---
title: Dual Lagrangian Products in Symplectic Geometry
url: https://www.emergentmind.com/topics/dual-lagrangian-products
type: topic
---

# Dual Lagrangian Products in Symplectic Geometry

to=arxiv_search 夫妻性生活影片  төлөв  天天中彩票appion code 一本道高清无码
{"query":"Dual Lagrangian Products symplectic capacities Lagrangian products duality arXiv","max_results":10,"sort_by":"submittedDate"}
Dual Lagrangian products are most explicitly realized in symplectic geometry as Lagrangian products of the form \(K\times_L K^\circ \subset \mathbb{R}^n_x\times_L \mathbb{R}^n_y\), where \(K\subset\mathbb{R}^n\) is a convex body containing the origin and \(K^\circ\) is its polar dual. More generally, the same literature studies \(K\times_L T\) for arbitrary convex bodies \(K,T\), as well as functionally dual variants such as \(K_\Phi\times_L K_{\Phi^*}\) built from Young functions and their Legendre transforms. In broader mathematical and optimization usage, closely related constructions treat “duality” as a bilinear pairing, a coupling function, a doubled relation, or a factored dual variable, all of which generate Lagrangian or dual objects by product-like operations [2507.20222], [2505.07572].

## 1. Symplectic definition and canonical models

The ambient symplectic space is the standard vector space
\[
(\mathbb{R}^{2n},\omega_0),\qquad \omega_0=\sum_{i=1}^n dy_i\wedge dx_i,
\]
identified with the cotangent bundle \(T^*\mathbb{R}^n\cong \mathbb{R}^n_x\times \mathbb{R}^n_y\). For subsets \(A\subset \mathbb{R}^n_x\) and \(B\subset \mathbb{R}^n_y\), the Lagrangian product is
\[
A\times_L B := A\times B \subset \mathbb{R}^{2n}.
\]
The notation emphasizes that the two factors lie in dual Lagrangian directions, namely configuration space and momentum space. A dual Lagrangian product is then a domain
\[
K\times_L K^\circ,
\]
with \(K\) convex and \(K^\circ\) its polar dual [2507.20222].

Two model examples recur throughout the literature. For the Euclidean ball
\[
D^n(r):=\{x\in\mathbb{R}^n\mid x_1^2+\dots+x_n^2\le r^2\},
\]
the polar dual of \(D^n(1)\) is again \(D^n(1)\), so the Lagrangian bidisk \(D^n(1)\times_L D^n(1)\) is a dual Lagrangian product. For the unit cube \(\square^n(1,\dots,1)\), the polar dual is the cross-polytope \(\diamond^n(1)\), yielding the cube–cross-polytope product \(\square^n(1,\dots,1)\times_L\diamond^n(1)\) [2507.20222].

A related basic object is the disk cotangent bundle \(A\times_L D^n(1)\). This suggests that dual Lagrangian products belong to a larger class of phase-space domains whose geometry is controlled by a convex body in the base and another convex body in the fiber. In the special dual case \(K\times_L K^\circ\), the interaction between the two factors is fixed by convex polarity rather than chosen independently.

## 2. Capacities, billiards, and Lagrangian barriers

The modern symplectic theory of dual Lagrangian products is organized around symplectic capacities. For any normalized symplectic capacity \(c\), the Lagrangian bidisk satisfies
\[
c(D^n(1)\times_L D^n(1))=4,
\]
and the cube–cross-polytope product satisfies
\[
c(\square^n(1,\dots,1)\times_L\diamond^n(1))=4,
\]
the latter because it is symplectomorphic to \(B^{2n}(4)\) [2507.20222].

Puncturing the base factor produces a sharp capacity drop in the model cases. For the annular bidisk,
\[
c\big((D^n(1)\setminus D^n(\delta))\times_L D^n(1)\big)=2(1-\delta),
\]
and for the punctured cube–cross-polytope product,
\[
c\big((\square^n(1,\dots,1)\setminus\{0\})\times_L\diamond^n(1)\big)=2.
\]
Thus the full products have capacity \(4\), while removing the Lagrangian plane over the origin halves the capacity in the punctured case. This underlies the notion of a Lagrangian barrier: the subsets \(\{0\}\times_L D^n(1)\) and \(\{0\}\times_L\diamond^n(1)\) obstruct symplectic embeddings that would otherwise fit into the unpunctured domain [2507.20222].

For a convex centrally symmetric body \(K\), the general holed inequality is
\[
c\big((K\setminus \delta K)\times_L K^\circ\big)\ge \frac{1-\delta}{2}\,c(K\times_L K^\circ),
\qquad 0\le \delta<1.
\]
For arbitrary convex \(K\) containing the origin, with
\[
\widehat{k}:=\max_{x\in\partial K}\|x\|,\qquad
\widecheck{k}:=\min_{x\in\partial K}\|x\|,
\]
normalized capacities satisfy
\[
4\sqrt{\frac{\widecheck{k}}{\widehat{k}}}\le c(K\times_L K^\circ)\le 4\sqrt{\frac{\widehat{k}}{\widecheck{k}}},
\]
and
\[
2\sqrt{\frac{\widecheck{k}}{\widehat{k}}}\le c\big((K\setminus\{0\})\times_L K^\circ\big)\le 2\sqrt{\frac{\widehat{k}}{\widecheck{k}}}.
\]
In dimension two, if \(K\) is symmetrically convex and \(\alpha\)-pinched with
\[
\alpha=\left(\frac{4}{\pi}\right)^{2/3}\approx 1.1751,
\]
then
\[
c\big((K\setminus\{0\})\times_L K^\circ\big)
<
\sqrt{\operatorname{Vol}(K)\operatorname{Vol}(K^\circ)}
\le
c(K\times_L K^\circ),
\]
so \(\{0\}\times_L K^\circ\) is a Lagrangian barrier in \(K\times_L K^\circ\) [2507.20222].

A second major structural result identifies symplectic capacity with billiard dynamics. For arbitrary convex bodies \(K,T\subset\mathbb{R}^n\), the Ekeland–Hofer–Zehnder capacity of the convex Lagrangian product \(K\times T\subset\mathbb{R}^n_q\times\mathbb{R}^n_p\) satisfies
\[
c_{EHZ}(K\times T)
=
\min_{q\in F_{n+1}^{cp}(K)}\ell_T(q)
=
\min_{p\in F_{n+1}^{cp}(T)}\ell_K(p)
=
\min_{q\in M_{n+1}(K,T)} \ell_T(q).
\]
Here \(\ell_T\) is the \(T\)-length induced by the gauge of \(T^\circ\), \(F_{n+1}^{cp}(K)\) is the set of closed polygonal curves with at most \(n+1\) vertices that cannot be translated into \(\operatorname{int}K\), and \(M_{n+1}(K,T)\) is the set of closed strong \((K,T)\)-Minkowski billiard trajectories with at most \(n+1\) bouncing points. This result removes the smoothness and strict convexity assumptions that had previously been required, and it applies in particular to the dual case \(K\times K^\circ\) [2203.01718].

## 3. Functional duality and polar duality in Young-function products

A distinct but closely related theory compares two notions of duality for Lagrangian products generated by an \(n\)-tuple of Young functions \(\Phi=(\Phi_1,\dots,\Phi_n)\). The associated Luxemburg unit ball is
\[
K_\Phi=\left\{x\in\mathbb{R}^n\;\middle|\;\Phi_1(|x_1|)+\dots+\Phi_n(|x_n|)\le 1\right\}.
\]
There are then two dual constructions. The functional-dual product is
\[
K_\Phi\times_L K_{\Phi^*},
\]
where \(\Phi^*=(\Phi_1^*,\dots,\Phi_n^*)\) is the Legendre dual \(n\)-tuple. The polar-dual product is
\[
K_\Phi\times_L K_\Phi^\circ,
\]
where \(K_\Phi^\circ\) is the polar body of \(K_\Phi\) [2505.07572].

For the functional-dual product, all normalized symplectic capacities agree exactly. Writing
\[
\alpha_i := \Phi_i^{-1}(1)\,(\Phi_i^*)^{-1}(1),\qquad
\alpha_{\min}:=\min_i \alpha_i,
\]
one has
\[
c\bigl(K_\Phi\times_L K_{\Phi^*}\bigr)=4\,\alpha_{\min}.
\]
Since \(\alpha_i\le 2\), this yields
\[
c\bigl(K_\Phi\times_L K_{\Phi^*}\bigr)\le 8.
\]
This is an exact strong-Viterbo-type statement for the full class of functional-dual products [2505.07572].

For the polar-dual product, the general result is weaker but still explicit:
\[
2\alpha_{\min}\le c\bigl(K_\Phi\times_L K_\Phi^\circ\bigr)\le 4.
\]
Because \(\alpha_{\min}>1\), this implies
\[
c\bigl(K_\Phi\times_L K_\Phi^\circ\bigr)>2.
\]
Under the additional condition that there is an index \(I\) such that \(\alpha_{\min}=\Phi_I^{-1}(1)(\Phi_I^*)^{-1}(1)\) and \(\Phi_I'(1)=1\), the bounds match and
\[
c(K_\Phi\times_L K_\Phi^\circ)=4
\]
for every normalized symplectic capacity [2505.07572].

Two examples illustrate the dichotomy. If \(\Phi_i(t)=t^p/p\) and \(\Phi_i^*(t)=t^q/q\) with \(1/p+1/q=1\), then
\[
c\bigl(K_\Phi\times_L K_{\Phi^*}\bigr)=4\,p^{1/p}q^{1/q},
\]
and after the natural homothety this recovers
\[
c\bigl(D_p(1)\times_L D_q(1)\bigr)=4.
\]
If \(\Phi_i(t)=(e^t-1)/e\), then
\[
c\bigl(K_\Phi\times_L K_{\Phi^*}\bigr)
=
4\,e^{W(1-1/e)}\log(e+1),
\]
while the polar-dual product still satisfies
\[
c(K_\Phi\times_L K_\Phi^\circ)=4
\]
because \(\Phi_i'(1)=1\) [2505.07572].

## 4. Duality as coupling: Lagrangians, perturbations, and generalized convexity

Outside symplectic convexity, a more abstract interpretation treats the “dual product” as the pairing or coupling that links primal and dual variables inside a Lagrangian. In Rockafellarian duality, the basic objects are an optimization problem
\[
\inf_{x\in\mathcal X} f(x),
\]
a perturbation or Rockafellian
\[
R:\mathcal X\times\mathcal U\to\bar{\mathbb R},\qquad f(x)=R(x,0),
\]
a perturbation function
\[
\varphi(u)=\inf_{x\in\mathcal X}R(x,u),
\]
and a coupling \(c:\mathcal U\times\mathcal V\to\bar{\mathbb R}\), which reduces to the bilinear pairing \(\langle u,v\rangle\) in the classical Fenchel setting [2211.13083].

Given a Rockafellian \(R\), the associated Lagrangian is
\[
L(x,v)=\inf_{u\in\mathcal U}\bigl\{R(x,u)\UppPlus(-c(u,v))\bigr\},
\]
and the dual function is
\[
\psi(v)=\inf_{x\in\mathcal X}L(x,v).
\]
Conversely, given \(L\), one reconstructs a Rockafellian by
\[
R(x,u)=\sup_{v\in\mathcal V}\bigl\{L(x,v)\LowPlus c(u,v)\bigr\}.
\]
For each fixed \(x\), the partial functions satisfy
\[
-L_x = R_x^c,\qquad R_x = (-L_x)^{c'},
\]
and the perturbation and dual functions obey
\[
-\psi=\varphi^c
\]
when one starts from \(R\), while in the reverse direction only
\[
\varphi\ge (-\psi)^{c'}
\]
is guaranteed [2211.13083].

The paper formalizes this symmetry through Lagrangian–Rockafellian couples \((L,R)\), defined by the minimality of \((-L,R)\) in
\[
(-L(x,v))\UppPlus R(x,u)\ge c(u,v).
\]
Its main theorem gives equivalent characterizations by mutual inf–sup formulas, by conjugacy,
\[
-L_x=R_x^c,\qquad R_x=(-L_x)^{c'},
\]
and by generalized convexity through biconjugacy. In this language, a “dual Lagrangian product” is best understood not as a Cartesian product of bodies, but as the coupling \(c(u,v)\) or \(\langle u,v\rangle\) that generates dual Lagrangians, perturbation functions, and conjugate value functions [2211.13083].

## 5. Computational and optimization reinterpretations

In computational optimization, product-type dual Lagrangian structures appear as explicit parametrizations of dual variables. In semidefinite programming, one such representation replaces the dual positive-semidefinite variable \(Z\in\mathbb S^n_+\) by a factorization
\[
Z=VV^\top,\qquad V\in\mathbb R^{n\times r},
\]
which eliminates the explicit cone constraint and turns the dual augmented-Lagrangian subproblem into an unconstrained but nonconvex maximization in \((y,V)\). The resulting augmented Lagrangian is
\[
L_\sigma(y,V;X)
=
b^Ty-\langle VV^\top-C+\mathcal A^T y,X\rangle
-\frac{\sigma}{2}\|VV^\top-C+\mathcal A^T y\|^2,
\]
and the stationarity condition with respect to \(V\) is
\[
-2\Big(X+\sigma(VV^\top-C+\mathcal A^T y)\Big)V=0.
\]
This is the core “dual product” representation in that setting: the dual matrix is encoded as the product \(VV^\top\) inside the augmented Lagrangian [1710.04869].

For multistage stochastic mixed-integer programming, Lagrangian Dual Decision Rules restrict the multiplier policies rather than the primal policies. In the stagewise dual, the multiplier–constraint product is
\[
\pi_t(\boldsymbol{\xi}^t)^\top\big(A_t(\boldsymbol{\xi}^t)x_t(\boldsymbol{\xi}^t)+B_t(\boldsymbol{\xi}^t)x_{t-1}(\boldsymbol{\xi}^{t-1})-b_t(\boldsymbol{\xi}^t)\big),
\]
while in the nonanticipative dual it is
\[
\gamma_t(\boldsymbol{\xi})^\top\big(y_t(\boldsymbol{\xi})-\mathbb E[y_t(\boldsymbol{\xi})\mid \boldsymbol{\xi}^t]\big).
\]
Restricting \(\pi_t\) and \(\gamma_t\) to basis-function families yields finite-dimensional restricted duals, and the restricted nonanticipative dual can be at least as strong as the restricted stagewise dual when its basis is chosen appropriately [2001.00761].

Constraint programming uses essentially the same decomposition principle. Variables are duplicated per constraint, and linking equalities \(X_i=X_1\) are relaxed with multipliers \(\mu_i\), giving a dual bound
\[
\mathcal B(\mu)=\Phi(X_1)+\sum_{i=2}^m \Psi_i(X_i),
\]
with
\[
\Phi(X_1)=\max\Big\{f(X_1)+\Big(\sum_{i=2}^m \mu_i\Big)\cdot X_1\ \Big|\ C_1(X_1)\Big\},
\]
and
\[
\Psi_i(X_i)=\max\{-\mu_i\cdot X_i\mid C_i(X_i)\}.
\]
A self-supervised graph neural network is then trained to predict multipliers \(\mu=\Omega_\theta(G)\) directly, while preserving validity of the dual bound because any multiplier vector defines a legitimate Lagrangian relaxation [2408.12695].

Sparse linear programming yields yet another explicit dual formula. For
\[
\min\{c^Tx\mid Ax=b,\ 0\le x\le l,\ \|x\|_0\le r\},
\]
the Lagrangian dual is
\[
\max_{y\in\mathbb R^m}
\;
b^Ty-\|\Pi_{S(r)\cap\mathbb R^n_+}(l\circ(A^Ty-c))\|_1,
\]
an unconstrained piecewise-linear convex program. Its nonsmooth term is governed by the projection onto the sparse nonnegative set and therefore by a vector Ky-Fan norm structure. The paper proves strong duality and uses a semi-proximal ADMM on a split form of the dual to recover primal optimal solutions [1805.11942].

## 6. Categorical, geometric, and representation-theoretic extensions

A categorical version of dual Lagrangian products appears in the theory of Lagrangian relations. Over a field \(k\), the prop \(\LagRel_k\) has objects \(k^{2n}\) with the standard symplectic form and morphisms given by Lagrangian subspaces of the corresponding input–output phase space. Its symmetric monoidal product is direct sum, and a linear relation \(V\) generates a pure Lagrangian relation by doubling:
\[
V \longmapsto V^\perp\oplus V.
\]
The category is described as a doubled, CPM-style category of linear relations, with orthogonal complement playing the role of conjugation; for prime fields, \(\LagRel_{F_p}\) is exactly a CPM construction on linear relations [2105.06244].

Derived symplectic geometry supplies a higher product construction. If \(L_1,\dots,L_m\to S\) are Lagrangians in an \(n\)-shifted symplectic derived stack, then the \(m\)-fold fiber product
\[
L_1\times_S\cdots\times_S L_m
\]
is itself Lagrangian in the cyclic product
\[
(L_1\times_S L_2)\times(L_2\times_S L_3)\times\cdots\times(L_m\times_S L_1).
\]
Here the target is a product of pairwise intersections, each carrying an \((n-1)\)-shifted symplectic structure, so the multi-intersection becomes a Lagrangian object in a product of dual pairings [1309.0596].

Lie-theoretic mechanics produces another product interpretation. For a Lie group \(G\), the iterated tangent group satisfies
\[
TTG\cong \mathfrak g\bowtie T^2G,
\]
and for a matched pair \(G\bowtie H\) one has
\[
T^2(G\bowtie H)\cong T^2G\bowtie T^2H.
\]
This yields second-order Euler–Lagrange and matched Euler–Poincaré systems on double cross product groups, where two Lagrangian subsystems mutually act on one another rather than appearing in a one-sided semidirect product [1909.10375].

Product ambient manifolds also support rigid Lagrangian geometries. In the para-Kähler products of Lorentzian surfaces \((\Sigma_1\times\Sigma_2,G^\varepsilon,J,\Omega^\varepsilon)\), any Lagrangian surface with parallel mean curvature vector is, under the paper’s curvature assumptions, locally a product \(\gamma_1\times\gamma_2\) of constant-curvature curves, while in \(d\mathbb S^2\times d\mathbb S^2\) any \(\Omega^+\)-Lagrangian with non-null parallel mean curvature is again locally such a product [1403.0305]. In the Kähler product \(\mathbb H^2\times\mathbb H^2\), the key scalar invariant \(T\) satisfies \(0\le T^2\le 1/4\); the case \(T^2=0\) characterizes products of curves, and \(T^2=1/4\) characterizes the diagonal immersion. Under constant curvature, parallel second fundamental form, or completeness plus a strict bound \(T^2\le 1/4-\varepsilon\), the only Lagrangian surfaces are products of geodesics or the diagonal [2106.13975].

Finally, derived-category theory on Lagrangian Grassmannians exhibits a representation-theoretic duality of blocks rather than a symplectic product of domains. On \(LG(n,2n)\), the Kuznetsov–Polishchuk exceptional blocks \(B_h\cong Y_{h,n-h}\) admit graded left dual exceptional collections indexed by transposed Young diagrams, and the block decomposition of \(D^b(LG(n,2n))\) becomes self-dual under dualization and twist. This suggests that “dual Lagrangian product” language can extend even to settings where the relevant product is combinatorial or categorical rather than geometric [2301.02941].

Source: https://www.emergentmind.com/topics/dual-lagrangian-products