Operator Algebra of Institutional Alignment

Both operators are projections, and that settles a lot

Working in the space of square-integrable cycle functions, Δ is the orthogonal projection onto the fragility-invariant subspace K* — the capital behaviours that do not vary with the fragility cycles — and Λ is the projection onto the mission subspace M, the cadence the institution’s purpose actually requires. The alignment transform is their composition, A = ΛΔ: carry capital first out of the image of the fragility cycles, then into the mission subspace.

🔴 A is not itself a projection. A product of two orthogonal projections is a projection if and only if they commute, and these do not — which is not an inconvenience discovered here but the generic case. It matters because the source asserts a direct-sum decomposition K = A ⊕ E and calls it the most rigorous statement in the line of work. As written it is a tautology: with E defined as I − A, the identity K = A(K) + E(K) holds for any operator whatever. A genuine direct sum needs complementary subspaces and a unique decomposition, which requires A idempotent.

What survives is better. Δ alone gives a real structure theorem — capital splits uniquely into a fragility-invariant part and a fragility-dependent one, K = K* ⊕ K*^⊥. That is the decomposition the prose was reaching for.

Where the degrees actually live

Δ and Λ are projections, so each has spectrum {0, 1} and nothing else. There are no degrees in either operator individually — which is why the source’s “eigenvalues representing degrees of alignment” could not have been theirs. The degrees live in the relation between the two subspaces, and the operator that sees it is A*A: self-adjoint, positive, spectrum in [0, 1], with eigenvalues cos²θ_k where θ_k are the principal angles between K* and M.

The index follows directly: Align(K) = ‖A K‖ / ‖K‖ in the L² norm, range [0, 1] by construction. It is one exactly when K lies in the intersection of the two subspaces, and zero exactly when the transform annihilates it.

None of the mathematics is new, and saying so makes it stronger

Principal angles between subspaces are Björck and Golub’s object, defined and computed there, with the numerically stable construction and the perturbation analysis this paper would need if it ever instantiated the index. The structure of a pair of orthogonal projections — which is the whole subject here — is Halmos’s: two subspaces in generic position decompose the space into 2×2 blocks parameterised by their angles.

Halmos settles the earlier point rather than corroborating it. Two projections commute only when their subspaces are compatible, so A’s failure to be a projection is generic, and the angles are precisely the invariants measuring that failure. Which makes the central finding — alignment admits degrees because the composition is not a projection — a rediscovery of the two-subspace normal form in institutional language. Stating it that way costs nothing and gives the reader somewhere to check the argument.

Competing mandates, stated formally

Different domains have different alignment operators, and the commutator [Λ_a, Λ_b] is non-zero exactly when their subspaces are incompatible. What that says precisely: aligning capital to domain a and then to domain b does not give what aligning to b and then to a gives. The order of institutional attention changes the outcome, and no single deployment satisfies both cadences.

That is sharper than “mandates compete”. It says the competition is structural rather than budgetary — it survives having enough capital for both. It also gives a formal reading of mission-class compression: applying one authority regime uniformly across domains whose operators do not commute silently privileges whichever cadence is applied first.

And the interference is itself measurable. The principal angles between the two mission subspaces quantify how much of one domain’s admissible capital the other can also use, and the norm of the commutator is bounded by their sines.

The objection to publishing an index at all

An institutional alignment score is exactly the object Goodhart warns about, and here the mechanism is available in closed form rather than as an analogy. Align(K) is maximised by moving capital into the intersection of the two subspaces — but an institution that cannot change its capital can raise the same score by changing M, because the mission subspace is declared rather than observed. Narrowing the mission to the cadence the capital already has scores a perfect one.

The institutional form of the same point is Power’s: a measurement apparatus does not passively observe an activity, it reorganises the activity into something auditable. An alignment index in use would make institutions legible as (K*, M) pairs — a representation nobody currently owes anyone.

The defence is structural and belongs in the definition, not in guidance. M must be fixed by an instrument that predates the measurement and cannot be revised to improve it — a mandate, an asset schedule, a statutory obligation — so that only K is free. An index whose second argument is chosen by the party being scored measures nothing.