Symplectic Eigenvalues

• Symplectic eigenvalues are canonical invariants defined via Williamson’s theorem, encapsulating the intrinsic structure of positive-definite matrices under symplectic transformations.
• They play a critical role in continuous-variable quantum information, Hamiltonian dynamics, and operator theory, informing variational principles and optimization methods.
• Recent advancements extend symplectic eigenvalue analysis to infinite-dimensional operators, optimizing numerical methods and refining inequality frameworks across diverse applications.

A symplectic eigenvalue is a canonical invariant arising from the diagonalization of real positive-definite matrices (and certain classes of operators) under congruence by symplectic transformations, as formalized by Williamson’s theorem. Symplectic eigenvalues encode structure that is central in areas ranging from continuous-variable quantum information theory to classical Hamiltonian systems, operator theory, majorization theory, and Riemannian geometry on matrix spaces. This entry synthesizes recent developments encompassing finite and infinite-dimensions, variational and majorization principles, optimization, sensitivity theory, and connections to operator inequalities.

1. Definition, Characterizations, and Williamson’s Theorem

Let $A \in \mathbb{R}^{2n \times 2n}$ be symmetric positive-definite and let $$J := \begin{pmatrix} 0 & I_n \ -I_n & 0 \end{pmatrix}$$. The real symplectic group $\mathrm{Sp}(2n)$ consists of those $S$ with $S^T J S = J$.

Williamson’s Theorem (1936): There exists $S \in \mathrm{Sp}(2n)$ and $D = \mathrm{diag}(d_1, \ldots, d_n)$ with $d_1 \leq \cdots \leq d_n > 0$ such that

$$S^T A S = \begin{pmatrix} D & 0 \ 0 & D \end{pmatrix}.$$

The unique positive numbers $d_1, \ldots, d_n$ are the symplectic eigenvalues of $A$ (Bhatia et al., 2018, Son, 9 Jun 2025).

Equivalent characterizations include:

• $iJA$ has eigenvalues $\pm d_1, \ldots, \pm d_n$.
• $i A^{1/2} J A^{1/2}$ is Hermitian with spectrum $\{\pm d_1, \ldots, \pm d_n\}$ (Mishra, 2020).
• Each symplectic eigenvalue $d_i$ appears with multiplicity two in the spectrum of $A$.
• For positive-semidefinite $A$, the notion generalizes if $\ker A$ is a symplectic subspace (Son et al., 2022).

Symplectic eigenvectors come in pairs: $A u_i = d_i J v_i$, $A v_i = -d_i J u_i$ with $u_i^T J v_j = \delta_{ij}$.

In the operator-theoretic infinite-dimensional setting, if $T$ is a positive, invertible operator on $\mathcal{H} \oplus \mathcal{H}$ (with $T - \alpha I$ compact), the symplectic spectrum of $T$ consists of the positive eigenvalues of $M := i \sqrt{T} J \sqrt{T}$ and is at most countable, accumulating only at $\alpha$ (John et al., 2022, Kumar et al., 2023).

2. Variational Principles and Optimization

Symplectic eigenvalues support variational characterizations mirroring classical eigenvalue theory:

• Wielandt-Type Min-Max Principle: The sum of selected symplectic eigenvalues is given by a max-min formula over symplectically orthonormal pairs in nested subspaces (Jain, 2021).
• Trace-Minimization Theorem: For $k \leq n$,

$$\sum_{j=1}^k d_j(A) = \min_{S \in \mathrm{Sp}(2n, 2k)} \frac{1}{2} \operatorname{tr}(S^T A S),$$

where $\mathrm{Sp}(2n, 2k)$ is the symplectic Stiefel manifold, and the minimum is achieved by symplectic eigenvector sets (Son et al., 2021, Son, 9 Jun 2025, Son et al., 2022).

• Critical points of the Brockett-type or trace objectives under symplecticity constraints correspond to symplectic eigenvectors, and Riemannian optimization methods (e.g., trust-region, gradient descent on symplectic Stiefel manifolds) are effective for large-scale computation (Son et al., 2021, Son, 9 Jun 2025).

For infinite-dimensional operators, finite-rank truncations converge to the symplectic spectrum under mild hypotheses, with explicit control available for certain operator classes (e.g., block diagonal or block circulant forms) (Kumar et al., 2023).

3. Inequalities: Majorization, Interlacing, and Schur–Horn Theory

Fundamental inequalities for symplectic eigenvalues have close analogs to—but often crucial differences from—those for ordinary Hermitian eigenvalues.

Majorization and Supermajorization

For $A \in \mathbb{R}^{2n \times 2n}$ positive-definite, let $d(A) = (d_1(A),...,d_n(A))$:

• Weak Supermajorization (Schur–Horn for Symplectic Eigenvalues):

$\Delta^s(A) \prec^w d(A),$

where $\Delta^s_j(A) = \sqrt{[A_{11}]_{jj} [A_{22}]_{jj}}$ and blocks are as in $$A = \begin{pmatrix} A_{11} & A_{12} \ A_{12}^T & A_{22} \end{pmatrix}$$ (Bhatia et al., 2020, Huang et al., 2024).

• For fixed $y \in \mathbb{R}^n_{>0}$, all symplectic diagonals $x$ arise from $d(A) = y$ if and only if $x \prec^w y$.
• Full majorization $\prec$ (i.e., equality in all partial sums) holds precisely when $A$ is orthosymplectically diagonalizable (Huang et al., 2024).

Interlacing with Ordinary Eigenvalues

For $A$ real positive-definite with eigenvalues $\lambda_1 \leq \cdots \leq \lambda_{2n}$,

$\lambda_j(A) \leq d_j(A) \leq \lambda_{n+j}(A),$

and $$d(A) \prec_{\log} (\lambda_1(A),..., \lambda_{2n}(A))$$ (Bhatia et al., 2018).

Sums and Products—Multiplicative and Additive Lidskii Inequalities

For $A,B > 0$,

• Symplectic Lidskii Inequality: For any $k$ and index sets,

$$\sum_{r=1}^k d_{i_r}(A+B) \geq \sum_{r=1}^k d_{i_r}(A) + \sum_{r=1}^k d_r(B),$$

with equality if and only if certain common symplectic eigenvector conditions hold (Mishra, 2023, Jain et al., 2020).

• Multiplicative Inequalities: For geometric means,

$d(A \#_t B) \prec_{\log} d(A)^{1-t} d(B)^t,$

and for Riemannian/Cartan barycenters (geometric means of random matrices),

$\log d(G(\mu)) \prec_w \int \log d(A) \, d\mu(A).$

(Hiai et al., 2017, Bhatia et al., 2018, Jain, 2021)

Horn-Type and Trace Inequalities

The Horn cone for symplectic eigenvalues satisfies all classical Horn inequalities, except trace equality $$\operatorname{Tr}(C) = \operatorname{Tr}(A) + \operatorname{Tr}(B)$$ is replaced by $$\operatorname{Tr}(C) \geq \operatorname{Tr}(A) + \operatorname{Tr}(B)$$ (Paradan, 2022).

4. Analyticity, Sensitivity, and Differential Properties

• Continuity: The maps $A \mapsto d_j(A)$ are continuous on the positive-definite cone.
• Differentiability: The functions $d_m(A)$ are not everywhere differentiable (unless the eigenvalue is simple), but always directionally differentiable. If $d_m(A)$ is simple:

$$D d_m(A)[E] = -\frac{1}{2} \langle x x^T + y y^T, E \rangle,$$

where $(x, y)$ is the corresponding normalized symplectic eigenpair of $A$ (Mishra, 2020, Jain et al., 2020).

• Subdifferential Structure: Clarke and Michel–Penot subdifferentials of $-d_m$ are convex hulls of matrices of the form $-\frac{1}{2}(x x^T + y y^T)$ for all associated symplectic eigenpairs (Mishra, 2020).
• Analytic Paths: Along $t \mapsto A(t)$ real-analytic, symplectic eigenvalues are piecewise real-analytic; their derivatives obey explicit Rayleigh-quotient-type formulas (Jain et al., 2020).

5. Infinite-Dimensional Extensions

A class of infinite-dimensional positive, invertible operators $$T: \mathcal{H} \oplus \mathcal{H} \rightarrow \mathcal{H} \oplus \mathcal{H}$$, with $T-\alpha I$ compact, admits a symplectic spectrum via the spectrum of $i\sqrt{T} J \sqrt{T}$. Main results include:

• Symplectic eigenvalues to the right (resp. left) of $\alpha$ are always less than (resp. greater than) the corresponding ordinary eigenvalues (John et al., 2022).
• For infinite-mode bosonic Gaussian covariance operators, symplectic eigenvalues determine entropic and operational quantities; finite-rank truncations converge to the true spectrum under mild assumptions (Kumar et al., 2023).
• The symplectic Szegő theorem gives limiting distributions for symplectic spectra of block Toeplitz operators and provides entropy rates for stationary quantum Gaussian processes (Bhatia et al., 2020).

6. Applications, Numerical Aspects, and Generalizations

Symplectic eigenvalues are indispensable in quantum optics, continuous-variable quantum information (notably for Gaussian states), Hamiltonian dynamics, operator theory, and sensitivity analysis. Key contexts include:

• Gaussian Covariance Matrices: A $2n \times 2n$ covariance matrix is physical if and only if all symplectic eigenvalues satisfy $d_j \geq \frac{1}{2}$, with explicit entropic formulas and operational relevance (Bhatia et al., 2018, Kumar et al., 2023, Bhatia et al., 2020).
• Fisher Information: Symplectic decomposition provides insight into sensitivity for parameter pairs (e.g., $(q_j, p_j)$), complementing ordinary eigenanalysis (Yang, 2022).
• Optimization: Riemannian and symplectic-structure-preserving algorithms target minimal symplectic eigenvalues and eigenvectors, with competitive computational properties (Son et al., 2021, Son, 9 Jun 2025).
• Log-Majorization and Cartan Means: Symplectic eigenvalue maps are monotone, homogeneous, and Lipschitz for trace metric; log-majorization bounds are preserved under Riemannian/Cartan barycenters (Hiai et al., 2017).

7. Open Problems and Outlook

Notable open directions include:

• Generalization of interlacing and majorization inequalities to broader or singular infinite-dimensional settings (e.g., non-invertible operators, uncountable spectrum).
• Characterization of equality cases for symplectic-analytic inequalities beyond Schur–Horn, Ky Fan, and Lidskii types (Huang et al., 2024, Mishra, 2023).
• Deeper understanding of variational and entropy-like principles for symplectic spectra in non-classical or infinite-component systems.
• Extension of truncation and approximation methods to new classes of infinite-dimensional quantum systems (John et al., 2022, Kumar et al., 2023).

Symplectic eigenvalues thus constitute a rich, unifying invariant at the interface of algebraic, analytic, and geometric analysis, with ongoing developments in optimization, infinite-dimensional analysis, and quantum information theory.