then drowns it. This is not caught by rank alone; you need the observability Gramian to grade the strength, and it explains why detectability is the softer fallback the Riccati steady state actually requires.

WHY cover all three? The rank test is binary, but reality lives on a spectrum from "blind" to "crystal clear." A designer who only checks rank will be blindsided by case (c).

PICTURE. Three stacked panels: (a) a gear in a sealed box — no beam; (b) a hidden gear whose motion ripples into the lit gear; (c) a gear faintly lit — barely observable, noise-vulnerable.

---

The one-picture summary

Everything on this page compresses into one flow: watch $y$ → take derivatives → stack the $C{A}^k$ lenses into $\mathcal{O}$ → check its rank → out pops (or doesn't) a unique $x(0)$.

Recall Feynman: the whole walkthrough in plain words

You've got a locked box of spinning gears and one little window. You film the window for a while. First frame tells you a bit. How the picture moves (its speed) tells you more. How that motion curves tells you even more — each is a fresh clue. But you can't keep gaining clues forever: after $n$ of them (Cayley–Hamilton's promise) every new clue is just an old clue in disguise. So you gather exactly $n$ clues, stack them into one big clue-sheet called $\mathcal{O}$, and ask: do these clues, together, nail down every gear's starting position? If yes — full rank — the box is observable and a Kalman Filter can rebuild the whole inside from the window alone. If some gear left no fingerprint anywhere in the clues, it hides in $\ker (\mathcal{O})$, and no filter on Earth can find it. And even when every gear leaves some mark, a mark too faint to read through the noise (weak observability) is the trap that sends you to the Gramian and to detectability.