THE OBSERVABILITY.

Can you reconstruct the hidden state from what you measure? A linear system is observable when the map from initial state to output history is one-to-one — when nothing inside can move without eventually showing at the sensor. The test is a rank: build Om = [C; CA; CA²; …; CAⁿ⁻¹] and check rank(Om) = n. Rendered, not quoted.

source Kálmán, R. E. — “On the General Theory of Control Systems”, Proc. First International Congress on Automatic Control (IFAC), Butterworth, London, 1960, pp. 481–492. The paper that introduced controllability and observability. No stable open full text — cited by author/title/year, marked AMBER. Reprint listing: IEEE Xplore 1104873.

Blue Team · builds & defends
3

The Model

A continuous LTI plant, state x∈ℝⁿ, scalar sensor:

ẋ = A x y = C x sampled: y(k) = C Aᵤ x₀

We carry two 2×2 systems with identical eigenvalues −1, −2. One is observable; in the other the −2 mode is wired so it never reaches C. Same dynamics, different visibility — observability is a property of the pair (A, C), not of A alone.

5

The Lineage

Can you see it? Observability is the exact mathematical dual of the-controllability: (A, C) is observable iff (Aᵀ, Cᵀ) is controllable. Reachability reversed in time.

It is the precondition every state estimator needs. the-luenberger-observer can only place its error dynamics if the pair is observable; the-kalman-filter can only converge on the observable subspace. Unobservable modes are invisible to every observer ever built — you can stabilise them (if stable) but never estimate them.

7

The Witness

Live re-check of the panel. Re-runs the rank test, the Gramian reconstruction and the duality identity against the planted tamper flag. Flips red the instant reported rank diverges from true rank.

The Machine
4

in ↓Data In

Two constructed pairs and one known initial state to recover:

compute Om · rank · invert Gramian
0

LITThe Panel booting…

proven result · out
8

out ↓Data Out

Red Team · attacks & breaks
1

wallThe Adversary

Observability is a rank test on an exact model. It is binary; estimation is analog. A system can pass and still be practically blind: if the observability Gramian is nearly singular, a tiny measurement noise explodes into enormous state error — the direction is seen but only barely.

Rank ignores conditioning, noise, model error and sample timing. “Observable” is necessary, never sufficient, for a good estimate. Nonlinear systems break it further: local observability can appear and vanish with the trajectory.

2

The Graveyard

“If I can measure the output, I can know the state.” Only if rank(Om)=n. A hidden mode leaves the state ambiguous forever — infinitely many x₀ produce the identical output. “Observability guarantees a good estimate.” It guarantees uniqueness in the noiseless limit, not accuracy. Near-unobservable pairs are ill-conditioned. (conditioning caveat: AMBER — system-specific) “More sensors always help.” Adding rows to C can never lower the rank, but redundant rows add nothing. Observability depends on direction, not sensor count.
6

The Tamper

The disclosed planted void: force the code to claim full rank for the unobservable pair — report the −2 mode as visible when it never reaches C. The Witness (7) recomputes the true rank and the Gramian and catches the lie live.