Definitions of integrability compared

What does calling a system integrable allow us to conclude? The answer depends on the mathematical setting and the structure established. Finite-dimensional Liouville integrability has a precise count of independent, commuting integrals. A differential Lax representation, a quantum transfer family and an exact stochastic formula make different claims. This reference separates those claims, gives the relevant hypotheses, and shows explicitly why counting commuting finite matrices cannot provide a useful universal test.

Required background. The Hamiltonian bracket bridge explains conservation and involution; the linear algebra bridge explains eigenvectors, projectors and commutators. Helpful background. A first picture of integrability introduces the questions before these distinctions become necessary.

Integrability in four settings

The table compares the meanings used in the site’s worked examples. It does not propose one definition covering every field.

Setting	Structure being established	Conclusion that needs additional work
Autonomous Hamiltonian system on a $2n$-dimensional symplectic manifold	$n$ functionally independent integrals in involution	A compact regular torus, global action–angle coordinates, or periodic motion
KdV on the rapidly decaying line	A specified Lax pair and an inverse-scattering construction; explicit finite-pole reconstruction in the worked articles	A theorem for arbitrary initial data, operator evolution, or long-time asymptotic completeness
Finite periodic XXX spin chain	A local Yang–Baxter relation producing a commuting transfer family containing the Hamiltonian	Independence and locality of specified charges, or completeness of a Bethe construction
Finite-ring exclusion dynamics	Exact formulas for a stated stationary law, current or dynamical mode	Every transition probability, every observable, or a complete spectral resolution

“Exactly solvable” must likewise name the requested object: a trajectory, spectrum, eigenvector, distribution or observable. A formula for one object need not solve all the others.

Liouville integrability counts differential rank

Let $(M,\omega)$ be a smooth symplectic manifold of dimension $2n$, with a smooth real autonomous Hamiltonian $H$. In canonical coordinates our bracket is

$$\{f,g\}=\sum_{j=1}^n \left( \frac{\partial f}{\partial q_j}\frac{\partial g}{\partial p_j} -\frac{\partial f}{\partial p_j}\frac{\partial g}{\partial q_j} \right).$$

The formulation of Liouville integrability used here asks for smooth functions $F_1=H,F_2,\ldots,F_n$ with

$$\begin{gathered} \{F_i,F_j\}=0\quad\text{for every }i,j,\\ dF_1\wedge\cdots\wedge dF_n\ne0. \end{gathered}$$

The first condition, involution, holds on $M$. The second, functional independence, is required on an open dense regular subset of $M$. Conservation follows because $\dot F_i=\{F_i,H\}=0$. These are the conditions in Cannas da Silva, January 2006 revision, § 18.4, Theorem 18.9 and Definition 18.10, pp. 109–110, PDF. If a claim concerns only an open region, that region must be specified. For a degenerate Poisson bracket, first identify the symplectic leaf; its dimension determines the count.

Independence is a rank condition, not a count of different expressions. For any $H$,

$$\{H,H^2\}=0, \qquad d(H^2)=2H\,dH, \qquad dH\wedge d(H^2)=0.$$

Thus writing down $H$ and $H^2$ does not establish integrability of a two-degree-of-freedom system. Conversely, nonlinear equations can be integrable: one degree of freedom needs only the energy, wherever $dH\ne0$. The first-picture lesson works through a nonlinear quartic oscillator.

The torus conclusion has extra hypotheses

Take a connected component $\Lambda$ of a common level $F^{-1}(c)$. If it is compact and regular, the commuting Hamiltonian flows are complete there, and the Liouville–Arnold theorem identifies it as an $n$-torus. A neighborhood of this torus has canonical actions and circle-valued angles with

$$\dot I_j=0, \qquad \dot\theta_j=\Omega_j(I) =\frac{\partial H}{\partial I_j}.$$

See Cannas da Silva, Lemma 18.11 and Theorem 18.12, pp. 110–111, PDF and the Liouville–Arnold reference for the precise local neighborhood statement. Compactness is not part of the definition of Liouville integrability. Regularity must hold on the component actually being used.

Even on such a torus, an orbit need not close. A nonstationary orbit is periodic precisely when some $T\gt0$ obeys $T\Omega\in2\pi\mathbb Z^n$. Rationally independent frequencies instead give an orbit dense in that torus. Neither case supplies a global real-valued angle chart through singular fibers. The harmonic oscillator makes the singular origin explicit; the free particle in the theorem reference shows why a noncompact regular level need not be a torus.

A Lax equation specifies an algebraic or analytic claim

For differentiable finite matrices, a Lax equation

$$\dot L=[B,L]$$

preserves spectral invariants. For every positive integer $m$, cyclicity of the ordinary trace gives

$$\begin{aligned} \frac{d}{dt}\operatorname{tr}L^m &=m\operatorname{tr}\bigl(L^{m-1}[B,L]\bigr)\\ &=m\bigl(\operatorname{tr}(BL^m)-\operatorname{tr}(L^mB)\bigr)=0. \end{aligned}$$

That calculation has not used a Poisson bracket. It therefore cannot, by itself, prove involution of the invariants. Their independence and sufficiency also need proof. The open Toda construction supplies these missing steps for its specified Hamiltonian phase space. A vacuous choice such as $L=I$, $B=0$ satisfies a Lax equation alongside any dynamics and records no information about that dynamics.

For differential operators, further distinctions matter. In the site’s smooth real, rapidly decaying KdV regime,

$$\begin{aligned} u_t+6uu_x+u_{xxx}&=0,\qquad x\in\mathbb R,\\ L&=-D^2-u,\qquad D=\partial_x,\\ A&=-4D^3-6uD-3u_x. \end{aligned}$$

On Schwartz test functions, the differential-expression calculation gives

$$[A,L]f=(6uu_x+u_{xxx})f, \qquad L_tf=-u_tf.$$

Thus $L_t=[A,L]$ is formally equivalent to KdV. The Lax-pair derivation verifies the cancellations. For bounded smooth real $u$, the Schrödinger operator $L$ has the usual self-adjoint realization on $L^2(\mathbb R)$ with domain $H^2(\mathbb R)$, the functions whose weak derivatives through order two are square integrable. This does not make every product involving the third-order operator $A$ defined on that same domain. An operator evolution theorem needs domains and evolution maps, not just equality of differential expressions.

An inverse-scattering solution also needs admissible spectral data, an inverse map, and hypotheses ensuring the reconstructed field solves the initial-value problem. Aktosun 2009, §§ III and VIII–IX describes that reconstruction framework; his KdV field is the negative of ours. The site’s one-pole and two-pole calculations establish explicit reflectionless sectors. Those calculations alone are not existence or asymptotic-completeness theorems for arbitrary initial data, and they do not transfer unchanged to periodic or boundary-driven problems.

Every finite Hermitian matrix has commuting projectors

Let $H$ be any Hermitian operator on a complex Hilbert space of finite dimension $d$. Choose an orthonormal eigenbasis $|a\rangle$, with $H|a\rangle=E_a|a\rangle$, and define

$$P_a=|a\rangle\langle a|.$$

Direct multiplication gives

$$\begin{gathered} P_aP_b=\delta_{ab}P_a,\qquad \sum_{a=1}^d P_a=I,\\ HP_a=P_aH=E_aP_a,\qquad [P_a,P_b]=0. \end{gathered}$$

These $d$ operators are linearly independent: applying $\sum_a z_aP_a=0$ to $|b\rangle$ gives $z_b|b\rangle=0$, hence every $z_b=0$. If an eigenvalue is degenerate, the basis within its eigenspace is not unique; the rank-one projectors refine its spectral projection. The construction still works.

Consequently, merely asking for many linearly independent commuting operators accepts every finite Hermitian Hamiltonian. This is the spectral-projector obstruction discussed in Caux and Mossel 2011, § 3, printed pp. 4–5, arXiv v1 PDF. It explains why a useful claim must say more about how the operators are constructed and how they relate to the physical model.

Nor should linear independence be silently replaced by algebraic independence: each projector already obeys $P_a^2-P_a=0$, and every finite matrix obeys its characteristic polynomial. For a nondegenerate spectrum, the projectors are even polynomials in $H$:

$$P_a=\prod_{b\ne a}\frac{H-E_bI}{E_a-E_b}.$$

Evaluating this polynomial on each eigenvector proves the formula. Its denominators require distinct eigenvalues; it does not choose a basis inside a degenerate eigenspace.

For a concrete check, take $J\gt0$ and

$$H=J\begin{pmatrix}1&-i\\ i&1\end{pmatrix}.$$

Its eigenvalues are $0$ and $2J$, with

$$P_{2J}=\frac{H}{2J} =\frac12\begin{pmatrix}1&-i\\ i&1\end{pmatrix}, \qquad P_0=I-P_{2J}.$$

Since $H^2=2JH$, these matrices satisfy $P_{2J}^2=P_{2J}$ and $P_0P_{2J}=0$. Their existence illustrates the general theorem; it does not identify a special many-body mechanism.

The XXX transfer construction supplies additional structure

For the periodic spin-½ XXX chain on $(\mathbb C^2)^{\otimes N}$, $N\ge3$, let $P_{n,n+1}$ swap neighboring spins, $N+1\equiv1$, and

$$H=\frac J2\sum_{n=1}^N(I-P_{n,n+1}),\qquad J\gt0.$$

The operative construction starts with the local matrix $R(v)=vI+iP$, where $v$ is a dimensionless spectral parameter. Set

$$\begin{aligned} L_{an}(\lambda)&=R_{an}(\lambda-i/2),\\ \tau(\lambda)&=\operatorname{tr}_a \bigl(L_{aN}(\lambda)\cdots L_{a1}(\lambda)\bigr). \end{aligned}$$

Here $a$ is a two-dimensional auxiliary factor and the displayed ordering matters. The Yang–Baxter and RTT relations imply

$$[\tau(\lambda),\tau(\mu)]=0, \qquad H=\frac{JN}{2}I-\frac{iJ}{2}\tau_0^{-1}\tau'_0, \qquad \tau_0=\tau(i/2),$$

where the prime differentiates with respect to $\lambda$ at $i/2$. The commuting-transfer proof establishes both identities, including invertibility at that point, the periodic closing bond and the sign of the Hamiltonian. It follows that $[H,\tau(\lambda)]=0$.

This construction uses the same local tensor rule for every finite chain length; it does not first require diagonalizing the whole Hamiltonian to define its projectors. It is the specific algebraic structure meant in these articles. Which derived charges are independent, which have local densities, and what persists in an infinite-chain limit are further questions.

A Bethe construction produces eigenvalues and nonzero eigenvectors from admissible parameters. Completeness additionally requires those states, with any necessary limiting or symmetry constructions, to span the stated Hilbert space or sector. A list of solutions of algebraic Bethe equations is not automatically a basis. The singular-root benchmark shows why this distinction matters: the formally singular pair $\pm i/2$ requires a specified limiting vector and physicality check even in a short chain.

The four-site XXX sector supplies a complete finite example: six orthonormal states resolve the sector identity. Omitting its singular state leaves every retained eigenvector equation and every distinct energy intact, while losing one dimension of the span.

A stationary law leaves dynamical questions open

For a finite continuous-time Markov chain, use probability columns and

$$\dot p=Qp, \qquad Q_{ij}\ge0\ (i\ne j), \qquad \sum_i Q_{ij}=0.$$

A stationary probability vector satisfies $Q\pi=0$, $\pi_i\ge0$, and $\sum_i\pi_i=1$. This describes an invariant distribution. It does not specify the remaining eigenvalues or the evolution from another initial distribution.

For example, every member of the two-state family

$$Q_r=r\begin{pmatrix}-1&1\\1&-1\end{pmatrix}, \qquad r\gt0,$$

has the same stationary vector $\pi=(1/2,1/2)^T$. Yet with $p(0)=(1,0)^T$,

$$p_r(t)=\frac12 \begin{pmatrix}1+e^{-2rt}\\1-e^{-2rt}\end{pmatrix}, \qquad t\ge0.$$

The stationary component has eigenvalue zero; the difference of the two probabilities decays with eigenvalue $-2r$. Direct differentiation verifies $\dot p_r=Q_rp_r$. The rate $r$ has units of inverse time. The identical stationary law therefore coexists with different relaxation times.

Writing $p(t)=e^{tQ}p(0)$ is valid for every finite generator and, by itself, does not distinguish an integrable family. The finite-ring TASEP lesson derives a uniform stationary law and its current; the coordinate-Bethe article and computation lab also establish a particular dynamical mode. Each result has its own initial conditions, sector and observable. None of those results alone is a complete formula for all stochastic questions.

Numerical evidence has a declared target

A small commutator residual checks chosen matrices at chosen parameters and precision. A finite spectrum checks that finite system. Level-spacing or level-crossing patterns are diagnostics with their own symmetry and sampling requirements; they cannot serve as universal definitions. Caux and Mossel discuss the limitations of such criteria in § 3, printed pp. 6–7, arXiv v1 PDF.

For a useful integrability statement, name the model and regime, the structure proved, and the calculation it enables. “The periodic XXX transfer matrices commute” and “this reflectionless KdV field reconstructs two scattering poles” are assessable claims. Increasing their scope to all eigenstates, all initial data or all observables requires the corresponding additional argument.

References

• Aktosun, Tuncay. “Inverse Scattering Transform and the Theory of Solitons.” In Encyclopedia of Complexity and Systems Science, edited by Robert A. Meyers, pp. 4960–4971. Springer, 2009. DOI. Author version arXiv:0905.4746v1; HTML. The reconstruction discussion uses §§ III and VIII–IX, with the field-sign conversion stated above.
• Cannas da Silva, Ana. Lectures on Symplectic Geometry. Lecture Notes in Mathematics 1764. Springer, 2001. DOI. Author revision January 2006, open PDF. Locators refer to § 18.4, pp. 109–111 in that revision.
• Caux, Jean-Sébastien, and Jorn Mossel. “Remarks on the notion of quantum integrability.” Journal of Statistical Mechanics: Theory and Experiment (2011), P02023. DOI. Author version arXiv:1012.3587v1, submitted 16 December 2010; open PDF. Citations refer to § 3, printed pp. 4–7 of that version, rather than the PDF’s later compilation date.