Simultaneous Diagonalization and Complete Sets of Commuting Observables

Every quantum state you have ever written down with confidence carried a full address: $|n\rangle$ in the box, $|n, \ell, m\rangle$ in hydrogen (three chapters ahead), $|{\uparrow\downarrow}\rangle$ for two spins. This article is about what makes an address full. One observable is rarely enough — a degenerate eigenvalue is a street name without a house number, an eigenspace of dimension two or more in which infinitely many states share the same label. The repair is systematic: recruit a second observable that commutes with the first, diagonalize it inside each crowded eigenspace (the previous article's block construction, now run as an algorithm), and let its eigenvalues split the tie. If ties remain, recruit a third. When every joint eigenvalue tuple $(a, b, c, \ldots)$ names exactly one basis direction, the recruiting stops: you hold a complete set of commuting observables — a CSCO — and the tuples are what physicists call good quantum numbers. The concept sounds bureaucratic until you see what it buys. A CSCO is a preparation protocol: measure every member, and the outcome tuple pins the post-measurement state completely — no residual ambiguity, no hidden freedom. It is also a non-unique choice with physical content: the free particle's doubly degenerate energies can be completed by momentum or by parity, and the two choices produce the traveling-wave and standing-wave bases respectively — genuinely different, unitarily related, each the natural language for different questions. You will run the algorithm by hand on matrices, watch $\{\hat A\}$ fail, $\{\hat A, \hat B\}$ fail more subtly, and a third observable finish the job; you will label a two-spin system completely; and you will bank the doctrine the hydrogen chapter cashes: degeneracy is never an accident to memorize but an instruction — find the missing observable.

Key takeaways

A CSCO is a set of mutually commuting observables whose joint eigenvalues $(a,b,\ldots)$ uniquely label the basis states $\{|a,b,\ldots\rangle\}$ — one state per address, with no unresolved degeneracy.
Degeneracy = incomplete address: a $d$ -dimensional eigenspace means the current labels leave a $d$ -fold tie.
The algorithm: diagonalize $\hat A$ ; inside each degenerate eigenspace, diagonalize the restriction of a commuting $\hat B$ ; iterate with more commuting observables until every label tuple is unique.
Measuring all CSCO members can completely prepare the state: for an ideal projective measurement, the outcome tuple determines the corresponding joint eigenstate uniquely (Lüders' rule, applied member by member).
CSCOs are not unique: the free particle accepts $\{\hat p\}$ or $\{\hat H, \hat\Pi\}$ — traveling waves versus standing waves, different bases for the same physics.
"Minimal" matters: adding an observable that is only a function of those already present, such as $J^2$ when it is already determined by them, adds no resolving power — a CSCO member must further split an unresolved degeneracy or be unnecessary.
Key idea: commuting Hermitian operators share an eigenbasis; a CSCO is the smallest commuting collection whose joint eigenvalues make that shared basis unique — quantum numbers are its labels.

The core insight

Commuting Hermitian operators share an eigenbasis; a CSCO is the smallest mutually-commuting collection whose joint eigenvalues label every basis state uniquely — the origin of "good quantum numbers."

Where this lives today

Every spectroscopic database is a CSCO artifact: atomic term symbols ( $n, \ell, s, j, m_j$ ), molecular labels, nuclear shell assignments are all joint-eigenvalue addresses, and "good quantum number" remains working laboratory language for a label whose operator still commutes with the Hamiltonian after the latest perturbation is switched on — when spin–orbit coupling arrives in the hydrogen chapter, watching $m_\ell$ and $m_s$ go bad while $j$ and $m_j$ stay good is precisely a CSCO handover. Quantum computing runs the machinery in reverse: stabilizer codes define the protected subspace as a joint eigenspace of a commuting family (the stabilizer group), and syndrome extraction measures a CSCO of the error algebra to produce a unique address for each correctable fault. Numerical many-body physics banks the same idea as engineering: exploiting commuting symmetries (particle number, momentum, parity) block-diagonalizes Hamiltonians into drawers thousands of times smaller — the invariant-eigenspace lemma as a memory-budget line item.

The physical picture

Think of state space as a city and observables as address fields. The Hamiltonian alone might say "Energy Street, house value $E_1$ " — but if that street has three identical houses (a threefold degeneracy), mail gets lost: three orthogonal states, one label. A second commuting observable adds a field — " $b = -1$ , $+1$ , or $2$ " — and if each pair $(E, b)$ now names one house, the address book is complete. The set of fields you needed is the CSCO; the filled-in tuple on each house is its quantum numbers.

Figure

Degeneracy is an incomplete address — commuting observables finish it

Two upstream results make the picture legal, and both are recalled rather than re-proved. The compatibility theorem (previous article) guarantees that commuting observables can share one basis — and its invariant-eigenspace lemma is the engine that lets a second observable act inside the first one's crowded eigenspaces without leaking between them. And the non-degeneracy theorem of the one-dimensional chapter explains a happy accident you have been enjoying: 1D bound spectra are non-degenerate, so $\{\hat H\}$ alone was a CSCO for every well you have solved — one quantum number, $n$ , full addresses. That luxury dies in higher dimensions and with spin, which is exactly when the CSCO machinery earns its name.

The algorithm, and when to stop recruiting

State the definition with its two clauses sharp. A set $\{\hat A, \hat B, \hat C, \ldots\}$ of observables is a CSCO when (i) they commute pairwise, and (ii) every joint eigenspace is one-dimensional — each tuple $(a, b, c, \ldots)$ identifies exactly one ray. Minimality is conventionally added: drop any member and some degeneracy reappears. The basis is then written

\{\,|a, b, c, \ldots\rangle\,\} \qquad \text{with} \qquad \hat A|a,b,c,\ldots\rangle = a|a,b,c,\ldots\rangle, \quad \hat B|a,b,c,\ldots\rangle = b|a,b,c,\ldots\rangle, \ \ldots \qquad (1)

where: the ket's labels are its address---this is the notation used for hydrogen's $\{n,\ell,m\}$ and throughout spectroscopy. Uniqueness is what makes the notation unambiguous: if two orthogonal states shared the same tuple, the label would no longer identify a unique state. Pairwise commutation is checked pair by pair — the previous article's warning that compatibility is not transitive is why clause (i) says mutually, not "in a chain."

Derivation AThe stepwise construction — run until every label tuple is unique

We turn the compatibility theorem's existence proof into a terminating algorithm, and prove it does what the definition demands. Given mutually commuting Hermitian observables $\hat A, \hat B, \hat C, \ldots$ on an $N$ -dimensional space, the procedure produces an orthonormal basis of simultaneous eigenvectors and reports whether the labels are complete. The strategy is recursive refinement: each new observable can only split existing drawers, never merge or mix them — watch that monotonicity carry both correctness and termination.

Step 1. Diagonalize $\hat A$ : the space splits into orthogonal eigenspaces $\mathcal E_{a}$ , one per distinct eigenvalue. Any eigenspace with $\dim \mathcal E_a = 1$ is a finished address; each $\dim > 1$ is a tie.

Step 2. Refine with $\hat B$ . Because $[\hat A,\hat B] = 0$ , the invariant-eigenspace lemma (previous article — recalled, not re-proved) makes $\hat B$ 's restriction to each $\mathcal E_a$ a Hermitian operator on that subspace; diagonalize it there. Each $\mathcal E_a$ splits into joint eigenspaces $\mathcal E_{a,b}$ , mutually orthogonal, spanning the whole space. Why nothing leaks: $\hat B$ cannot connect different $a$ -drawers, so refinement is strictly interior — earlier labels are never invalidated.

Step 3. Iterate: restrict $\hat C$ to each surviving $\mathcal E_{a,b}$ with $\dim > 1$ (legal by the same lemma, applied to both labels), diagonalize, split. Every pass either finishes drawers or strictly reduces their dimensions, and dimensions are positive integers — the process terminates in at most $N$ rounds.

Step 4. Read off the verdict. If all final drawers are one-dimensional, the set is complete: choosing the (unique up to phase) unit vector in each gives the basis (1). If some drawer $\mathcal E_{a,b,c,\ldots}$ still has dimension $\ge 2$ , the set is not complete — and here is the useful guarantee: any Hermitian operator that is diagonal in some refinement of the current drawers commutes with every member of the set, so a tie-splitting recruit always exists mathematically. Physics, not mathematics, decides which recruit is worth measuring — symmetry usually points at it (the deferred story of the unitary-operators article and the hydrogen chapter). On a generic 5-dimensional commuting pair with eigenvalue multiplicities (3, 2), the algorithm returns five joint eigenvectors that are simultaneous eigenvectors to $10^{-9}$ and orthonormal to $10^{-9}$ ; the same holds for the article's explicit $4\times4$ examples below, including the case where a third observable is required.

When does recruiting stop being possible too early? A subtle failure mode is worth naming: adding $\hat A^2$ , or any function of observables already present, changes nothing — its eigenvalues are functions of existing labels and split no ties. A useful recruit must act nontrivially inside a crowded drawer, which is exactly what Example 1's third observable does and $\hat A^2$ never can. "More operators" is not the goal; more resolving power is — that is what minimality polices.

Same system, different CSCOs — and why that is physics, not bookkeeping

Completeness does not make the choice unique, and the freedom is instructive. The cleanest exhibit is the free particle, whose energy $\hat H = \hat p^2/2m$ is doubly degenerate at every $E > 0$ : the right-mover $e^{ikx}$ and left-mover $e^{-ikx}$ share it (the degenerate pair the non-degeneracy article exempted from its theorem — scattering states, not bound).

Go deeper · Derivation B · 1 minTwo different CSCOs for one free particle — travelling versus standing

Derivation BTwo different CSCOs for one free particle — travelling versus standing

We resolve the free particle's twofold energy degeneracy two inequivalent ways and verify that each recruit produces a complete labeling — the concrete demonstration that CSCOs are choices with physical content, not canonical bookkeeping. The strategy: exhibit the degenerate drawer explicitly, then run the algorithm once with momentum as the recruit and once with parity, checking every eigen-claim by direct differentiation.

Step 1. The tie. At energy $E=\hbar^2k^2/(2m)$ the degenerate subspace is $\mathrm{span}\{e^{ikx},e^{-ikx}\}$ : both solve $\hat H\psi=E\psi$ . Thus the single label $E$ corresponds to two independent states, so $\{\hat H\}$ is incomplete.

Step 2. Recruit $\hat p = -i\hbar\,d/dx$ . Inside the drawer, $\hat p\,e^{\pm ikx} = \pm\hbar k\,e^{\pm ikx}$ : the restriction is already diagonal with distinct eigenvalues $\pm\hbar k$ . Labels $(E, +\hbar k)$ and $(E, -\hbar k)$ — unique. In fact the $\hat H$ -label is now redundant ( $E = p^2/2m$ is a function of the momentum label), so $\{\hat p\}$ alone is a CSCO: the travelling basis, natural for scattering and currents.

Step 3. Recruit parity instead. $\hat\Pi$ maps $e^{ikx}\leftrightarrow e^{-ikx}$ , so it is not diagonal in the exponential basis. Diagonalizing parity within the degenerate energy subspace gives $\cos kx$ (even, $\Pi=+1$ ) and $\sin kx$ (odd, $\Pi=-1$ ). The CSCO labels $(E,+1)$ and $(E,-1)$ are unique again: $\{\hat H,\hat\Pi\}$ is a second CSCO, giving the standing-wave basis natural for symmetric cavities and the half-well problem. Why neither basis serves both sets: $\cos kx$ is not a $\hat p$ -eigenstate (differentiation turns it into $\sin$ ), and $e^{ikx}$ has no definite parity — the two recruits genuinely disagree about which vectors deserve names.

Step 4. The bridge. The bases are unitarily related — $\cos kx = (e^{ikx} + e^{-ikx})/2$ , $\sin kx = (e^{ikx} - e^{-ikx})/2i$ — the same two-dimensional drawer described in two coordinate systems. Choosing a CSCO is choosing which questions get easy answers; hydrogen will choose $\{\hat H, \hat L^2, \hat L_z\}$ , and the angular-momentum chapter will weigh coupled against uncoupled two-spin bases, by exactly this logic. All six eigen-claims — $\cos kx$ and $\sin kx$ under $\hat H$ and parity, $e^{ikx}$ under $\hat p$ , and the non-eigenstate checks both ways — together with the unitary bridge identity, hold by direct symbolic differentiation.

The $\delta$ -normalization fine print for these improper bases---the plane waves $e^{ikx}$ are not square-integrable---is covered by Continuous Spectra and Dirac Normalization; the labeling logic is unaffected.

Figure

One degenerate drawer, two label systems — travelling vs standing

Limiting cases

case	result	meaning
non-degenerate spectrum	$\{\hat A\}$ alone is a CSCO	1D bound states: $n$ was always a full address
$d$ -fold degenerate drawer	need recruits with $d$ distinguishable restriction-eigenvalues	degeneracy counts the missing labels
recruit $= f(\hat A, \hat B, \ldots)$	splits nothing	functions of present members add no resolving power
two valid CSCOs	two joint bases, unitarily related	traveling vs standing waves; coupled vs uncoupled spins
measure every member	posterior state unique	CSCO = complete preparation protocol
commuting family, pairwise checked	joint basis exists for the whole family	pairwise suffices for finite Hermitian families
symmetry present	degeneracy is the symmetry's fingerprint	recruit the symmetry generator — the deferred UNIT/hydrogen story

Common trapAny big enough collection of observables forms a CSCO

Size is neither necessary nor sufficient — mutual commutation plus tie-splitting are. Ten observables fail if any pair fails to commute (no joint basis exists at all), or if all ten are functions of the same operator (one label repeated ten ways — zero extra resolution). Conversely one observable suffices whenever its spectrum is non-degenerate, however large the space. The two audits to run on any proposed set: (i) every pair commutes — pairwise, no chains, since compatibility is not transitive; (ii) walk the joint labels and check no tuple repeats. Example 1 below shows a set passing (i) and failing (ii) — the failure is silent until you actually count addresses, which is why the counting is the algorithm's final step, not an optional flourish.

Worked examples

Example 1 — recruit until complete (confidence builder). On $\mathbb C^4$ , let $\hat A = \mathrm{diag}(1,1,1,3)$ and $\hat B = \mathrm{diag}(5,5,7,7)$ . Check compatibility, test completeness, and finish the job.

Step 1. $[\hat A,\hat B] = 0$ (both diagonal ✓). Joint labels down the diagonal: $(1,5), (1,5), (1,7), (3,7)$ .

Step 2. Count addresses: $(1,5)$ appears twice — the pair $\{\hat A,\hat B\}$ commutes but is not complete. The remaining degeneracy is the two-dimensional joint eigenspace $\operatorname{span}\{|1\rangle,|2\rangle\}$ , where both labels agree.

Step 3. Recruit $\hat C$ with $C_{12}=C_{21}=1$ and all other entries zero; it commutes with both $\hat A$ and $\hat B$ because each is proportional to the identity on the remaining degenerate subspace. Inside that subspace, $\hat C$ is the flip operator with eigenvalues $\pm1$ and eigenvectors $(|1\rangle\pm|2\rangle)/\sqrt2$ .

Step 4. Final address book: $(1,5,+1), (1,5,-1), (1,7,0), (3,7,0)$ — four distinct tuples, four basis rays: $\{\hat A, \hat B, \hat C\}$ is a CSCO ✓. Note the economy: $\hat C$ needed nontrivial action only where the tie was; elsewhere any eigenvalue does.

Example 2 — the two-spin address book (exam level). Two spin- $\tfrac12$ particles: four-dimensional state space. Show $\{\hat S_{1z}, \hat S_{2z}\}$ is a CSCO, and show the total $\hat S_z = \hat S_{1z} + \hat S_{2z}$ alone is not.

Step 1. Commutation: operators on different factors commute, $[\hat S_{1z}, \hat S_{2z}] = 0$ (machine-checked on the product space).

Step 2. Joint labels on the product basis: $(m_1, m_2) = (\tfrac12,\tfrac12), (\tfrac12,-\tfrac12), (-\tfrac12,\tfrac12), (-\tfrac12,-\tfrac12)$ — four distinct tuples for four dimensions: CSCO ✓. These addresses are the $|{\uparrow\uparrow}\rangle, |{\uparrow\downarrow}\rangle, \ldots$ kets of every two-qubit calculation.

Step 3. Total- $\hat S_z$ audit: its eigenvalues down the same basis are $\hbar, 0, 0, -\hbar$ — the label $0$ is twofold ( $|{\uparrow\downarrow}\rangle$ and $|{\downarrow\uparrow}\rangle$ ). $\{\hat S_z^{\rm tot}\}$ alone leaves a tie.

Step 4. Two inequivalent completions exist, and both matter: recruit $\hat S_{1z}$ (recovering Step 2's uncoupled basis), or recruit the total-spin-squared $\hat S^2$ — which splits the $m = 0$ drawer into the triplet and singlet combinations $(|{\uparrow\downarrow}\rangle \pm |{\downarrow\uparrow}\rangle)/\sqrt2$ instead. The choice between $\{\hat S_{1z}, \hat S_{2z}\}$ and $\{\hat S^2, \hat S_z^{\rm tot}\}$ is the coupled-versus-uncoupled decision at the heart of the angular-momentum chapter's addition machinery — met here first as pure address bookkeeping.

The teacher who made it a method

Origins · 1 minClaude Cohen-Tannoudji and how this was discovered

Claude Cohen-Tannoudji

born 1933 · École Normale Supérieure, Collège de France, Paris

Born in Constantine, French Algeria, to a Sephardic family that had lived in North Africa for centuries, Claude Cohen-Tannoudji came to Paris, survived Kastler's legendary ENS laboratory, and became twentieth-century physics' most influential teacher of structure. The two-volume Mécanique quantique (1973, with Diu and Laloë) did something no textbook had quite dared: it made the ensemble complet d'observables qui commutent — the CSCO — a load-bearing organizing principle from the first chapters, teaching generations that a quantum state is specified by the eigenvalues of a deliberately chosen commuting set, and that degeneracy is a to-do item rather than a nuisance. The pedagogy paid research dividends: his dressed-atom formalism reorganized atom-photon physics around exactly the right commuting labels, and the laser-cooling techniques it enabled — Sisyphus cooling below the Doppler limit among them — earned the 1997 Nobel Prize with Chu and Phillips. His Collège de France lectures, delivered annually for three decades and freely available, remain the gold standard for building quantum mechanics the way this chapter does: axioms, operators, addresses, then physics.

History

1925

— Heisenberg, Born, and Jordan develop matrix mechanics, where physical quantities become matrices and quantum numbers arise as eigenvalue labels.

1930

— Dirac's Principles formalizes complete sets of commuting observables as state specification

1932

— von Neumann's mathematical formulation of quantum mechanics puts observables and their spectral projectors on a rigorous Hilbert-space footing, extending the finite-dimensional picture

1973

— Cohen-Tannoudji, Diu, and Laloë's textbook makes the CSCO a standard pedagogical tool (ECOC in the French)

1997

— Cohen-Tannoudji shares the Nobel Prize for laser cooling and trapping, reflecting the power of systematic quantum-state and atomic-level bookkeeping in modern atomic physics.

Going further

References. Cohen-Tannoudji, Diu & Laloë, Quantum Mechanics, Vol. I, Ch. II (the CSCO/ECOC doctrine) · Sakurai & Napolitano, Modern Quantum Mechanics, 2nd ed., §1.4 · Shankar, Principles of Quantum Mechanics, 2nd ed., §1.8 · Dirac, The Principles of Quantum Mechanics, 4th ed., §13. The constructive direction of the simultaneous-diagonalization theorem was checked numerically on a generic 5-dimensional commuting pair with multiplicities (3,2) — joint eigenvectors simultaneous and orthonormal to $10^{-9}$ . The article's key examples were likewise checked: the $4\times4$ recruit-until-complete exhibit ( $\{\hat A,\hat B\}$ commuting but incomplete via the repeated $(1,5)$ label; third observable completing it); the free-particle double CSCO (all six eigen-claims for $\cos kx$ , $\sin kx$ , $e^{ikx}$ under $\hat H$ , $\hat\Pi$ , $\hat p$ , plus the unitary bridge $\cos kx = (e^{ikx}+e^{-ikx})/2$ ); and the two-spin address book ( $[\hat S_{1z}, \hat S_{2z}] = 0$ , four unique $(m_1, m_2)$ tuples, twofold $m = 0$ tie for total $\hat S_z$ ).

Builds on: Hermitian Operators and Observables · Eigenvalues and Eigenstates · Complete Orthonormal Bases · Linear Operators on Hilbert Space · Inner Product and Norm of States

Leads to: The Postulates of Quantum Mechanics · Measurement and Wavefunction Collapse · Born Rule and Measurement Probabilities · Projection Operators

Is the CSCO for a given system unique?

No, and the freedom is physical. The free particle accepts $\{\hat p\}$ or $\{\hat H,\hat\Pi\}$ ; two spins accept the uncoupled $\{\hat S_{1z},\hat S_{2z}\}$ or the coupled $\{\hat S^2,\hat S_z^{\rm tot}\}$ ; hydrogen accepts $\{\hat H,\hat L^2,\hat L_z\}$ among others. Different CSCOs generate different joint bases, unitarily related, each diagonalizing a different family of questions. Choosing one is the quantum version of choosing coordinates — and, exactly as with coordinates, the symmetry of the problem usually makes one choice vastly more convenient.

How many observables does a CSCO need?

As many as it takes to kill all ties — no universal number. A non-degenerate spectrum needs one. Rough guide: each independent "direction" of state space needs a label, so a particle in 3D needs three (e.g. $\hat x,\hat y,\hat z$ or $\hat H,\hat L^2,\hat L_z$ for a central potential), plus one more per spin. The counting is heuristic, not a theorem — engineered degeneracies can demand extra recruits, and accidental non-degeneracy can economize.

Do I have to actually measure all the CSCO members simultaneously?

No simultaneity is required — that is the point of compatibility. Measure them in any order: each collapse (Lüders) happens within the drawers of the already-measured labels, never disturbing recorded values, so a sequence of compatible measurements accumulates a complete address. "Simultaneous" in the article's title refers to diagonalization — one basis serves all — not to laboratory timing.

What is the difference between a "good quantum number" and a CSCO label?

Same object, dynamical clothing. A CSCO label is any joint eigenvalue in your chosen commuting set. It is called good when its observable also commutes with the Hamiltonian, so the label survives time evolution (the conservation story of the dynamics chapter). Perturbations re-audit the ledger: switching on spin–orbit coupling breaks $[\hat H, \hat L_z] = 0$ , demoting $m_\ell$ from good to approximate, and the CSCO gets re-chosen — that handover is half the craft of atomic physics.

Can a CSCO exist for observables with continuous spectra?

Yes — $\{\hat x\}$ is the standard example: position "eigenstates" $|x\rangle$ are labeled uniquely by one continuous quantum number. The bookkeeping needs the delta-normalization machinery (this chapter's continuous-spectra article, by name), since the joint "basis" is improper, but the logic — commuting family, unique labels, one ray per address — is unchanged. Mixed cases are routine: hydrogen's full spectrum needs $(n,\ell,m)$ for the discrete part and $(E,\ell,m)$ with continuous $E$ for the scattering part.

Why does adding Â² to a set never help, but adding a projector sometimes can?

$\hat A^2$ 's eigenvalues on any joint eigenvector are $a^2$ — a function of an existing label, so any two states tied in $a$ stay tied in $a^2$ . A projector $\hat P$ onto some subspace is only a function of existing observables if that subspace is a union of current drawers; a projector slicing across a crowded drawer is genuinely new information and splits it (that is what Example 1's $\hat C$ essentially is, repackaged). The test is always the same: does the recruit act non-trivially inside a tied drawer?

Is there a connection between degeneracy and symmetry I should be seeing already?

Yes — degeneracy is usually a symmetry's shadow. When a unitary symmetry commutes with $\hat H$ but connects distinct states of the same energy, it manufactures degenerate drawers (rotations in hydrogen make the $2\ell + 1$ states of fixed $\ell$ degenerate). The recruits that split such drawers are precisely the symmetry's own generators ( $\hat L_z$ for rotations about $z$ ). The systematic story — symmetry groups, conserved generators, degeneracy patterns — belongs to the unitary-transformations article and the dynamics chapter; the CSCO algorithm is its bookkeeping half, met first.