Introduction
Models of pedestrian dynamics have applications in urban planning, event management, and public safety. A key goal is to understand and optimize stationary crowd flows—persistent, large-scale movement patterns that are statistically constant over time. Such flows are described by a macroscopic density m ( x ) m(x) m ( x ) and a velocity field v ( x ) v(x) v ( x ) , governed by the continuity equation
div ( m v ) = 0 , \operatorname{div}(mv) = 0, div ( m v ) = 0 ,
which expresses conservation of agents.
Crowd flow models include the Hughes continuum framework [11 ] and its analytical and numerical developments such as in [9 ] and references therein. Mean-field-game formulations account for anticipation, crowd aversion, and congestion [15 ] , while optimal-control formulations incorporate nonlinear mobility in evacuation problems [3 ] . Non-separable mean-field games extend the Hughes framework to describe movement through or around congested regions [7 ] . For dense crowds, density-constrained models restrict motion through admissible-velocity projections [16 ] , while hybrid agent–continuum simulations use unilateral incompressibility to model crowding and facilitate collision avoidance [17 ] . A related granular approach obtains congested crowd motion as the p → ∞ p \to\infty p → ∞ limit of Hele–Shaw-type problems with p p p -Laplacian diffusion and drift [12 ] .
Variational formulations for first-order stationary mean-field games with congestion provide a basis for existence results and numerical approximation [5 ] . For stationary crowd entry–exit dynamics on bounded domains, mixed boundary conditions prescribe inflow and regulate exit through an exit-cost complementarity condition, while a variational approach yields existence of solutions [1 ] . Prescribed boundary inflow has also been studied in time-dependent first-order mean-field games through a relaxed variational formulation and Fenchel–Rockafellar duality [8 ] . In these formulations, costs depend on density and individual motion rather than directly on the symmetric velocity gradient. More generally, continuity and incompressibility constraints alone do not penalize local shear or directional compression.
Anticipatory pedestrian models use projected time to collision [14 ] , while vision-based kinetic and macroscopic models use bearing-angle changes and time to interaction [4 ] . Kinetic models also describe collision avoidance through sidestepping [6 ] , while microscopic and macroscopic contact models estimate the forces and pressure generated in dense crowds [13 ] . In the deterministic kinematic interpretation adopted here, agents in a smooth stationary velocity field follow the streamlines x ˙ ( t ) = v ( x ( t ) ) \dot{x}(t)=v(x(t)) x ˙ ( t ) = v ( x ( t )) , and collision risk refers to the instantaneous approach of nearby agents, rather than intersections of point trajectories. Such approach depends on the local directional compression and shear encoded in the velocity gradient and can occur even at moderate density. At the same density, rigid rotation produces no local approach, while shear and directional compression can do so. This motivates using the symmetric velocity gradient to compare stationary flow geometries under prescribed boundary flux.
We introduce a collision functional and develop a variational framework for its minimization, addressing two questions: how should collision intensity be quantified from the pair ( m , v ) (m,v) ( m , v ) , and, in the constant-density regime under prescribed boundary flux, which incompressible flow geometries minimize it? Our derivation of the local collision intensity adapts the classical Saffman–Turner inward-relative-flux construction for passive particles in fluid flows [19 ] and its two-dimensional eigenvalue evaluation in [10 Appendix A.1] . In the aerosol and turbulence literature, this construction is used to estimate collision rates of finite-radius particles advected in random velocity fields. The Saffman–Turner approximation describes an initial transient; long-time first-contact rates must also account for trajectory re-entry and, in compressible flows, clustering [10 ] . Here, by contrast, we pass to the small-radius limit to define a pointwise encounter-intensity density for a deterministic stationary flow on a bounded domain, and study the variational boundary-value problem of minimizing the resulting total collision functional under prescribed boundary flux.
While the collision functional is derived for a general density m m m and velocity field v v v , our variational existence and relaxation theory focuses on constant-density incompressible flows (m ≡ 1 m \equiv1 m ≡ 1 , div v = 0 \operatorname{div}v=0 div v = 0 ) on bounded simply connected Lipschitz planar domains. By introducing a stream function ψ \psi ψ such that v = D ⊥ ψ v=D^{\perp}\psi v = D ⊥ ψ , we show that the dense collision functional becomes, up to the constant factor 1 / π 1/\pi 1/ π , the convex functional
ψ ↦ ∫ Ω 4 ψ y 2 + ( ψ x x − ψ y y ) 2 d x d y , \psi\mapsto\int_{\Omega} \sqrt{4\psi_y^2+(\psi_{xx}-\psi_{yy})^2}\,dx\,dy, ψ ↦ ∫ Ω 4 ψ y 2 + ( ψ xx − ψ y y ) 2 d x d y ,
to be minimized over stream functions with the boundary trace prescribed by the flux. Under a compatible boundary flux and a finite-energy lifting condition (Assumptions 2 and 3), we establish the existence of relaxed minimizers and identify the relaxed functional with the lower semicontinuous envelope of the unrelaxed functional in the strong W 1 , 1 W^{1,1} W 1 , 1 topology. Using convex duality and integration by parts, we also obtain a dual representation of the minimum, weak Euler–Lagrange conditions, and a smooth sufficient optimality criterion involving a natural boundary condition. We apply these conditions to prove that radial flow is a global minimizer in an annular sector, exhibit a family of minimizers on the square with different peak local collision intensities, and show how domain geometry rules out optimality of the saddle flow on a disk.
The proof of existence relies on the direct method in the calculus of variations. The energy controls T D 2 ψ = ( 2 ψ x y , ψ x x − ψ y y ) TD^{2}\psi=(2\psi_{xy},\psi_{xx}-\psi_{yy}) T D 2 ψ = ( 2 ψ x y , ψ xx − ψ y y ) , but not the L 1 L^{1} L 1 norm of the full Hessian, and is therefore not coercive in W 2 , 1 ( Ω ) W^{2,1}(\Omega) W 2 , 1 ( Ω ) ; see [18 ] . Since the controlled derivatives may concentrate along bounded-energy sequences, we relax the problem by allowing T D 2 ψ TD^{2}\psi T D 2 ψ to be a vector-valued Radon measure.
The prescribed scalar trace removes every zero-energy perturbation except on disks, where the zero-trace kernel is one-dimensional and corresponds to rigid rotation about the center. We prove W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) -coercivity after subtracting the projection onto this kernel. This preserves the prescribed trace and the entire measure T D 2 ψ TD^{2}\psi T D 2 ψ , although it may change the velocity. Using the compact embedding of the space of functions of bounded deformation, B D ( Ω ) ↪ ↪ L 1 ( Ω ; R 2 ) BD(\Omega)\hookrightarrow\hookrightarrow L^{1}(\Omega;\mathbb{R}^{2}) B D ( Ω ) ↪↪ L 1 ( Ω ; R 2 ) , recalled in [2 Theorem 2.10] , we show that relaxed energy sublevels are compact in strong W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) without normalization on domains other than disks, and after normalization on disks. Related relaxation problems for linear-growth symmetric-gradient energies are studied in [2 ] . The dense collision energy is a constant multiple of the total variation of the symmetric velocity gradient; our analysis incorporates incompressibility, prescribed normal flux, and the resulting geometric kernel.
For exact relaxation, we use interior mollification with a locally finite partition of unity. The smoothing scales are chosen so that the difference from the original stream function belongs to W 0 1 , 1 ( Ω ) W^{1,1}_{0}(\Omega) W 0 1 , 1 ( Ω ) , while the collision energy converges. Thus the prescribed trace is preserved exactly, and incompressibility holds automatically throughout the approximation.
The paper is organized as follows. Section 2 fixes notation. Section 3 derives the collision-intensity functional and expresses it in terms of the density and the eigenvalues of the symmetric velocity gradient. Section 4 applies the model to dense incompressible flows, introduces the boundary flux and lifting assumptions, and formulates the stream-function variational problem. Section 5 proves existence of relaxed minimizers, identifies the relaxed problem with the lower semicontinuous envelope of the unrelaxed one, and derives duality and optimality conditions. Section 6 studies radial flow in an annular region, characterizes zero-energy minimizers, exhibits a one-parameter family of minimizers on the square, and compares the saddle flow on a disk.
Notation
Unless otherwise specified, Ω ⊂ R 2 \Omega\subset\mathbb{R}^{2} Ω ⊂ R 2 denotes a bounded simply connected domain with Lipschitz boundary Γ = ∂ Ω \Gamma=\partial\Omega Γ = ∂ Ω , and d s ds d s denotes arclength measure. Along Γ \Gamma Γ , we denote the outer unit normal by ν = ( ν 1 , ν 2 ) \nu=(\nu_{1},\nu_{2}) ν = ( ν 1 , ν 2 ) and the positively oriented unit tangent vector by
τ : = ( − ν 2 , ν 1 ) . \tau:=(-\nu_{2},\nu_{1}). τ := ( − ν 2 , ν 1 ) .
For a scalar function f f f smooth up to Γ \Gamma Γ , we write D f = ( ∂ x f , ∂ y f ) Df=(\partial_{x}f,\partial_{y}f) D f = ( ∂ x f , ∂ y f ) , set
D ⊥ f : = ( ∂ y f , − ∂ x f ) , D^{\perp}f:=(\partial_{y}f,-\partial_{x}f), D ⊥ f := ( ∂ y f , − ∂ x f ) ,
and define the tangential derivative of f f f along Γ \Gamma Γ by
∂ τ f : = D f ⋅ τ = D ⊥ f ⋅ ν . \partial_{\tau}f:=Df\cdot\tau=D^{\perp}f\cdot\nu. ∂ τ f := D f ⋅ τ = D ⊥ f ⋅ ν .
For boundary functions in W 1 , 1 ( Γ ) W^{1,1}(\Gamma) W 1 , 1 ( Γ ) , tangential derivatives are understood weakly with respect to arclength. The positive and negative parts of a real-valued quantity a a a are defined by
a + : = max { a , 0 } , a − : = max { − a , 0 } . a^{+} := \max\{a,0\}, \qquad a^{-} := \max\{-a,0\}. a + := max { a , 0 } , a − := max { − a , 0 } .
Let B r ( x ) B_r(x) B r ( x ) denote the open ball of radius r > 0 r>0 r > 0 centered at x ∈ R 2 x \in\mathbb{R}^{2} x ∈ R 2 . Its boundary ∂ B r ( x ) \partial B_r(x) ∂ B r ( x ) has outer unit normal ν ( y ) = ( y − x ) / r \nu(y) = (y-x)/r ν ( y ) = ( y − x ) / r , and we parametrize ∂ B r ( x ) \partial B_r(x) ∂ B r ( x ) by y = x + r e ( θ ) y=x+r e(\theta) y = x + r e ( θ ) , where
e ( θ ) : = ( cos θ , sin θ ) , θ ∈ [ 0 , 2 π ) . (1) e(\theta) := (\cos\theta,\sin\theta), \qquad\theta\in[0,2\pi). \tag*{(1)} e ( θ ) := ( cos θ , sin θ ) , θ ∈ [ 0 , 2 π ) . (1)
We use standard notation for Lebesgue and Sobolev spaces, L p ( Ω ) L^{p}(\Omega) L p ( Ω ) and W k , p ( Ω ) W^{k,p}(\Omega) W k , p ( Ω ) . On Lipschitz boundaries, traces of W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) functions (denoted simply by ψ \psi ψ on Γ \Gamma Γ ) are understood in the L 1 ( Γ ) L^{1}(\Gamma) L 1 ( Γ ) sense, and W 0 1 , 1 ( Ω ) W^{1,1}_{0}(\Omega) W 0 1 , 1 ( Ω ) denotes the subspace of functions in W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) with zero trace. We use M ( Ω ; R k ) \mathcal{M}(\Omega;\mathbb{R}^{k}) M ( Ω ; R k ) for the space of finite R k \mathbb{R}^{k} R k -valued Radon measures on Ω \Omega Ω . For μ ∈ M ( Ω ; R k ) \mu\in\mathcal{M}(\Omega;\mathbb{R}^{k}) μ ∈ M ( Ω ; R k ) , ∣ μ ∣ |\mu| ∣ μ ∣ denotes its total variation measure and ∣ μ ∣ ( Ω ) |\mu|(\Omega) ∣ μ ∣ ( Ω ) its total mass. Weak-∗ * ∗ convergence in M ( Ω ; R k ) \mathcal{M}(\Omega;\mathbb{R}^{k}) M ( Ω ; R k ) , denoted by ⇀ ∗ \overset{*}{\rightharpoonup} ⇀ ∗ , is understood in duality with C 0 ( Ω ; R k ) C_{0}(\Omega;\mathbb{R}^{k}) C 0 ( Ω ; R k ) . We identify an integrable field f f f with the measure f L 2 f\mathcal{L}^{2} f L 2 , where L 2 \mathcal{L}^{2} L 2 denotes planar Lebesgue measure.
For w ∈ L 1 ( Ω ; R 2 ) w \in L^{1}(\Omega;\mathbb{R}^{2}) w ∈ L 1 ( Ω ; R 2 ) , let E w = ( D w + D w T ) / 2 Ew=(Dw+Dw^{\mathsf{T}})/2 E w = ( D w + D w T ) /2 in the sense of distributions, and set
B D ( Ω ) : = { w ∈ L 1 ( Ω ; R 2 ) : E w ∈ M ( Ω ; R s y m 2 × 2 ) } . BD(\Omega):=\{w\in L^{1}(\Omega;\mathbb{R}^{2}): Ew\in\mathcal{M}(\Omega;\mathbb{R}^{2\times2}_{\mathrm{sym}})\}. B D ( Ω ) := { w ∈ L 1 ( Ω ; R 2 ) : E w ∈ M ( Ω ; R sym 2 × 2 )} .
Here R s y m 2 × 2 \mathbb{R}^{2\times2}_{\mathrm{sym}} R sym 2 × 2 carries the Frobenius norm, and ∥ w ∥ B D ( Ω ) : = ∥ w ∥ L 1 ( Ω ) + ∣ E w ∣ ( Ω ) \|w\|_{BD(\Omega)}:=\|w\|_{L^{1}(\Omega)}+|Ew|(\Omega) ∥ w ∥ B D ( Ω ) := ∥ w ∥ L 1 ( Ω ) + ∣ E w ∣ ( Ω ) .
Throughout, generic positive constants may change from line to line and depend only on the fixed problem data.
The collision functional
Throughout this section, we work under the assumption:
Assumption 1 (Local regularity). The velocity field and density satisfy
v ∈ C l o c 1 , α ( Ω ; R 2 ) , m ∈ C l o c 0 , 1 ( Ω ) , m ≥ 0 , v\in C^{1,\alpha}_{\mathrm{loc}}(\Omega;\mathbb{R}^{2}), \qquad m\in C^{0,1}_{\mathrm{loc}}(\Omega), \qquad m\ge0, v ∈ C loc 1 , α ( Ω ; R 2 ) , m ∈ C loc 0 , 1 ( Ω ) , m ≥ 0 , for some α ∈ ( 0 , 1 ] \alpha\in(0,1] α ∈ ( 0 , 1 ] .
We quantify collision risk through the instantaneous approach of nearby agents, rather than intersections of point trajectories. To isolate this effect from uniform drift, we measure the influx of agents into a small ball B r ( x ) B_r(x) B r ( x ) , $0<r<\operatorname{dist}(x,\partial\Omega), r e l a t i v e t o t h e a g e n t a t i t s c e n t e r , relative to the agent at its center , r e l a t i v e t o t h e a g e n t a t i t sce n t er x\in\Omega:
F r ( x ) : = ∫ ∂ B r ( x ) m ( y ) ( − ( v ( y ) − v ( x ) ) ⋅ ν ( y ) ) + d s . F_r(x):=\int_{\partial B_r(x)}m(y)\bigl(-(v(y)-v(x))\cdot\nu(y)\bigr)^{+}\,ds. F r ( x ) := ∫ ∂ B r ( x ) m ( y ) ( − ( v ( y ) − v ( x )) ⋅ ν ( y ) ) + d s .
The positive part retains only inward relative motion.
Lemma 3.1 (Small-radius asymptotics of the influx). Under Assumption 1, for every x ∈ Ω x\in\Omega x ∈ Ω , as r ↓ 0 r\downarrow0 r ↓ 0 ,
F r ( x ) = r 2 m ( x ) ∫ 0 2 π ( e ( θ ) T D v ( x ) e ( θ ) ) − d θ + O ( r 2 + α ) , (2) F_r(x)=r^{2}m(x)\int_{0}^{2\pi}\bigl(e(\theta)^{\mathsf{T}}Dv(x)e(\theta)\bigr)^{-}\,d\theta+O(r^{2+\alpha}), \tag*{(2)} F r ( x ) = r 2 m ( x ) ∫ 0 2 π ( e ( θ ) T D v ( x ) e ( θ ) ) − d θ + O ( r 2 + α ) , (2) where, for every compact K ⋐ Ω K\Subset\Omega K ⋐ Ω , the O ( r 2 + α ) O(r^{2+\alpha}) O ( r 2 + α ) remainder is uniform for x ∈ K x\in K x ∈ K and 0 < r < 1 2 dist ( K , ∂ Ω ) 0<r<\frac{1}{2}\operatorname{dist}(K,\partial\Omega) 0 < r < 2 1 dist ( K , ∂ Ω ) . Consequently,
lim r → 0 F r ( x ) π r 2 = m ( x ) π ∫ 0 2 π ( e ( θ ) T D v ( x ) e ( θ ) ) − d θ . \lim_{r\to0}\frac{F_r(x)}{\pi r^{2}}=\frac{m(x)}{\pi}\int_{0}^{2\pi}\bigl(e(\theta)^{\mathsf{T}}Dv(x)e(\theta)\bigr)^{-}\,d\theta. r → 0 lim π r 2 F r ( x ) = π m ( x ) ∫ 0 2 π ( e ( θ ) T D v ( x ) e ( θ ) ) − d θ .
Proof. Fix a compact set K ⋐ Ω K \Subset\Omega K ⋐ Ω and set ρ K : = 1 2 dist ( K , ∂ Ω ) \rho_K := \frac{1}{2}\operatorname{dist}(K,\partial\Omega) ρ K := 2 1 dist ( K , ∂ Ω ) , so that
K ρ K : = { z ∈ Ω : dist ( z , K ) ≤ ρ K } ⋐ Ω . K_{\rho_K} := \{z \in\Omega: \operatorname{dist}(z,K) \le\rho_K\} \Subset\Omega. K ρ K := { z ∈ Ω : dist ( z , K ) ≤ ρ K } ⋐ Ω.
For x ∈ K x \in K x ∈ K and 0 < r < ρ K 0 < r < \rho_K 0 < r < ρ K , B r ( x ) ⊂ K ρ K B_r(x) \subset K_{\rho_K} B r ( x ) ⊂ K ρ K , and with the parametrization of § 2 , ν ( y ) = e ( θ ) \nu(y) = e(\theta) ν ( y ) = e ( θ ) and d s = r d θ ds = r\,d\theta d s = r d θ . Since v ∈ C l o c 1 , α v \in C^{1,\alpha}_{\mathrm{loc}} v ∈ C loc 1 , α , there is a constant C K C_K C K , depending only on the C 1 , α C^{1,\alpha} C 1 , α -seminorm of v v v on K ρ K K_{\rho_K} K ρ K , with
∣ v ( y ) − v ( x ) − D v ( x ) ( y − x ) ∣ ≤ C K ∣ y − x ∣ 1 + α = C K r 1 + α . \lvert v(y)-v(x)-Dv(x)(y-x)\rvert\le C_K\lvert y-x\rvert^{1+\alpha}=C_Kr^{1+\alpha}. ∣ v ( y ) − v ( x ) − D v ( x ) ( y − x )∣ ≤ C K ∣ y − x ∣ 1 + α = C K r 1 + α .
Writing a ( θ ) : = − r e ( θ ) T D v ( x ) e ( θ ) a(\theta):=-r e(\theta)^{\mathsf T}Dv(x)e(\theta) a ( θ ) := − r e ( θ ) T D v ( x ) e ( θ ) and using ν ( y ) = e ( θ ) \nu(y)=e(\theta) ν ( y ) = e ( θ ) ,
− ( v ( y ) − v ( x ) ) ⋅ ν ( y ) = a ( θ ) + b ( θ ) , ∣ b ( θ ) ∣ ≤ C K r 1 + α . -(v(y)-v(x))\cdot\nu(y)=a(\theta)+b(\theta),\qquad\lvert b(\theta)\rvert\le C_Kr^{1+\alpha}. − ( v ( y ) − v ( x )) ⋅ ν ( y ) = a ( θ ) + b ( θ ) , ∣ b ( θ )∣ ≤ C K r 1 + α .
The map s ↦ s + s\mapsto s^+ s ↦ s + is 1-Lipschitz, and ( − r t ) + = r t − (-rt)^+=rt^- ( − r t ) + = r t − for r > 0 r>0 r > 0 ; hence
( − ( v ( y ) − v ( x ) ) ⋅ ν ( y ) ) + = r ( e ( θ ) T D v ( x ) e ( θ ) ) − + O ( r 1 + α ) , \left(-(v(y)-v(x))\cdot\nu(y)\right)^+=r\left(e(\theta)^{\mathsf T}Dv(x)e(\theta)\right)^-+O(r^{1+\alpha}), ( − ( v ( y ) − v ( x )) ⋅ ν ( y ) ) + = r ( e ( θ ) T D v ( x ) e ( θ ) ) − + O ( r 1 + α ) ,
uniformly in θ \theta θ and in x ∈ K x\in K x ∈ K . Since m ∈ C l o c 0 , 1 m\in C^{0,1}_{\mathrm{loc}} m ∈ C loc 0 , 1 , we have m ( y ) = m ( x ) + O ( r ) m(y)=m(x)+O(r) m ( y ) = m ( x ) + O ( r ) , with the remainder controlled by the Lipschitz seminorm of m m m on K ρ K K_{\rho_K} K ρ K . The quantity ( e ( θ ) T D v ( x ) e ( θ ) ) − \left(e(\theta)^{\mathsf T}Dv(x)e(\theta)\right)^- ( e ( θ ) T D v ( x ) e ( θ ) ) − is uniformly bounded for x ∈ K x\in K x ∈ K and θ ∈ [ 0 , 2 π ] \theta\in[0,2\pi] θ ∈ [ 0 , 2 π ] . Multiplying and using α ≤ 1 \alpha\le1 α ≤ 1 ,
m ( y ) ( − ( v ( y ) − v ( x ) ) ⋅ ν ( y ) ) + = r m ( x ) ( e ( θ ) T D v ( x ) e ( θ ) ) − + O ( r 1 + α ) . m(y)\left(-(v(y)-v(x))\cdot\nu(y)\right)^+=rm(x)\left(e(\theta)^{\mathsf T}Dv(x)e(\theta)\right)^-+O(r^{1+\alpha}). m ( y ) ( − ( v ( y ) − v ( x )) ⋅ ν ( y ) ) + = r m ( x ) ( e ( θ ) T D v ( x ) e ( θ ) ) − + O ( r 1 + α ) .
Integrating against d s = r d θ ds=r\,d\theta d s = r d θ , over θ ∈ [ 0 , 2 π ] \theta\in[0,2\pi] θ ∈ [ 0 , 2 π ] , yields (2) , with a remainder constant depending only on ρ K \rho_K ρ K , ∥ v ∥ C 1 , α ( K ρ K ) \lVert v\rVert_{C^{1,\alpha}(K_{\rho_K})} ∥ v ∥ C 1 , α ( K ρ K ) , and ∥ m ∥ C 0 , 1 ( K ρ K ) \lVert m\rVert_{C^{0,1}(K_{\rho_K})} ∥ m ∥ C 0 , 1 ( K ρ K ) , hence uniform on K K K . Dividing by π r 2 \pi r^2 π r 2 and letting r → 0 r\to0 r → 0 gives the stated limit. □
We model local collision risk by combining the inward relative flux per unit area, lim r → 0 F r ( x ) / ( π r 2 ) \lim_{r\to0}F_r(x)/(\pi r^2) lim r → 0 F r ( x ) / ( π r 2 ) , with the local density of agents, m ( x ) m(x) m ( x ) . Since
e ( θ ) T D v ( x ) e ( θ ) = e ( θ ) T A ( x ) e ( θ ) , e(\theta)^{\mathsf T}Dv(x)e(\theta)=e(\theta)^{\mathsf T}A(x)e(\theta), e ( θ ) T D v ( x ) e ( θ ) = e ( θ ) T A ( x ) e ( θ ) ,
where A ( x ) : = 1 2 ( D v ( x ) + D v ( x ) T ) A(x):=\frac{1}{2}(Dv(x)+Dv(x)^{\mathsf T}) A ( x ) := 2 1 ( D v ( x ) + D v ( x ) T ) is the symmetric velocity gradient, we define the local collision intensity by
C ( x ) : = m ( x ) ( lim r → 0 F r ( x ) π r 2 ) = m 2 ( x ) π ∫ 0 2 π ( e ( θ ) T A ( x ) e ( θ ) ) − d θ . C(x):=m(x)\left(\lim_{r\to0}\frac{F_r(x)}{\pi r^2}\right)=\frac{m^2(x)}{\pi}\int_0^{2\pi}\left(e(\theta)^{\mathsf T}A(x)e(\theta)\right)^-\,d\theta. C ( x ) := m ( x ) ( r → 0 lim π r 2 F r ( x ) ) = π m 2 ( x ) ∫ 0 2 π ( e ( θ ) T A ( x ) e ( θ ) ) − d θ .
The factor m ( x ) m(x) m ( x ) corresponds to the product approximation m ( x ) m ( y ) m(x)m(y) m ( x ) m ( y ) for the density of ordered agent pairs, neglecting pair correlations, and the division by π r 2 \pi r^2 π r 2 normalizes out the interaction area, so C ( x ) C(x) C ( x ) is a normalized encounter intensity. Counting unordered pairs would introduce a factor 1 / 2 1/2 1/2 , which does not affect the minimization. The local collision intensity is defined at every interior point without prescribing boundary values. We define the total collision functional, possibly taking the value + ∞ +\infty + ∞ , by
C Ω = ∫ Ω C ( x ) d x = 1 π ∫ Ω m 2 ( x ) ∫ 0 2 π ( e ( θ ) T A ( x ) e ( θ ) ) − d θ d x . (3) \mathcal{C}_{\Omega}=\int_{\Omega}C(x)\,dx=\frac{1}{\pi}\int_{\Omega}m^2(x)\int_0^{2\pi}\left(e(\theta)^{\mathsf T}A(x)e(\theta)\right)^-\,d\theta\,dx. \tag*{(3)} C Ω = ∫ Ω C ( x ) d x = π 1 ∫ Ω m 2 ( x ) ∫ 0 2 π ( e ( θ ) T A ( x ) e ( θ ) ) − d θ d x . (3)
We now compute the local collision intensity in terms of the eigenvalues of A A A . The small-radius inward-flux reduction to the symmetric velocity gradient follows the classical Saffman–Turner approximation [19 ] , and the resulting two-dimensional angular integral was evaluated in [10 Appendix A.1] ; we include the short proof for completeness.
Proposition 3.2 (Explicit form of the collision integral). Let A A A be a real symmetric 2 × 2 2 \times2 2 × 2 matrix and set
I ( A ) : = ∫ 0 2 π ( e ( θ ) T A e ( θ ) ) − d θ . I(A) := \int_{0}^{2\pi} \left(e(\theta)^{\mathsf T} A e(\theta)\right)^{-} \, d\theta. I ( A ) := ∫ 0 2 π ( e ( θ ) T A e ( θ ) ) − d θ . Let λ 1 , λ 2 \lambda_1,\lambda_2 λ 1 , λ 2 be the eigenvalues of A A A , labeled so that λ 1 ≥ λ 2 \lambda_1 \ge\lambda_2 λ 1 ≥ λ 2 . Then I ( A ) I(A) I ( A ) depends only on ( λ 1 , λ 2 ) (\lambda_1,\lambda_2) ( λ 1 , λ 2 ) and
I ( A ) = { 0 , λ 1 ≥ λ 2 ≥ 0 , − π ( λ 1 + λ 2 ) , 0 ≥ λ 1 ≥ λ 2 , 2 − λ 1 λ 2 − ( λ 1 + λ 2 ) arccos ( λ 1 + λ 2 λ 1 − λ 2 ) , λ 1 > 0 > λ 2 . I(A) =
\begin{cases}
0, & \lambda_1 \ge\lambda_2 \ge0,\\
-\pi(\lambda_1+\lambda_2), & 0 \ge\lambda_1 \ge\lambda_2,\\
2\sqrt{-\lambda_1\lambda_2}-(\lambda_1+\lambda_2)\arccos\left(\frac{\lambda_1+\lambda_2}{\lambda_1-\lambda_2}\right), & \lambda_1>0>\lambda_2.
\end{cases} I ( A ) = ⎩ ⎨ ⎧ 0 , − π ( λ 1 + λ 2 ) , 2 − λ 1 λ 2 − ( λ 1 + λ 2 ) arccos ( λ 1 − λ 2 λ 1 + λ 2 ) , λ 1 ≥ λ 2 ≥ 0 , 0 ≥ λ 1 ≥ λ 2 , λ 1 > 0 > λ 2 . Proof. By the spectral theorem, A = Q T Λ Q A=Q^{\mathsf T}\Lambda Q A = Q T Λ Q with Q ∈ O ( 2 ) Q \in O(2) Q ∈ O ( 2 ) and Λ = diag ( λ 1 , λ 2 ) \Lambda=\operatorname{diag}(\lambda_1,\lambda_2) Λ = diag ( λ 1 , λ 2 ) . Then
e ( θ ) T ( Q T Λ Q ) e ( θ ) = ( Q e ( θ ) ) T Λ ( Q e ( θ ) ) . e(\theta)^{\mathsf T}(Q^{\mathsf T}\Lambda Q)e(\theta)=(Qe(\theta))^{\mathsf T}\Lambda(Qe(\theta)). e ( θ ) T ( Q T Λ Q ) e ( θ ) = ( Q e ( θ ) ) T Λ ( Q e ( θ )) . In two dimensions, Q Q Q is either a rotation or a reflection, so there exists ϕ ∈ R \phi\in\mathbb{R} ϕ ∈ R such that Q e ( θ ) = e ( ϕ ± θ ) Qe(\theta)=e(\phi\pm\theta) Q e ( θ ) = e ( ϕ ± θ ) . A change of variable ϑ = ϕ ± θ \vartheta=\phi\pm\theta ϑ = ϕ ± θ and 2 π 2\pi 2 π -periodicity yield
I ( A ) = I ( Q T Λ Q ) = ∫ 0 2 π ( e ( ϕ ± θ ) T Λ e ( ϕ ± θ ) ) − d θ = ∫ 0 2 π ( e ( ϑ ) T Λ e ( ϑ ) ) − d ϑ = I ( Λ ) . I(A)=I(Q^{\mathsf T}\Lambda Q)=\int_{0}^{2\pi}\left(e(\phi\pm\theta)^{\mathsf T}\Lambda e(\phi\pm\theta)\right)^{-}\,d\theta=\int_{0}^{2\pi}\left(e(\vartheta)^{\mathsf T}\Lambda e(\vartheta)\right)^{-}\,d\vartheta=I(\Lambda). I ( A ) = I ( Q T Λ Q ) = ∫ 0 2 π ( e ( ϕ ± θ ) T Λ e ( ϕ ± θ ) ) − d θ = ∫ 0 2 π ( e ( ϑ ) T Λ e ( ϑ ) ) − d ϑ = I ( Λ ) . Thus, I ( A ) I(A) I ( A ) depends only on ( λ 1 , λ 2 ) (\lambda_1,\lambda_2) ( λ 1 , λ 2 ) . In particular,
I ( A ) = I ( Λ ) = ∫ 0 2 π ( λ 1 cos 2 θ + λ 2 sin 2 θ ) − d θ . I(A)=I(\Lambda)=\int_{0}^{2\pi}\left(\lambda_1\cos^2\theta+\lambda_2\sin^2\theta\right)^{-}\,d\theta. I ( A ) = I ( Λ ) = ∫ 0 2 π ( λ 1 cos 2 θ + λ 2 sin 2 θ ) − d θ . Next, set
f ( θ ) = λ 1 cos 2 θ + λ 2 sin 2 θ = λ 1 + λ 2 2 + λ 1 − λ 2 2 cos ( 2 θ ) = : c + d cos ( 2 θ ) . f(\theta)=\lambda_1\cos^2\theta+\lambda_2\sin^2\theta=\frac{\lambda_1+\lambda_2}{2}+\frac{\lambda_1-\lambda_2}{2}\cos(2\theta)=:c+d\cos(2\theta). f ( θ ) = λ 1 cos 2 θ + λ 2 sin 2 θ = 2 λ 1 + λ 2 + 2 λ 1 − λ 2 cos ( 2 θ ) =: c + d cos ( 2 θ ) . If λ 1 , λ 2 ≥ 0 \lambda_1,\lambda_2\ge0 λ 1 , λ 2 ≥ 0 , then f ( θ ) ≥ 0 f(\theta)\ge0 f ( θ ) ≥ 0 for all θ \theta θ , hence f ( θ ) − = 0 f(\theta)^{-}=0 f ( θ ) − = 0 and I ( Λ ) = 0 I(\Lambda)=0 I ( Λ ) = 0 .
If λ 1 , λ 2 ≤ 0 \lambda_1,\lambda_2\le0 λ 1 , λ 2 ≤ 0 , then f ( θ ) ≤ 0 f(\theta)\le0 f ( θ ) ≤ 0 for all θ \theta θ , so f ( θ ) − = − f ( θ ) f(\theta)^{-}=-f(\theta) f ( θ ) − = − f ( θ ) and
I ( Λ ) = − ∫ 0 2 π ( c + d cos ( 2 θ ) ) d θ = − 2 π c = − π ( λ 1 + λ 2 ) . I(\Lambda)=-\int_{0}^{2\pi}(c+d\cos(2\theta))\,d\theta=-2\pi c=-\pi(\lambda_1+\lambda_2). I ( Λ ) = − ∫ 0 2 π ( c + d cos ( 2 θ )) d θ = − 2 π c = − π ( λ 1 + λ 2 ) . If λ 1 > 0 > λ 2 \lambda_1>0>\lambda_2 λ 1 > 0 > λ 2 , then ∣ c ∣ < d |c|<d ∣ c ∣ < d (since d 2 − c 2 = − λ 1 λ 2 > 0 d^2-c^2=-\lambda_1\lambda_2>0 d 2 − c 2 = − λ 1 λ 2 > 0 ), so f f f changes sign. Using t = 2 θ t=2\theta t = 2 θ and the 2 π 2\pi 2 π -periodicity of cos t \cos t cos t ,
I ( Λ ) = ∫ 0 2 π ( c + d cos ( 2 θ ) ) − d θ = ∫ 0 2 π ( c + d cos t ) − d t . I(\Lambda)=\int_{0}^{2\pi}(c+d\cos(2\theta))^{-}\,d\theta=\int_{0}^{2\pi}(c+d\cos t)^{-}\,dt. I ( Λ ) = ∫ 0 2 π ( c + d cos ( 2 θ ) ) − d θ = ∫ 0 2 π ( c + d cos t ) − d t . Let θ 0 : = arccos ( − c / d ) ∈ ( 0 , π ) \theta_0:=\arccos(-c/d)\in(0,\pi) θ 0 := arccos ( − c / d ) ∈ ( 0 , π ) ; then { t ∈ [ 0 , 2 π ] : c + d cos t < 0 } = ( θ 0 , 2 π − θ 0 ) \{t\in[0,2\pi]:c+d\cos t<0\}=(\theta_0,2\pi-\theta_0) { t ∈ [ 0 , 2 π ] : c + d cos t < 0 } = ( θ 0 , 2 π − θ 0 ) . Hence
I ( Λ ) = ∫ θ 0 2 π − θ 0 − ( c + d cos t ) d t = − 2 c ( π − θ 0 ) + 2 d sin θ 0 . I(\Lambda)=\int_{\theta_0}^{2\pi-\theta_0}-(c+d\cos t)\,dt=-2c(\pi-\theta_0)+2d\sin\theta_0. I ( Λ ) = ∫ θ 0 2 π − θ 0 − ( c + d cos t ) d t = − 2 c ( π − θ 0 ) + 2 d sin θ 0 . Since sin θ 0 = 1 − ( c / d ) 2 = d 2 − c 2 / d \sin\theta_0=\sqrt{1-(c/d)^2}=\sqrt{d^2-c^2}/d sin θ 0 = 1 − ( c / d ) 2 = d 2 − c 2 / d and π − θ 0 = arccos ( c / d ) \pi-\theta_0=\arccos(c/d) π − θ 0 = arccos ( c / d ) , we obtain
I ( Λ ) = 2 d 2 − c 2 − 2 c arccos ( c d ) = 2 − λ 1 λ 2 − ( λ 1 + λ 2 ) arccos ( λ 1 + λ 2 λ 1 − λ 2 ) . I(\Lambda)=2\sqrt{d^2-c^2}-2c\arccos\left(\frac{c}{d}\right)=2\sqrt{-\lambda_1\lambda_2}-(\lambda_1+\lambda_2)\arccos\left(\frac{\lambda_1+\lambda_2}{\lambda_1-\lambda_2}\right). I ( Λ ) = 2 d 2 − c 2 − 2 c arccos ( d c ) = 2 − λ 1 λ 2 − ( λ 1 + λ 2 ) arccos ( λ 1 − λ 2 λ 1 + λ 2 ) .
Remark 3.3. In the traceless case tr A = 0 \operatorname{tr} A=0 tr A = 0 , the eigenvalues satisfy λ 1 = − λ 2 = − det A ≥ 0 \lambda_1=-\lambda_2=\sqrt{-\det A}\ge0 λ 1 = − λ 2 = − det A ≥ 0 , so the formula gives I ( A ) = 2 − det A = 2 λ 1 I(A)=2\sqrt{-\det A}=2\lambda_1 I ( A ) = 2 − det A = 2 λ 1 . This is precisely the situation encountered in the dense crowd model of § 4 .
Dense incompressible stream-function formulation
We now specialize the collision functional to the dense incompressible regime
m ≡ 1 , div v = 0 in Ω . m \equiv1,\qquad\operatorname{div} v = 0 \quad\text{in } \Omega. m ≡ 1 , div v = 0 in Ω.
Here m ≡ 1 m \equiv1 m ≡ 1 represents a fixed reference density in normalized units. The variational analysis below is carried out on bounded simply connected Lipschitz domains.
Dense regime and boundary fluxes
On a simply connected domain, every sufficiently smooth divergence-free velocity field admits a stream function, unique up to an additive constant, such that
v = D ⊥ ψ = ( ψ y , − ψ x ) , v = D^\perp\psi= (\psi_y,-\psi_x), v = D ⊥ ψ = ( ψ y , − ψ x ) ,
and v ⋅ ν = ∂ τ ψ v \cdot\nu= \partial_\tau\psi v ⋅ ν = ∂ τ ψ .
Assumption 2 (Lipschitz domain and boundary flux). The domain Ω ⊂ R 2 \Omega\subset\mathbb{R}^2 Ω ⊂ R 2 is bounded and simply connected, with Lipschitz boundary Γ = ∂ Ω \Gamma= \partial\Omega Γ = ∂ Ω . The prescribed signed boundary flux j ∈ L 1 ( Γ ) j \in L^1(\Gamma) j ∈ L 1 ( Γ ) satisfies the compatibility condition
∫ Γ j d s = 0. (4) \int_\Gamma j\,\mathrm{d}s = 0. \tag*{(4)} ∫ Γ j d s = 0. (4)
The boundary condition v ⋅ ν = j v \cdot\nu= j v ⋅ ν = j prescribes inflow on { j < 0 } \{j < 0\} { j < 0 } , outflow on { j > 0 } \{j > 0\} { j > 0 } , and zero normal flow on { j = 0 } \{j = 0\} { j = 0 } , up to sets of arclength measure zero.
We parametrize Γ \Gamma Γ by arclength in the orientation induced by τ \tau τ . If j ∈ L 1 ( Γ ) j \in L^1(\Gamma) j ∈ L 1 ( Γ ) satisfies the compatibility condition (4.1), then there is a boundary primitive j ˉ ∈ W 1 , 1 ( Γ ) \bar{j} \in W^{1,1}(\Gamma) j ˉ ∈ W 1 , 1 ( Γ ) , unique up to an additive constant, with ∂ τ j ˉ = j \partial_\tau\bar{j} = j ∂ τ j ˉ = j weakly on Γ \Gamma Γ . We fix the additive constant in j ˉ \bar{j} j ˉ once and for all. The boundary flux condition is then encoded by the Dirichlet condition in the trace sense:
ψ = j ˉ on Γ . (5) \psi= \bar{j} \quad\text{on } \Gamma. \tag*{(5)} ψ = j ˉ on Γ. (5)
For nonsmooth stream functions, (4.2) is taken as the meaning of the prescribed boundary flux.
The dense collision energy
Let ψ ∈ C 2 ( Ω ‾ ) \psi\in C^2(\overline{\Omega}) ψ ∈ C 2 ( Ω ) . Since
D ( D ⊥ ψ ) = ( ψ x y ψ y y − ψ x x − ψ x y ) , D(D^\perp\psi)=
\begin{pmatrix}
\psi_{xy} & \psi_{yy}\\
-\psi_{xx} & -\psi_{xy}
\end{pmatrix}, D ( D ⊥ ψ ) = ( ψ x y − ψ xx ψ y y − ψ x y ) ,
the symmetric velocity gradient associated with D ⊥ ψ D^\perp\psi D ⊥ ψ is
A [ ψ ] : = 1 2 ( D ( D ⊥ ψ ) + D ( D ⊥ ψ ) T ) = ( ψ x y 1 2 ( ψ y y − ψ x x ) 1 2 ( ψ y y − ψ x x ) − ψ x y ) . A[\psi] := \frac{1}{2}\left(D(D^\perp\psi)+D(D^\perp\psi)^\mathsf{T}\right)
=
\begin{pmatrix}
\psi_{xy} & \frac{1}{2}(\psi_{yy}-\psi_{xx})\\
\frac{1}{2}(\psi_{yy}-\psi_{xx}) & -\psi_{xy}
\end{pmatrix}. A [ ψ ] := 2 1 ( D ( D ⊥ ψ ) + D ( D ⊥ ψ ) T ) = ( ψ x y 2 1 ( ψ y y − ψ xx ) 2 1 ( ψ y y − ψ xx ) − ψ x y ) .
This matrix is traceless, with eigenvalues
λ ± = ± λ , λ : = ψ x y 2 + 1 4 ( ψ x x − ψ y y ) 2 . \lambda_\pm= \pm\lambda,\qquad\lambda:= \sqrt{\psi_{xy}^2+\frac{1}{4}(\psi_{xx}-\psi_{yy})^2}. λ ± = ± λ , λ := ψ x y 2 + 4 1 ( ψ xx − ψ y y ) 2 .
By Proposition 3.2 and Remark 3.3,
∫ 0 2 π ( e ( θ ) T A [ ψ ] e ( θ ) ) − d θ = 2 λ = 4 ψ x y 2 + ( ψ x x − ψ y y ) 2 . \int_0^{2\pi}\left(e(\theta)^\mathsf{T}A[\psi]e(\theta)\right)^-\,\mathrm{d}\theta
=2\lambda
=\sqrt{4\psi_{xy}^2+(\psi_{xx}-\psi_{yy})^2}. ∫ 0 2 π ( e ( θ ) T A [ ψ ] e ( θ ) ) − d θ = 2 λ = 4 ψ x y 2 + ( ψ xx − ψ y y ) 2 .
Thus, in the dense incompressible regime, the total collision functional (3.2) reduces to
J [ ψ ] : = 1 π ∫ Ω 4 ψ x y 2 + ( ψ x x − ψ y y ) 2 d x d y . (6) J[\psi] := \frac{1}{\pi}\int_\Omega\sqrt{4\psi_{xy}^2+(\psi_{xx}-\psi_{yy})^2}\,\mathrm{d}x\,\mathrm{d}y. \tag*{(6)} J [ ψ ] := π 1 ∫ Ω 4 ψ x y 2 + ( ψ xx − ψ y y ) 2 d x d y . (6)
We introduce the second-order operator
T D 2 ψ : = ( 2 ψ x y , ψ x x − ψ y y ) , (7) T D^2\psi:= (2\psi_{xy},\psi_{xx}-\psi_{yy}), \tag*{(7)} T D 2 ψ := ( 2 ψ x y , ψ xx − ψ y y ) , (7)
so that
J [ ψ ] = 1 π ∫ Ω ∣ T D 2 ψ ∣ d x d y , (8) J[\psi] = \frac{1}{\pi}\int_{\Omega} \lvert T D^2\psi\rvert\,dx\,dy, \tag*{(8)} J [ ψ ] = π 1 ∫ Ω ∣ T D 2 ψ ∣ d x d y , (8)
a convex, positively one-homogeneous functional of T D 2 ψ T D^2\psi T D 2 ψ . Equivalently,
J [ ψ ] = 2 π ∫ Ω ∣ E ( D ⊥ ψ ) ∣ d x d y . J[\psi] = \frac{\sqrt{2}}{\pi}\int_{\Omega} \lvert E(D^\perp\psi)\rvert\,dx\,dy. J [ ψ ] = π 2 ∫ Ω ∣ E ( D ⊥ ψ )∣ d x d y .
The integrand is affine along positive rays, so convexity alone does not give uniqueness; Proposition 6.3 gives an explicit example with multiple minimizers.
The functional J J J controls only T D 2 ψ T D^2\psi T D 2 ψ . In general, J J J does not control ∥ D 2 ψ ∥ L 1 ( Ω ) \lVert D^2\psi\rVert_{L^1(\Omega)} ∥ D 2 ψ ∥ L 1 ( Ω ) , even for ψ ∈ C ∞ ( Ω ) \psi\in C^\infty(\Omega) ψ ∈ C ∞ ( Ω ) ; see [18 ] . This distinction is essential for the compactness and relaxation theory developed below.
The admissible set
Assumption 3 (Finite-energy lifting). The boundary primitive j ˉ \bar{j} j ˉ associated with the data in Assumption 2 admits a lifting satisfying
ψ b ∈ W 1 , 1 ( Ω ) , T D 2 ψ b ∈ L 1 ( Ω ; R 2 ) , ψ b = j ˉ on Γ . (9) \psi_b \in W^{1,1}(\Omega), \qquad T D^2\psi_b \in L^1(\Omega;\mathbb{R}^2), \qquad\psi_b = \bar{j} \quad\text{on } \Gamma. \tag*{(9)} ψ b ∈ W 1 , 1 ( Ω ) , T D 2 ψ b ∈ L 1 ( Ω ; R 2 ) , ψ b = j ˉ on Γ. (9) Here T D 2 ψ b T D^2\psi_b T D 2 ψ b is understood in the sense of distributions and the boundary equality is understood in the trace sense. We fix one such lifting throughout the dense variational analysis.
For example, Assumption 3 holds whenever j ˉ \bar{j} j ˉ is the boundary restriction of a C 2 C^2 C 2 function defined on a neighborhood of Ω ‾ \overline{\Omega} Ω . The admissible set is
A : = { ψ ∈ W 1 , 1 ( Ω ) : T D 2 ψ ∈ L 1 ( Ω ; R 2 ) , ψ − ψ b ∈ W 0 1 , 1 ( Ω ) } . (10) \mathcal{A} := \left\{ \psi\in W^{1,1}(\Omega) : T D^2\psi\in L^1(\Omega;\mathbb{R}^2),\ \psi-\psi_b \in W^{1,1}_0(\Omega) \right\}. \tag*{(10)} A := { ψ ∈ W 1 , 1 ( Ω ) : T D 2 ψ ∈ L 1 ( Ω ; R 2 ) , ψ − ψ b ∈ W 0 1 , 1 ( Ω ) } . (10)
The class A \mathcal{A} A requires only the distributional combination T D 2 ψ T D^2\psi T D 2 ψ to be integrable; the relaxation will allow T D 2 ψ T D^2\psi T D 2 ψ to be a measure. The dense variational problem is
inf ψ ∈ A J [ ψ ] . (11) \inf_{\psi\in\mathcal{A}} J[\psi]. \tag*{(11)} ψ ∈ A inf J [ ψ ] . (11)
Since ψ b ∈ A \psi_b \in\mathcal{A} ψ b ∈ A , 0 ≤ inf ψ ∈ A J [ ψ ] ≤ J [ ψ b ] < + ∞ 0 \le\inf_{\psi\in\mathcal{A}} J[\psi] \le J[\psi_b] < +\infty 0 ≤ inf ψ ∈ A J [ ψ ] ≤ J [ ψ b ] < + ∞ .
Geometry of the zero-trace kernel
We characterize the zero-trace kernel of T D 2 T D^2 T D 2 . It is trivial on domains other than disks. On a disk, it is one-dimensional and corresponds to rigid rotation about the center.
Lemma 4.1 (Distributional kernel of T D 2 T D^2 T D 2 ). Let Ω \Omega Ω be connected. If u ∈ W 1 , 1 ( Ω ) u \in W^{1,1}(\Omega) u ∈ W 1 , 1 ( Ω ) satisfies T D 2 u = 0 T D^2u = 0 T D 2 u = 0 in distributions, then
u ( x , y ) = α 2 ( x 2 + y 2 ) + β x + γ y + δ (12) u(x,y) = \frac{\alpha}{2}(x^2+y^2)+\beta x+\gamma y+\delta \tag*{(12)} u ( x , y ) = 2 α ( x 2 + y 2 ) + β x + γ y + δ (12) almost everywhere in Ω \Omega Ω , for constants α , β , γ , δ \alpha,\beta,\gamma,\delta α , β , γ , δ .
Proof. All derivatives below are understood in the sense of distributions. Differentiating u x y = 0 u_{xy}=0 u x y = 0 gives
u x x y = u x y y = 0. u_{xxy}=u_{xyy}=0. u xx y = u x y y = 0.
Differentiating u x x − u y y = 0 u_{xx}-u_{yy}=0 u xx − u y y = 0 with respect to x x x and y y y then yields
u x x x = u x y y = 0 , u y y y = u x x y = 0. u_{xxx}=u_{xyy}=0,\qquad u_{yyy}=u_{xxy}=0. u xxx = u x y y = 0 , u y y y = u xx y = 0.
Thus both u x x u_{xx} u xx and u y y u_{yy} u y y have zero distributional gradient. Since Ω \Omega Ω is connected and u x x = u y y u_{xx}=u_{yy} u xx = u y y , there is a constant α ∈ R \alpha\in\mathbb{R} α ∈ R such that
u x x = u y y = α , u x y = 0. u_{xx}=u_{yy}=\alpha,\qquad u_{xy}=0. u xx = u y y = α , u x y = 0.
It follows that
D ( u x − α x ) = 0 , D ( u y − α y ) = 0. D(u_x-\alpha x)=0,\qquad D(u_y-\alpha y)=0. D ( u x − α x ) = 0 , D ( u y − α y ) = 0.
Connectedness therefore gives constants β , γ ∈ R \beta,\gamma\in\mathbb{R} β , γ ∈ R such that
u x = α x + β , u y = α y + γ . u_x=\alpha x+\beta,\qquad u_y=\alpha y+\gamma. u x = α x + β , u y = α y + γ .
Consequently,
D ( u − α 2 ( x 2 + y 2 ) − β x − γ y ) = 0. D\left(u-\frac{\alpha}{2}(x^2+y^2)-\beta x-\gamma y\right)=0. D ( u − 2 α ( x 2 + y 2 ) − β x − γ y ) = 0.
Applying the same constant-distribution property once more, we obtain a constant δ ∈ R \delta\in\mathbb{R} δ ∈ R such that
u ( x , y ) = α 2 ( x 2 + y 2 ) + β x + γ y + δ u(x,y)=\frac{\alpha}{2}(x^2+y^2)+\beta x+\gamma y+\delta u ( x , y ) = 2 α ( x 2 + y 2 ) + β x + γ y + δ
almost everywhere in Ω \Omega Ω . This proves (4.9). □
We define
K 0 : = { q ∈ W 0 1 , 1 ( Ω ) : T D 2 q = 0 } . (13) \mathcal{K}_0:=\{q\in W^{1,1}_0(\Omega):TD^2q=0\}. \tag*{(13)} K 0 := { q ∈ W 0 1 , 1 ( Ω ) : T D 2 q = 0 } . (13)
Lemma 4.1 shows that K 0 \mathcal{K}_0 K 0 is finite-dimensional and consists of polynomials of the form (4.9).
Proposition 4.2 (Geometry of the zero-trace kernel). Let Ω ⊂ R 2 \Omega\subset\mathbb{R}^2 Ω ⊂ R 2 be a bounded connected Lipschitz domain. Then K 0 = { 0 } \mathcal{K}_0=\{0\} K 0 = { 0 } unless Ω \Omega Ω is an open disk. If
Ω = { ( x , y ) ∈ R 2 : ( x − x 0 ) 2 + ( y − y 0 ) 2 < R 2 } , R > 0 , \Omega=\{(x,y)\in\mathbb{R}^2:(x-x_0)^2+(y-y_0)^2<R^2\},\qquad R>0, Ω = {( x , y ) ∈ R 2 : ( x − x 0 ) 2 + ( y − y 0 ) 2 < R 2 } , R > 0 , then
K 0 = span { q 0 } , q 0 ( x , y ) = ( x − x 0 ) 2 + ( y − y 0 ) 2 − R 2 2 . (14) \mathcal{K}_0=\operatorname{span}\{q_0\},\qquad q_0(x,y)=\frac{(x-x_0)^2+(y-y_0)^2-R^2}{2}. \tag*{(14)} K 0 = span { q 0 } , q 0 ( x , y ) = 2 ( x − x 0 ) 2 + ( y − y 0 ) 2 − R 2 . (14)
Proof. Suppose 0 ≠ q ∈ K 0 0\ne q\in\mathcal{K}_0 0 = q ∈ K 0 . By Lemma 4.1, q q q has the form (4.9) and, by continuity and the zero trace, vanishes everywhere on Γ \Gamma Γ . Its quadratic coefficient α \alpha α is nonzero, since an affine function with zero boundary values on a bounded domain must vanish identically.
Completing the square gives
q ( x , y ) = α 2 ( ( x − x 0 ) 2 + ( y − y 0 ) 2 + c ) q(x,y)=\frac{\alpha}{2}\left((x-x_0)^2+(y-y_0)^2+c\right) q ( x , y ) = 2 α ( ( x − x 0 ) 2 + ( y − y 0 ) 2 + c )
for some ( x 0 , y 0 ) ∈ R 2 (x_0,y_0)\in\mathbb{R}^2 ( x 0 , y 0 ) ∈ R 2 and c ∈ R c\in\mathbb{R} c ∈ R . Since q = 0 q=0 q = 0 on Γ \Gamma Γ and Γ \Gamma Γ contains more than one point, we must have c = − R 2 c=-R^2 c = − R 2 for some R > 0 R>0 R > 0 , and hence Γ ⊂ ∂ B R ( x 0 , y 0 ) \Gamma\subset\partial B_R(x_0,y_0) Γ ⊂ ∂ B R ( x 0 , y 0 ) .
Let B = B R ( x 0 , y 0 ) B=B_R(x_0,y_0) B = B R ( x 0 , y 0 ) , so that Γ = ∂ Ω ⊂ ∂ B \Gamma=\partial\Omega\subset\partial B Γ = ∂ Ω ⊂ ∂ B . For q 0 q_0 q 0 defined by (4.11), we have q = α q 0 q=\alpha q_0 q = α q 0 , so q 0 = 0 q_0=0 q 0 = 0 on ∂ Ω \partial\Omega ∂ Ω . Since Δ q 0 = 2 > 0 \Delta q_0=2>0 Δ q 0 = 2 > 0 , the strong maximum principle gives q 0 < 0 q_0<0 q 0 < 0 throughout Ω \Omega Ω , and hence Ω ⊂ B \Omega\subset B Ω ⊂ B .
To see the reverse inclusion, note that B ∩ ∂ Ω = ∅ B\cap\partial\Omega=\varnothing B ∩ ∂ Ω = ∅ because ∂ Ω ⊂ ∂ B \partial\Omega\subset\partial B ∂ Ω ⊂ ∂ B . If there existed a point z ∈ B ∖ Ω z\in B\setminus\Omega z ∈ B ∖ Ω , then for any fixed p ∈ Ω ⊂ B p\in\Omega\subset B p ∈ Ω ⊂ B , the line segment connecting p p p to z z z would lie entirely in the convex set B B B and would have to intersect ∂ Ω \partial\Omega ∂ Ω , contradicting B ∩ ∂ Ω = ∅ B\cap\partial\Omega=\varnothing B ∩ ∂ Ω = ∅ . Thus B ⊂ Ω B\subset\Omega B ⊂ Ω , and consequently Ω = B \Omega=B Ω = B .
Conversely, on this disk, the function q 0 q_0 q 0 in (4.11) belongs to K 0 \mathcal{K}_0 K 0 . For any q ∈ K 0 q\in\mathcal{K}_0 q ∈ K 0 , subtracting α q 0 \alpha q_0 α q 0 , where α \alpha α is its quadratic coefficient, leaves an affine function with zero boundary values, hence the zero function. Therefore K 0 = span { q 0 } \mathcal{K}_0=\operatorname{span}\{q_0\} K 0 = span { q 0 } . □
Normalization and coercivity
Coercivity holds after removing the component in the zero-trace kernel K 0 K_0 K 0 . The compactness argument uses the relation between T D 2 ψ TD^2\psi T D 2 ψ and the symmetric gradient of D ⊥ ψ D^\perp\psi D ⊥ ψ . For every ψ ∈ W 1 , 1 ( Ω ) \psi\in W^{1,1}(\Omega) ψ ∈ W 1 , 1 ( Ω ) with T D 2 ψ ∈ M ( Ω ; R 2 ) TD^2\psi\in\mathcal{M}(\Omega;\mathbb{R}^2) T D 2 ψ ∈ M ( Ω ; R 2 ) , a direct computation gives D ⊥ ψ ∈ B D ( Ω ) D^\perp\psi\in BD(\Omega) D ⊥ ψ ∈ B D ( Ω ) and
∥ D ⊥ ψ ∥ B D ( Ω ) = ∥ D ψ ∥ L 1 ( Ω ) + 1 2 ∣ T D 2 ψ ∣ ( Ω ) . (15) \lVert D^\perp\psi\rVert_{BD(\Omega)}=\lVert D\psi\rVert_{L^1(\Omega)}+\frac{1}{\sqrt{2}}\lvert TD^2\psi\rvert(\Omega). \tag*{(15)} ∥ D ⊥ ψ ∥ B D ( Ω ) = ∥ D ψ ∥ L 1 ( Ω ) + 2 1 ∣ T D 2 ψ ∣ ( Ω ) . (15)
We use the compact embedding B D ( Ω ) ⋐ L 1 ( Ω ; R 2 ) BD(\Omega)\Subset L^1(\Omega;\mathbb{R}^2) B D ( Ω ) ⋐ L 1 ( Ω ; R 2 ) on bounded Lipschitz domains; see [2 Theorem 2.10] .
Example 4.3 (Failure of coercivity on a disk). Let Ω \Omega Ω be a disk and let q 0 q_0 q 0 be given by (14) . For the fixed lifting ψ b \psi_b ψ b in Assumption 3, the sequence ψ n = ψ b + n q 0 \psi_n=\psi_b+nq_0 ψ n = ψ b + n q 0 belongs to A \mathcal{A} A and satisfies
J [ ψ n ] = J [ ψ b ] , ∥ ψ n ∥ W 1 , 1 ( Ω ) ⟶ + ∞ . J[\psi_n]=J[\psi_b],\qquad\lVert\psi_n\rVert_{W^{1,1}(\Omega)}\longrightarrow+\infty. J [ ψ n ] = J [ ψ b ] , ∥ ψ n ∥ W 1 , 1 ( Ω ) ⟶ + ∞. Thus the collision energy does not control the W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) norm, even when the boundary trace is fixed.
Proposition 4.4 (Coercivity). Let Ω ⊂ R 2 \Omega\subset\mathbb{R}^2 Ω ⊂ R 2 be a bounded connected Lipschitz domain. Define
P 0 u : = { ∫ Ω u q 0 d x d y ∫ Ω q 0 2 d x d y q 0 , if Ω is a disk , 0 , otherwise , (16) P_{0}u:=\begin{cases}\frac{\int_\Omega uq_{0}\,dx\,dy}{\int_\Omega q_{0}^{2}\,dx\,dy}q_{0}, & \text{if } \Omega\text{ is a disk},\\0, & \text{otherwise},\end{cases} \tag*{(16)} P 0 u := { ∫ Ω q 0 2 d x d y ∫ Ω u q 0 d x d y q 0 , 0 , if Ω is a disk , otherwise , (16) where q 0 q_0 q 0 is given by (14) . Then P 0 : W 1 , 1 ( Ω ) → K 0 P_0:W^{1,1}(\Omega)\to K_0 P 0 : W 1 , 1 ( Ω ) → K 0 is a continuous linear projection, and there is a constant C 0 > 0 C_0>0 C 0 > 0 , depending only on Ω \Omega Ω , such that
∥ u − P 0 u ∥ W 1 , 1 ( Ω ) ≤ C 0 ∣ T D 2 u ∣ ( Ω ) (17) \lVert u-P_0u\rVert_{W^{1,1}(\Omega)}\le C_0\lvert TD^2u\rvert(\Omega) \tag*{(17)} ∥ u − P 0 u ∥ W 1 , 1 ( Ω ) ≤ C 0 ∣ T D 2 u ∣ ( Ω ) (17) for every u ∈ W 0 1 , 1 ( Ω ) u\in W^{1,1}_0(\Omega) u ∈ W 0 1 , 1 ( Ω ) with T D 2 u ∈ M ( Ω ; R 2 ) TD^2u\in\mathcal{M}(\Omega;\mathbb{R}^2) T D 2 u ∈ M ( Ω ; R 2 ) .
If ψ b ∈ W 1 , 1 ( Ω ) \psi_b\in W^{1,1}(\Omega) ψ b ∈ W 1 , 1 ( Ω ) with T D 2 ψ b ∈ L 1 ( Ω ; R 2 ) TD^2\psi_b\in L^1(\Omega;\mathbb{R}^2) T D 2 ψ b ∈ L 1 ( Ω ; R 2 ) is fixed, then there is a constant C > 0 C>0 C > 0 , depending only on Ω \Omega Ω and ψ b \psi_b ψ b , such that
∥ ψ − P 0 ψ ∥ W 1 , 1 ( Ω ) ≤ C ( 1 + ∣ T D 2 ψ ∣ ( Ω ) ) (18) \lVert\psi-P_0\psi\rVert_{W^{1,1}(\Omega)}\le C\bigl(1+\lvert TD^2\psi\rvert(\Omega)\bigr) \tag*{(18)} ∥ ψ − P 0 ψ ∥ W 1 , 1 ( Ω ) ≤ C ( 1 + ∣ T D 2 ψ ∣ ( Ω ) ) (18) for every ψ ∈ W 1 , 1 ( Ω ) \psi\in W^{1,1}(\Omega) ψ ∈ W 1 , 1 ( Ω ) satisfying T D 2 ψ ∈ M ( Ω ; R 2 ) TD^2\psi\in\mathcal{M}(\Omega;\mathbb{R}^2) T D 2 ψ ∈ M ( Ω ; R 2 ) and ψ − ψ b ∈ W 0 1 , 1 ( Ω ) \psi-\psi_b\in W^{1,1}_0(\Omega) ψ − ψ b ∈ W 0 1 , 1 ( Ω ) .
Proof. By Proposition 4.2, P 0 P_0 P 0 is a projection onto K 0 K_0 K 0 . Its continuity is immediate when P 0 = 0 P_0=0 P 0 = 0 , and follows from boundedness of q 0 q_0 q 0 in the disk case.
Suppose (4.14) fails. Subtracting the projection and normalizing gives a sequence z n ∈ W 0 1 , 1 ( Ω ) z_n\in W^{1,1}_0(\Omega) z n ∈ W 0 1 , 1 ( Ω ) such that
∥ z n ∥ W 1 , 1 ( Ω ) = 1 , P 0 z n = 0 , ∣ T D 2 z n ∣ ( Ω ) ⟶ 0. \lVert z_n\rVert_{W^{1,1}(\Omega)}=1,\qquad P_0z_n=0,\qquad\lvert TD^2z_n\rvert(\Omega)\longrightarrow0. ∥ z n ∥ W 1 , 1 ( Ω ) = 1 , P 0 z n = 0 , ∣ T D 2 z n ∣ ( Ω ) ⟶ 0. By (4.12), the fields D ⊥ z n D^\perp z_n D ⊥ z n are bounded in B D ( Ω ) BD(\Omega) B D ( Ω ) . The compact embedding therefore gives a subsequence for which D z n Dz_n D z n converges strongly in L 1 L^1 L 1 . Poincaré’s inequality then gives strong convergence in W 0 1 , 1 ( Ω ) W^{1,1}_0(\Omega) W 0 1 , 1 ( Ω ) to some z ∈ W 0 1 , 1 ( Ω ) z\in W^{1,1}_0(\Omega) z ∈ W 0 1 , 1 ( Ω ) . Distributional convergence and continuity of P 0 P_0 P 0 yield
∥ z ∥ W 1 , 1 ( Ω ) = 1 , T D 2 z = 0 , P 0 z = 0. \lVert z\rVert_{W^{1,1}(\Omega)}=1,\qquad TD^2z=0,\qquad P_0z=0. ∥ z ∥ W 1 , 1 ( Ω ) = 1 , T D 2 z = 0 , P 0 z = 0. Thus z ∈ K 0 z\in K_0 z ∈ K 0 , so z = P 0 z = 0 z=P_0z=0 z = P 0 z = 0 , a contradiction.
For the prescribed-trace estimate, apply (4.14) to u = ψ − ψ b u=\psi-\psi_b u = ψ − ψ b . By linearity of P 0 P_0 P 0 and the triangle inequality,
∥ ψ − P 0 ψ ∥ W 1 , 1 ( Ω ) ≤ ∥ ψ b − P 0 ψ b ∥ W 1 , 1 ( Ω ) + C 0 ( ∣ T D 2 ψ ∣ ( Ω ) + ∥ T D 2 ψ b ∥ L 1 ( Ω ) ) . \lVert\psi-P_0\psi\rVert_{W^{1,1}(\Omega)}
\le\lVert\psi_b-P_0\psi_b\rVert_{W^{1,1}(\Omega)}
+C_0\bigl(\lvert TD^2\psi\rvert(\Omega)+\lVert TD^2\psi_b\rVert_{L^1(\Omega)}\bigr). ∥ ψ − P 0 ψ ∥ W 1 , 1 ( Ω ) ≤ ∥ ψ b − P 0 ψ b ∥ W 1 , 1 ( Ω ) + C 0 ( ∣ T D 2 ψ ∣ ( Ω ) + ∥ T D 2 ψ b ∥ L 1 ( Ω ) ) . Absorbing the fixed lifting terms into C C C proves (4.15). □
For ψ \psi ψ satisfying the prescribed-trace hypotheses of Proposition 4.4, define
ψ ^ : = ψ − P 0 ψ . \widehat{\psi}:=\psi-P_0\psi. ψ := ψ − P 0 ψ .
Since P 0 ψ ∈ K 0 P_0\psi\in K_0 P 0 ψ ∈ K 0 ,
ψ ^ − ψ b ∈ W 0 1 , 1 ( Ω ) , T D 2 ψ ^ = T D 2 ψ , P 0 ψ ^ = 0. (19) \widehat{\psi}-\psi_b\in W^{1,1}_0(\Omega),\qquad TD^2\widehat{\psi}=TD^2\psi,\qquad P_0\widehat{\psi}=0. \tag*{(19)} ψ − ψ b ∈ W 0 1 , 1 ( Ω ) , T D 2 ψ = T D 2 ψ , P 0 ψ = 0. (19)
In particular, if ψ ∈ A \psi\in\mathcal{A} ψ ∈ A , then ψ ^ ∈ A \widehat{\psi}\in\mathcal{A} ψ ∈ A and J [ ψ ^ ] = J [ ψ ] J[\widehat{\psi}]=J[\psi] J [ ψ ] = J [ ψ ] . When Ω \Omega Ω is not a disk, the normalization is the identity.
Remark 4.5 (Interpretation of the disk normalization). On a disk, the normalized stream function satisfies
∫ Ω ψ ^ q 0 d x d y = 0. \int_{\Omega}\widehat{\psi}q_0\,dx\,dy=0. ∫ Ω ψ q 0 d x d y = 0.
Among the functions in ψ + K 0 \psi+K_0 ψ + K 0 , ψ ^ \widehat{\psi} ψ is the unique one satisfying this condition. Since
D ⊥ q 0 = ( y − y 0 , − ( x − x 0 ) ) , D^\perp q_0=(y-y_0,-(x-x_0)), D ⊥ q 0 = ( y − y 0 , − ( x − x 0 )) ,
the normalization subtracts a rigid rotation tangent to the circular boundary. It may change the velocity, but preserves the prescribed trace and the entire measure T D 2 ψ TD^2\psi T D 2 ψ . It is used to obtain compactness and does not select a preferred physical flow. Only zero-trace kernel elements are subtracted; a general element of ker T D 2 \ker TD^2 ker T D 2 need not preserve the prescribed trace or boundary flux.
The relaxed dense variational problem
Proposition 4.4 bounds every minimizing sequence for J J J in W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) after the normalization ψ ↦ ψ − P 0 ψ \psi\mapsto\psi-P_0\psi ψ ↦ ψ − P 0 ψ . This normalization preserves the prescribed trace and energy and is the identity when Ω \Omega Ω is not a disk.
The energy bounds the L 1 L^1 L 1 norm of T D 2 ψ TD^2\psi T D 2 ψ , but these derivatives may concentrate along sequences of bounded energy. We therefore retain the prescribed trace and allow T D 2 ψ TD^2\psi T D 2 ψ to be a finite vector-valued Radon measure. Compactness and existence follow from the normalized coercivity estimate. The recovery construction and the identification of the relaxation require no normalization.
The relaxed admissible class and energy
5.1. 5.1. 5.1. The relaxed admissible class and energy. We define
X T ( Ω ) : = { ψ ∈ W 1 , 1 ( Ω ) : T D 2 ψ ∈ M ( Ω ; R 2 ) } . (20) \mathcal{X}_T(\Omega):=\left\{\psi\in W^{1,1}(\Omega):TD^2\psi\in\mathcal{M}(\Omega;\mathbb{R}^2)\right\}. \tag*{(20)} X T ( Ω ) := { ψ ∈ W 1 , 1 ( Ω ) : T D 2 ψ ∈ M ( Ω ; R 2 ) } . (20)
The operator T D 2 TD^2 T D 2 is understood in the sense of distributions. Thus, for Φ = ( Φ 1 , Φ 2 ) ∈ C c ∞ ( Ω ; R 2 ) \Phi=(\Phi_1,\Phi_2)\in C^\infty_c(\Omega;\mathbb{R}^2) Φ = ( Φ 1 , Φ 2 ) ∈ C c ∞ ( Ω ; R 2 ) ,
⟨ T D 2 ψ , Φ ⟩ = ∫ Ω ψ ( 2 ∂ x y Φ 1 + ( ∂ x x − ∂ y y ) Φ 2 ) d x d y . \langle TD^2\psi,\Phi\rangle=\int_{\Omega}\psi\left(2\partial_{xy}\Phi_1+(\partial_{xx}-\partial_{yy})\Phi_2\right)\,dx\,dy. ⟨ T D 2 ψ , Φ ⟩ = ∫ Ω ψ ( 2 ∂ x y Φ 1 + ( ∂ xx − ∂ y y ) Φ 2 ) d x d y .
The boundary flux is imposed through the same trace condition as in the unrelaxed problem. We define
B : = { ψ ∈ X T ( Ω ) : ψ − ψ b ∈ W 0 1 , 1 ( Ω ) } . (21) \mathcal{B}:=\left\{\psi\in\mathcal{X}_T(\Omega):\psi-\psi_b\in W^{1,1}_0(\Omega)\right\}. \tag*{(21)} B := { ψ ∈ X T ( Ω ) : ψ − ψ b ∈ W 0 1 , 1 ( Ω ) } . (21)
The boundary condition is equivalent to ψ = j ˉ \psi=\bar{j} ψ = j ˉ almost everywhere on Γ \Gamma Γ in the trace sense. By Assumption 3, ψ b ∈ A ⊂ B \psi_b\in\mathcal{A}\subset\mathcal{B} ψ b ∈ A ⊂ B .
We extend the functional in (4.5) to X T ( Ω ) \mathcal{X}_T(\Omega) X T ( Ω ) by setting
J [ ψ ] : = 1 π ∣ T D 2 ψ ∣ ( Ω ) . J[\psi]:=\frac{1}{\pi}\lvert TD^2\psi\rvert(\Omega). J [ ψ ] := π 1 ∣ T D 2 ψ ∣ ( Ω ) .
If
T D 2 ψ = g ψ L 2 + μ ψ s TD^2\psi=g_\psi\mathcal{L}^2+\mu_\psi^s T D 2 ψ = g ψ L 2 + μ ψ s
is the Lebesgue decomposition of T D 2 ψ T D^2\psi T D 2 ψ , then
J [ ψ ] = 1 π ∫ Ω ∣ g ψ ∣ d x d y + 1 π ∣ μ ψ s ∣ ( Ω ) . \mathcal{J}[\psi] = \frac{1}{\pi}\int_{\Omega} |g_\psi|\,dx\,dy + \frac{1}{\pi}|\mu_\psi^s|(\Omega). J [ ψ ] = π 1 ∫ Ω ∣ g ψ ∣ d x d y + π 1 ∣ μ ψ s ∣ ( Ω ) .
For a relaxed stream function, the natural collision measure is π − 1 ∣ T D 2 ψ ∣ \pi^{-1}|T D^2\psi| π − 1 ∣ T D 2 ψ ∣ . Its absolutely continuous density is π − 1 ∣ g ψ ∣ \pi^{-1}|g_\psi| π − 1 ∣ g ψ ∣ , while its singular part records concentration in the relaxed limit. Thus J \mathcal{J} J extends the original dense energy: if ψ ∈ A \psi\in\mathcal{A} ψ ∈ A , then T D 2 ψ T D^2\psi T D 2 ψ is absolutely continuous and
J [ ψ ] = 1 π ∫ Ω ∣ T D 2 ψ ∣ d x d y = J [ ψ ] . \mathcal{J}[\psi] = \frac{1}{\pi}\int_{\Omega}|T D^2\psi|\,dx\,dy = J[\psi]. J [ ψ ] = π 1 ∫ Ω ∣ T D 2 ψ ∣ d x d y = J [ ψ ] .
The following lemma records the behavior of the energy along a sequence converging to a prescribed limit. It applies on every domain satisfying the standing assumptions and does not use normalization.
Lemma 5.1 (Limits of sequences with bounded energy). Under Assumptions 2 and 3, let { ψ n } ⊂ B \{\psi_n\}\subset\mathcal{B} { ψ n } ⊂ B satisfy
ψ n → ψ in W 1 , 1 ( Ω ) , sup n J [ ψ n ] < + ∞ . \psi_n \to\psi\quad\text{in } W^{1,1}(\Omega), \qquad\sup_n \mathcal{J}[\psi_n] < +\infty. ψ n → ψ in W 1 , 1 ( Ω ) , n sup J [ ψ n ] < + ∞. Then ψ ∈ B \psi\in\mathcal{B} ψ ∈ B ,
T D 2 ψ n ⇀ ∗ T D 2 ψ in M ( Ω ; R 2 ) , J [ ψ ] ≤ lim inf n → ∞ J [ ψ n ] . T D^2\psi_n \stackrel{*}{\rightharpoonup} T D^2\psi\quad\text{in } \mathcal{M}(\Omega;\mathbb{R}^2), \qquad\mathcal{J}[\psi]\leq\liminf_{n\to\infty}\mathcal{J}[\psi_n]. T D 2 ψ n ⇀ ∗ T D 2 ψ in M ( Ω ; R 2 ) , J [ ψ ] ≤ n → ∞ lim inf J [ ψ n ] . Moreover, the weak-* convergence holds for the entire sequence.
Proof. The energy bound gives a weak-* convergent subsequence of T D 2 ψ n T D^2\psi_n T D 2 ψ n . Strong W 1 , 1 W^{1,1} W 1 , 1 convergence implies distributional convergence, so the limiting measure is T D 2 ψ T D^2\psi T D 2 ψ . Thus ψ ∈ X T ( Ω ) \psi\in X_T(\Omega) ψ ∈ X T ( Ω ) . Since W 0 1 , 1 ( Ω ) W_0^{1,1}(\Omega) W 0 1 , 1 ( Ω ) is closed and ψ n − ψ b ∈ W 0 1 , 1 ( Ω ) \psi_n-\psi_b\in W_0^{1,1}(\Omega) ψ n − ψ b ∈ W 0 1 , 1 ( Ω ) , we also have ψ − ψ b ∈ W 0 1 , 1 ( Ω ) \psi-\psi_b\in W_0^{1,1}(\Omega) ψ − ψ b ∈ W 0 1 , 1 ( Ω ) . Hence ψ ∈ B \psi\in\mathcal{B} ψ ∈ B .
Every subsequence of T D 2 ψ n T D^2\psi_n T D 2 ψ n has a further weak-* convergent subsequence by the uniform energy bound. Distributional convergence identifies every such limit with T D 2 ψ T D^2\psi T D 2 ψ . Hence the entire sequence converges weak-* to T D 2 ψ T D^2\psi T D 2 ψ . For every Φ ∈ C c ( Ω ; R 2 ) \Phi\in C_c(\Omega;\mathbb{R}^2) Φ ∈ C c ( Ω ; R 2 ) with ∥ Φ ∥ ∞ ≤ 1 \|\Phi\|_\infty\leq1 ∥Φ ∥ ∞ ≤ 1 , we have
∫ Ω Φ ⋅ d ( T D 2 ψ ) = lim n → ∞ ∫ Ω Φ ⋅ d ( T D 2 ψ n ) ≤ lim inf n → ∞ ∣ T D 2 ψ n ∣ ( Ω ) . \int_{\Omega}\Phi\cdot d(T D^2\psi)=\lim_{n\to\infty}\int_{\Omega}\Phi\cdot d(T D^2\psi_n)\leq\liminf_{n\to\infty}|T D^2\psi_n|(\Omega). ∫ Ω Φ ⋅ d ( T D 2 ψ ) = n → ∞ lim ∫ Ω Φ ⋅ d ( T D 2 ψ n ) ≤ n → ∞ lim inf ∣ T D 2 ψ n ∣ ( Ω ) .
Taking the supremum over such Φ \Phi Φ , by the dual characterization of the total variation of a vector-valued Radon measure, gives
∣ T D 2 ψ ∣ ( Ω ) ≤ lim inf n → ∞ ∣ T D 2 ψ n ∣ ( Ω ) . |T D^2\psi|(\Omega)\leq\liminf_{n\to\infty}|T D^2\psi_n|(\Omega). ∣ T D 2 ψ ∣ ( Ω ) ≤ n → ∞ lim inf ∣ T D 2 ψ n ∣ ( Ω ) .
Dividing by π \pi π proves the energy inequality. □
Theorem 5.6 will identify J \mathcal{J} J on B \mathcal{B} B with the relaxation of J J J on A \mathcal{A} A for strong W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) convergence.
Compactness and existence
We prove compactness after normalization using Proposition 4.4 and the compact embedding of B D ( Ω ) BD(\Omega) B D ( Ω ) into L 1 ( Ω ; R 2 ) L^1(\Omega;\mathbb{R}^2) L 1 ( Ω ; R 2 ) . The direct method then gives a minimizer of J \mathcal{J} J over the full class B \mathcal{B} B .
Proposition 5.2 (Compactness after normalization). Under Assumptions 2 and 3, let { ψ n } ⊂ B \{\psi_n\}\subset\mathcal{B} { ψ n } ⊂ B satisfy
sup n ∣ T D 2 ψ n ∣ ( Ω ) < + ∞ . \sup_n |T D^2\psi_n|(\Omega)<+\infty. n sup ∣ T D 2 ψ n ∣ ( Ω ) < + ∞. Set
ψ ^ n : = ψ n − P 0 ψ n , \widehat{\psi}_n:=\psi_n-P_0\psi_n, ψ n := ψ n − P 0 ψ n ,
where P 0 P_0 P 0 is given by (16) . Then, for every n n n ,
ψ ^ n ∈ B , T D 2 ψ ^ n = T D 2 ψ n , J [ ψ ^ n ] = J [ ψ n ] . (22) \widehat{\psi}_n \in\mathcal{B}, \qquad T D^2\widehat{\psi}_n = T D^2\psi_n, \qquad\mathcal{J}[\widehat{\psi}_n] = \mathcal{J}[\psi_n]. \tag*{(22)} ψ n ∈ B , T D 2 ψ n = T D 2 ψ n , J [ ψ n ] = J [ ψ n ] . (22)
There are ψ ∈ B \psi\in\mathcal{B} ψ ∈ B , with P 0 ψ = 0 P_0\psi= 0 P 0 ψ = 0 , and a subsequence, not relabeled, such that
ψ ^ n → ψ in W 1 , 1 ( Ω ) , T D 2 ψ n ⇀ ∗ T D 2 ψ in M ( Ω ; R 2 ) . \widehat{\psi}_n \to\psi\quad\text{in } W^{1,1}(\Omega), \qquad T D^2\psi_n \mathrel{\overset{*}{\rightharpoonup}} T D^2\psi\quad\text{in } \mathcal{M}(\Omega;\mathbb{R}^2). ψ n → ψ in W 1 , 1 ( Ω ) , T D 2 ψ n ⇀ ∗ T D 2 ψ in M ( Ω ; R 2 ) .
If Ω \Omega Ω is not a disk, then ψ ^ n = ψ n \widehat{\psi}_n = \psi_n ψ n = ψ n .
Proof. Equation (19) gives (22) and P 0 ψ ^ n = 0 P_0\widehat{\psi}_n = 0 P 0 ψ n = 0 . Proposition 4.4 bounds ψ ^ n \widehat{\psi}_n ψ n in W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) . Together with (15) and the energy bound, this shows that D ⊥ ψ ^ n D^\perp\widehat{\psi}_n D ⊥ ψ n is bounded in B D ( Ω ) BD(\Omega) B D ( Ω ) . The compact embedding gives a subsequence for which D ψ ^ n D\widehat{\psi}_n D ψ n converges strongly in L 1 L^1 L 1 .
Since ψ ^ n − ψ ^ m ∈ W 0 1 , 1 ( Ω ) \widehat{\psi}_n-\widehat{\psi}_m \in W^{1,1}_0(\Omega) ψ n − ψ m ∈ W 0 1 , 1 ( Ω ) , Poincaré’s inequality gives
∥ ψ ^ n − ψ ^ m ∥ L 1 ( Ω ) ≤ C ∥ D ψ ^ n − D ψ ^ m ∥ L 1 ( Ω ) . \|\widehat{\psi}_n-\widehat{\psi}_m\|_{L^1(\Omega)} \leq C\|D\widehat{\psi}_n-D\widehat{\psi}_m\|_{L^1(\Omega)}. ∥ ψ n − ψ m ∥ L 1 ( Ω ) ≤ C ∥ D ψ n − D ψ m ∥ L 1 ( Ω ) .
Thus this subsequence converges strongly in W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) to some ψ \psi ψ . Lemma 5.1 gives ψ ∈ B \psi\in\mathcal{B} ψ ∈ B and the asserted weak-* convergence, since T D 2 ψ ^ n = T D 2 ψ n T D^2\widehat{\psi}_n = T D^2\psi_n T D 2 ψ n = T D 2 ψ n . Finally, continuity of P 0 P_0 P 0 gives P 0 ψ = 0 P_0\psi= 0 P 0 ψ = 0 .
Corollary 5.3 (Compactness of relaxed energy sublevels). Under Assumptions 2 and 3, for M ≥ 0 M \geq0 M ≥ 0 , define
B M : = { ψ ∈ B : J [ ψ ] ≤ M } , B ^ M : = { ψ ∈ B M : P 0 ψ = 0 } . \mathcal{B}_M := \{\psi\in\mathcal{B} : \mathcal{J}[\psi] \leq M\}, \qquad\widehat{\mathcal{B}}_M := \{\psi\in\mathcal{B}_M : P_0\psi= 0\}. B M := { ψ ∈ B : J [ ψ ] ≤ M } , B M := { ψ ∈ B M : P 0 ψ = 0 } . Then B ^ M \widehat{\mathcal{B}}_M B M is compact in the strong W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) topology. If Ω \Omega Ω is not a disk, then B ^ M = B M \widehat{\mathcal{B}}_M = \mathcal{B}_M B M = B M . If Ω \Omega Ω is a disk, every nonempty B M \mathcal{B}_M B M is unbounded in W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) .
Proof. Let { ψ n } ⊂ B ^ M \{\psi_n\} \subset\widehat{\mathcal{B}}_M { ψ n } ⊂ B M . Since P 0 ψ n = 0 P_0\psi_n = 0 P 0 ψ n = 0 for every n n n , the normalized sequence satisfies ψ ^ n = ψ n \widehat{\psi}_n = \psi_n ψ n = ψ n . Proposition 5.2 gives a strongly convergent subsequence whose limit belongs to B \mathcal{B} B and satisfies P 0 ψ = 0 P_0\psi= 0 P 0 ψ = 0 . Lemma 5.1 preserves the energy bound, proving compactness of B ^ M \widehat{\mathcal{B}}_M B M . If Ω \Omega Ω is not a disk, then P 0 = 0 P_0 = 0 P 0 = 0 , so B ^ M = B M \widehat{\mathcal{B}}_M = \mathcal{B}_M B M = B M . If Ω \Omega Ω is a disk and ψ ∈ B M \psi\in\mathcal{B}_M ψ ∈ B M , then ψ + t q 0 ∈ B M \psi+ t q_0 \in\mathcal{B}_M ψ + t q 0 ∈ B M for every t ∈ R t \in\mathbb{R} t ∈ R , because q 0 q_0 q 0 has zero trace and T D 2 q 0 = 0 T D^2q_0 = 0 T D 2 q 0 = 0 . However,
∥ ψ + t q 0 ∥ W 1 , 1 ( Ω ) ≥ ∣ t ∣ ∥ q 0 ∥ W 1 , 1 ( Ω ) − ∥ ψ ∥ W 1 , 1 ( Ω ) ⟶ + ∞ as ∣ t ∣ → ∞ . \|\psi+tq_0\|_{W^{1,1}(\Omega)} \geq|t|\|q_0\|_{W^{1,1}(\Omega)}-\|\psi\|_{W^{1,1}(\Omega)} \longrightarrow+\infty\quad\text{as } |t| \to\infty. ∥ ψ + t q 0 ∥ W 1 , 1 ( Ω ) ≥ ∣ t ∣∥ q 0 ∥ W 1 , 1 ( Ω ) − ∥ ψ ∥ W 1 , 1 ( Ω ) ⟶ + ∞ as ∣ t ∣ → ∞. Thus every nonempty unnormalized sublevel on a disk is unbounded.
Theorem 5.4 (Existence of relaxed minimizers). Under Assumptions 2 and 3, J \mathcal{J} J attains its minimum on B \mathcal{B} B .
Proof. Since ψ b ∈ B \psi_b \in\mathcal{B} ψ b ∈ B and J [ ψ b ] = J [ ψ ^ b ] < + ∞ \mathcal{J}[\psi_b] = \mathcal{J}[\widehat{\psi}_b] < +\infty J [ ψ b ] = J [ ψ b ] < + ∞ , the infimum is finite and nonnegative. Let { ψ n } ⊂ B \{\psi_n\} \subset\mathcal{B} { ψ n } ⊂ B be a minimizing sequence. Then ∣ T D 2 ψ n ∣ ( Ω ) = π J [ ψ n ] |T D^2\psi_n|(\Omega) = \pi\mathcal{J}[\psi_n] ∣ T D 2 ψ n ∣ ( Ω ) = π J [ ψ n ] is uniformly bounded.
By Proposition 5.2, the normalized sequence ψ ^ n = ψ n − P 0 ψ n \widehat{\psi}_n = \psi_n-P_0\psi_n ψ n = ψ n − P 0 ψ n has a subsequence converging strongly in W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) to some ψ ∈ B \psi\in\mathcal{B} ψ ∈ B . Since normalization preserves the energy, Lemma 5.1 gives
J [ ψ ] ≤ lim inf n → ∞ J [ ψ ^ n ] = lim inf n → ∞ J [ ψ n ] = inf φ ∈ B J [ φ ] . \mathcal{J}[\psi] \leq\liminf_{n\to\infty}\mathcal{J}[\widehat{\psi}_n] = \liminf_{n\to\infty}\mathcal{J}[\psi_n] = \inf_{\varphi\in\mathcal{B}}\mathcal{J}[\varphi]. J [ ψ ] ≤ n → ∞ lim inf J [ ψ n ] = n → ∞ lim inf J [ ψ n ] = φ ∈ B inf J [ φ ] . Since ψ ∈ B \psi\in\mathcal{B} ψ ∈ B , it is a minimizer.
Strict recovery with the prescribed trace
We approximate every ψ ∈ B \psi\in\mathcal{B} ψ ∈ B by functions in A \mathcal{A} A with the same prescribed trace and convergent collision energy. The construction applies interior mollification directly to ψ \psi ψ and combines the local approximations using a partition of unity. No normalization is needed, whether or not Ω \Omega Ω is a disk.
Theorem 5.5 (Boundary-preserving strict T D 2 TD^2 T D 2 -recovery). Under Assumptions 2 and 3, for every ψ ∈ B \psi\in\mathcal{B} ψ ∈ B , there exists a sequence { ψ k } k ∈ N ⊂ A ∩ C ∞ ( Ω ) \{\psi_k\}_{k\in\mathbb{N}} \subset\mathcal{A} \cap C^\infty(\Omega) { ψ k } k ∈ N ⊂ A ∩ C ∞ ( Ω ) such that
ψ k − ψ ∈ W 0 1 , 1 ( Ω ) , ψ k → ψ in W 1 , 1 ( Ω ) , \psi_k-\psi\in W^{1,1}_0(\Omega), \qquad\psi_k \to\psi\quad\text{in } W^{1,1}(\Omega), ψ k − ψ ∈ W 0 1 , 1 ( Ω ) , ψ k → ψ in W 1 , 1 ( Ω ) , T D 2 ψ k ⇀ ∗ T D 2 ψ in M ( Ω ; R 2 ) , TD^2\psi_k \mathrel{\overset{*}{\rightharpoonup}} TD^2\psi\quad\text{in } \mathcal{M}(\Omega;\mathbb{R}^2), T D 2 ψ k ⇀ ∗ T D 2 ψ in M ( Ω ; R 2 ) , and
∣ T D 2 ψ k ∣ ( Ω ) → ∣ T D 2 ψ ∣ ( Ω ) . (23) |TD^2\psi_k|(\Omega) \to|TD^2\psi|(\Omega). \tag*{(23)} ∣ T D 2 ψ k ∣ ( Ω ) → ∣ T D 2 ψ ∣ ( Ω ) . (23) Consequently, J [ ψ k ] → J [ ψ ] J[\psi_k] \to J[\psi] J [ ψ k ] → J [ ψ ] .
Proof. Local smoothing and the product rule. Set μ : = T D 2 ψ \mu:= TD^2\psi μ := T D 2 ψ . Choose a nonnegative, locally finite partition of unity { φ i } i ≥ 1 ⊂ C c ∞ ( Ω ) \{\varphi_i\}_{i\geq1} \subset C^\infty_c(\Omega) { φ i } i ≥ 1 ⊂ C c ∞ ( Ω ) , with ∑ i φ i = 1 \sum_i \varphi_i = 1 ∑ i φ i = 1 on Ω \Omega Ω . Extend each φ i \varphi_i φ i smoothly by zero to R 2 \mathbb{R}^2 R 2 . Denote by ψ ~ \widetilde{\psi} ψ and μ ~ \widetilde{\mu} μ the zero extensions of ψ \psi ψ as an L 1 L^1 L 1 function and of μ \mu μ as a finite measure. The extension of ψ \psi ψ is used only for interior convolution; no derivative identity across Γ \Gamma Γ is required.
Let ρ \rho ρ be a nonnegative even standard mollifier supported in B 1 ( 0 ) B_1(0) B 1 ( 0 ) , and set
ψ δ : = ρ δ ∗ ψ ~ . \psi_\delta:= \rho_\delta* \widetilde{\psi}. ψ δ := ρ δ ∗ ψ .
As δ ↓ 0 \delta\downarrow0 δ ↓ 0 , we have ψ δ → ψ \psi_\delta\to\psi ψ δ → ψ in W l o c 1 , 1 ( Ω ) W^{1,1}_{\mathrm{loc}}(\Omega) W loc 1 , 1 ( Ω ) . For
0 < δ < 1 2 dist ( supp φ i , R 2 ∖ Ω ) , 0 < \delta< \frac{1}{2}\operatorname{dist}(\operatorname{supp}\varphi_i,\mathbb{R}^2 \setminus\Omega), 0 < δ < 2 1 dist ( supp φ i , R 2 ∖ Ω ) ,
the identity
T D 2 ψ δ = ρ δ ∗ μ ~ (24) TD^2\psi_\delta= \rho_\delta* \widetilde{\mu} \tag*{(24)} T D 2 ψ δ = ρ δ ∗ μ (24)
holds on a neighborhood of supp φ i \operatorname{supp}\varphi_i supp φ i .
For φ ∈ C c ∞ ( Ω ) \varphi\in C^\infty_c(\Omega) φ ∈ C c ∞ ( Ω ) and q ∈ W 1 , 1 ( Ω ) q \in W^{1,1}(\Omega) q ∈ W 1 , 1 ( Ω ) , define
R φ [ q ] : = ( 2 φ x q y + 2 φ y q x + 2 φ x y q , 2 φ x q x − 2 φ y q y + ( φ x x − φ y y ) q ) . R_\varphi[q] := \left(2\varphi_xq_y+2\varphi_yq_x+2\varphi_{xy}q,\
2\varphi_xq_x-2\varphi_yq_y+(\varphi_{xx}-\varphi_{yy})q\right). R φ [ q ] := ( 2 φ x q y + 2 φ y q x + 2 φ x y q , 2 φ x q x − 2 φ y q y + ( φ xx − φ y y ) q ) .
The distributional product rule gives
T D 2 ( φ q ) = φ T D 2 q + R φ [ q ] . (25) TD^2(\varphi q)=\varphi TD^2q+R_\varphi[q]. \tag*{(25)} T D 2 ( φq ) = φT D 2 q + R φ [ q ] . (25)
The remainder involves only q q q and its first derivatives.
For each k , i ≥ 1 k,i\geq1 k , i ≥ 1 , choose δ k , i > 0 \delta_{k,i}>0 δ k , i > 0 small enough for (5.5) to hold near supp φ i \operatorname{supp}\varphi_i supp φ i and for all three inequalities
∥ φ i ( ψ δ k , i − ψ ) ∥ W 1 , 1 ( Ω ) ≤ 2 − i k , (26) \left\|\varphi_i(\psi_{\delta_{k,i}}-\psi)\right\|_{W^{1,1}(\Omega)}\leq\frac{2^{-i}}{k}, \tag*{(26)} φ i ( ψ δ k , i − ψ ) W 1 , 1 ( Ω ) ≤ k 2 − i , (26)
∥ R φ i [ ψ δ k , i − ψ ] ∥ L 1 ( Ω ) ≤ 2 − i k , (27) \left\|R_{\varphi_i}[\psi_{\delta_{k,i}}-\psi]\right\|_{L^1(\Omega)}\leq\frac{2^{-i}}{k}, \tag*{(27)} R φ i [ ψ δ k , i − ψ ] L 1 ( Ω ) ≤ k 2 − i , (27)
∥ ρ δ k , i ∗ φ i − φ i ∥ L ∞ ( R 2 ) ≤ 2 − i k . (28) \left\|\rho_{\delta_{k,i}} * \varphi_i-\varphi_i\right\|_{L^\infty(\mathbb{R}^2)}\leq\frac{2^{-i}}{k}. \tag*{(28)} ρ δ k , i ∗ φ i − φ i L ∞ ( R 2 ) ≤ k 2 − i . (28)
The first two follow from local W 1 , 1 W^{1,1} W 1 , 1 convergence. The third follows from uniform convergence of mollifications of the fixed smooth compactly supported function φ i \varphi_i φ i .
Preservation of the scalar trace. Define
ψ k : = ∑ i = 1 ∞ φ i ψ δ k , i . (29) \psi_k := \sum_{i=1}^{\infty} \varphi_i \psi_{\delta_{k,i}} . \tag*{(29)} ψ k := i = 1 ∑ ∞ φ i ψ δ k , i . (29)
The sum is locally finite, so ψ k ∈ C ∞ ( Ω ) \psi_k \in C^\infty(\Omega) ψ k ∈ C ∞ ( Ω ) . Moreover,
ψ k − ψ = ∑ i φ i ( ψ δ k , i − ψ ) . \psi_k-\psi=\sum_i\varphi_i(\psi_{\delta_{k,i}}-\psi). ψ k − ψ = i ∑ φ i ( ψ δ k , i − ψ ) .
By (26) , this series converges absolutely in W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) , with
∥ ψ k − ψ ∥ W 1 , 1 ( Ω ) ≤ 1 k . \lVert\psi_k-\psi\rVert_{W^{1,1}(\Omega)}\leq\frac{1}{k}. ∥ ψ k − ψ ∥ W 1 , 1 ( Ω ) ≤ k 1 .
Every summand belongs to W 0 1 , 1 ( Ω ) W^{1,1}_0(\Omega) W 0 1 , 1 ( Ω ) , since it belongs to W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) and has compact support in Ω \Omega Ω . Closedness of W 0 1 , 1 ( Ω ) W^{1,1}_0(\Omega) W 0 1 , 1 ( Ω ) therefore gives
ψ k − ψ ∈ W 0 1 , 1 ( Ω ) , ψ k → ψ in W 1 , 1 ( Ω ) . \psi_k-\psi\in W^{1,1}_0(\Omega),\qquad\psi_k\to\psi\quad\text{in }W^{1,1}(\Omega). ψ k − ψ ∈ W 0 1 , 1 ( Ω ) , ψ k → ψ in W 1 , 1 ( Ω ) .
Since ψ ∈ B \psi\in B ψ ∈ B , we also have ψ k − ψ b ∈ W 0 1 , 1 ( Ω ) \psi_k-\psi_b\in W^{1,1}_0(\Omega) ψ k − ψ b ∈ W 0 1 , 1 ( Ω ) . Thus the prescribed scalar trace is preserved exactly.
Convergence of the energy. Local finiteness and ∑ i φ i = 1 \sum_i\varphi_i=1 ∑ i φ i = 1 give
∑ i D φ i = 0 , ∑ i D 2 φ i = 0. \sum_iD\varphi_i=0,\qquad\sum_iD^2\varphi_i=0. i ∑ D φ i = 0 , i ∑ D 2 φ i = 0.
Hence ∑ i R φ i [ ψ ] = 0 \sum_i\mathcal{R}_{\varphi_i}[\psi]=0 ∑ i R φ i [ ψ ] = 0 locally. By local finiteness, we may differentiate the sum defining ψ k \psi_k ψ k term by term. The product rule (25) and the local convolution identity (24) give
T D 2 ψ k = ∑ i T D 2 ( φ i ψ δ k , i ) = ∑ i φ i ( ρ δ k , i ∗ μ ~ ) + ∑ i R φ i [ ψ δ k , i ] . \begin{aligned}
TD^2\psi_k=\sum_iTD^2(\varphi_i\psi_{\delta_{k,i}}) \\
&=\sum_i\varphi_i(\rho_{\delta_{k,i}}*\widetilde{\mu})+\sum_i\mathcal{R}_{\varphi_i}[\psi_{\delta_{k,i}}].
\end{aligned} T D 2 ψ k = i ∑ T D 2 ( φ i ψ δ k , i ) = i ∑ φ i ( ρ δ k , i ∗ μ ) + i ∑ R φ i [ ψ δ k , i ] .
Since R φ i \mathcal{R}_{\varphi_i} R φ i is linear in its argument and ∑ i R φ i [ ψ ] = 0 \sum_i\mathcal{R}_{\varphi_i}[\psi]=0 ∑ i R φ i [ ψ ] = 0 , we have
∑ i R φ i [ ψ δ k , i ] = ∑ i R φ i [ ψ δ k , i − ψ ] . \sum_i\mathcal{R}_{\varphi_i}[\psi_{\delta_{k,i}}]
=\sum_i\mathcal{R}_{\varphi_i}[\psi_{\delta_{k,i}}-\psi]. i ∑ R φ i [ ψ δ k , i ] = i ∑ R φ i [ ψ δ k , i − ψ ] .
Thus, in distributions,
T D 2 ψ k = M k + r k , TD^2\psi_k=M_k+r_k, T D 2 ψ k = M k + r k ,
where
M k : = ∑ i φ i ( ρ δ k , i ∗ μ ~ ) , r k : = ∑ i R φ i [ ψ δ k , i − ψ ] . M_k:=\sum_i\varphi_i(\rho_{\delta_{k,i}}*\widetilde{\mu}),\qquad
r_k:=\sum_i\mathcal{R}_{\varphi_i}[\psi_{\delta_{k,i}}-\psi]. M k := i ∑ φ i ( ρ δ k , i ∗ μ ) , r k := i ∑ R φ i [ ψ δ k , i − ψ ] .
By (27) , the series defining r k r_k r k converges absolutely in L 1 ( Ω ; R 2 ) L^1(\Omega;\mathbb{R}^2) L 1 ( Ω ; R 2 ) , and
∥ r k ∥ L 1 ( Ω ) ≤ 1 k . \lVert r_k\rVert_{L^1(\Omega)}\leq\frac{1}{k}. ∥ r k ∥ L 1 ( Ω ) ≤ k 1 .
Since ρ \rho ρ is even,
∫ R 2 φ i ( w ) ρ δ k , i ( w − z ) d w = ( ρ δ k , i ∗ φ i ) ( z ) . \int_{\mathbb{R}^2}\varphi_i(w)\rho_{\delta_{k,i}}(w-z)\,dw
=(\rho_{\delta_{k,i}}*\varphi_i)(z). ∫ R 2 φ i ( w ) ρ δ k , i ( w − z ) d w = ( ρ δ k , i ∗ φ i ) ( z ) .
Note also that (28) gives, for every z ∈ Ω z\in\Omega z ∈ Ω ,
( ρ δ k , i ∗ φ i ) ( z ) ≤ φ i ( z ) + 2 − i k . (\rho_{\delta_{k,i}}*\varphi_i)(z)\leq\varphi_i(z)+\frac{2^{-i}}{k}. ( ρ δ k , i ∗ φ i ) ( z ) ≤ φ i ( z ) + k 2 − i .
Summing over i i i and using ∑ i φ i ( z ) = 1 \sum_i \varphi_i(z)=1 ∑ i φ i ( z ) = 1 and ∑ i = 1 ∞ 2 − i = 1 \sum_{i=1}^{\infty}2^{-i}=1 ∑ i = 1 ∞ 2 − i = 1 , we obtain
∑ i ( ρ δ k , i ∗ φ i ) ( z ) ≤ 1 + 1 k . \sum_i(\rho_{\delta_{k,i}}*\varphi_i)(z)\le1+\frac{1}{k}. i ∑ ( ρ δ k , i ∗ φ i ) ( z ) ≤ 1 + k 1 .
Using the convolution identity, the preceding bound, nonnegativity, and Tonelli’s theorem, we obtain
∫ Ω ∣ M k ∣ d x d y ≤ ∑ i ∫ R 2 φ i ( ρ δ k , i ∗ ∣ μ ~ ∣ ) d x d y = ∫ Ω ∑ i ( ρ δ k , i ∗ φ i ) ( z ) d ∣ μ ∣ ( z ) ≤ ( 1 + 1 k ) ∣ μ ∣ ( Ω ) . \begin{aligned}
\int_{\Omega}|M_k|\,dx\,dy
&\le\sum_i\int_{\mathbb{R}^2}\varphi_i(\rho_{\delta_{k,i}}*|\widetilde{\mu}|)\,dx\,dy\\
&=\int_{\Omega}\sum_i(\rho_{\delta_{k,i}}*\varphi_i)(z)\,d|\mu|(z)\\
&\le\left(1+\frac{1}{k}\right)|\mu|(\Omega).
\end{aligned} ∫ Ω ∣ M k ∣ d x d y ≤ i ∑ ∫ R 2 φ i ( ρ δ k , i ∗ ∣ μ ∣ ) d x d y = ∫ Ω i ∑ ( ρ δ k , i ∗ φ i ) ( z ) d ∣ μ ∣ ( z ) ≤ ( 1 + k 1 ) ∣ μ ∣ ( Ω ) .
Consequently,
∫ Ω ∣ T D 2 ψ k ∣ d x d y ≤ ( 1 + 1 k ) ∣ μ ∣ ( Ω ) + 1 k . (30) \int_{\Omega}|TD^2\psi_k|\,dx\,dy\le\left(1+\frac{1}{k}\right)|\mu|(\Omega)+\frac{1}{k}. \tag*{(30)} ∫ Ω ∣ T D 2 ψ k ∣ d x d y ≤ ( 1 + k 1 ) ∣ μ ∣ ( Ω ) + k 1 . (30)
In particular, T D 2 ψ k ∈ L 1 ( Ω ; R 2 ) TD^2\psi_k\in L^1(\Omega;\mathbb{R}^2) T D 2 ψ k ∈ L 1 ( Ω ; R 2 ) . Together with the trace condition, this gives ψ k ∈ A \psi_k\in\mathcal{A} ψ k ∈ A .
The bound (30) and Lemma 5.1 give
T D 2 ψ k ⇀ ∗ T D 2 ψ in M ( Ω ; R 2 ) , TD^2\psi_k\mathrel{\mathop{\stackrel{*}{\rightharpoonup}}}TD^2\psi\quad\text{in }\mathcal{M}(\Omega;\mathbb{R}^2), T D 2 ψ k ⇀ ∗ T D 2 ψ in M ( Ω ; R 2 ) ,
and lower semicontinuity yields
∣ μ ∣ ( Ω ) ≤ lim inf k → ∞ ∫ Ω ∣ T D 2 ψ k ∣ d x d y . |\mu|(\Omega)\le\liminf_{k\to\infty}\int_{\Omega}|TD^2\psi_k|\,dx\,dy. ∣ μ ∣ ( Ω ) ≤ k → ∞ lim inf ∫ Ω ∣ T D 2 ψ k ∣ d x d y .
Together with (30) , this proves (23) . Finally,
J [ ψ k ] = 1 π ∣ T D 2 ψ k ∣ ( Ω ) ⟶ 1 π ∣ T D 2 ψ ∣ ( Ω ) = J [ ψ ] . J[\psi_k]=\frac{1}{\pi}|TD^2\psi_k|(\Omega)\longrightarrow\frac{1}{\pi}|TD^2\psi|(\Omega)=J[\psi]. J [ ψ k ] = π 1 ∣ T D 2 ψ k ∣ ( Ω ) ⟶ π 1 ∣ T D 2 ψ ∣ ( Ω ) = J [ ψ ] .
The approximation argument applies to every ψ ∈ X T ( Ω ) \psi\in\mathcal{X}_T(\Omega) ψ ∈ X T ( Ω ) on an arbitrary bounded open set Ω \Omega Ω and gives ψ k − ψ ∈ W 0 1 , 1 ( Ω ) \psi_k-\psi\in W^{1,1}_0(\Omega) ψ k − ψ ∈ W 0 1 , 1 ( Ω ) . On a Lipschitz domain, this means that the scalar traces are equal.
The approximants are smooth in Ω \Omega Ω and have integrable T D 2 ψ k TD^2\psi_k T D 2 ψ k . Neither integrability of the full Hessian nor smoothness up to ∂ Ω \partial\Omega ∂ Ω is asserted. This is precisely the regularity needed for A \mathcal{A} A .
Identification of the relaxation
For ψ ∈ W 1 , 1 ( Ω ) \psi\in W^{1,1}(\Omega) ψ ∈ W 1 , 1 ( Ω ) , define
J r e l [ ψ ] : = inf { lim inf k → ∞ J [ ψ k ] : ψ k ∈ A , ψ k → ψ in W 1 , 1 ( Ω ) } . (31) J_{\mathrm{rel}}[\psi]:=\inf\left\{\liminf_{k\to\infty}J[\psi_k]:\ \psi_k\in\mathcal{A},\ \psi_k\to\psi\text{ in }W^{1,1}(\Omega)\right\}. \tag*{(31)} J rel [ ψ ] := inf { k → ∞ lim inf J [ ψ k ] : ψ k ∈ A , ψ k → ψ in W 1 , 1 ( Ω ) } . (31)
We use the convention inf ∅ = + ∞ \inf\varnothing=+\infty inf ∅ = + ∞ . We also set
J ‾ [ ψ ] : = { J [ ψ ] , ψ ∈ B , + ∞ , ψ ∉ B . \overline{J}[\psi]:=
\begin{cases}
J[\psi], & \psi\in\mathcal{B},\\
+\infty, & \psi\notin\mathcal{B}.
\end{cases} J [ ψ ] := { J [ ψ ] , + ∞ , ψ ∈ B , ψ ∈ / B .
The following theorem identifies J ‾ \overline{J} J with the relaxation of J J J on A \mathcal{A} A in the strong W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) topology.
Theorem 5.6 (Relaxation identity). _Under Assumptions 2 and 3_, for every ψ ∈ W 1 , 1 ( Ω ) \psi\in W^{1,1}(\Omega) ψ ∈ W 1 , 1 ( Ω ) ,
J r e l [ ψ ] = J ‾ [ ψ ] . (32) J_{\mathrm{rel}}[\psi]=\overline{J}[\psi]. \tag*{(32)} J rel [ ψ ] = J [ ψ ] . (32)
Consequently,
min ψ ∈ B J [ ψ ] = inf ψ ∈ A J [ ψ ] . (33) \min_{\psi\in B} \mathcal{J}[\psi] = \inf_{\psi\in A} J[\psi]. \tag*{(33)} ψ ∈ B min J [ ψ ] = ψ ∈ A inf J [ ψ ] . (33)
Proof. We first prove the lower bound. Let { ψ k } ⊂ A \{\psi_k\} \subset A { ψ k } ⊂ A converge to ψ \psi ψ strongly in W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) , and set
L : = lim inf k → ∞ J [ ψ k ] . L := \liminf_{k\to\infty} J[\psi_k]. L := k → ∞ lim inf J [ ψ k ] .
If L = + ∞ L=+\infty L = + ∞ , then J ‾ [ ψ ] ≤ L \overline{\mathcal{J}}[\psi]\leq L J [ ψ ] ≤ L is immediate. Otherwise, pass to a subsequence, not relabeled, such that J [ ψ k ] → L J[\psi_k]\to L J [ ψ k ] → L . The total variations ∣ T D 2 ψ k ∣ ( Ω ) = π J [ ψ k ] |TD^2\psi_k|(\Omega)=\pi J[\psi_k] ∣ T D 2 ψ k ∣ ( Ω ) = π J [ ψ k ] are then uniformly bounded.
Lemma 5.1 gives ψ ∈ B \psi\in B ψ ∈ B and
T D 2 ψ k ⇀ ∗ T D 2 ψ in M ( Ω ; R 2 ) , TD^2\psi_k \mathrel{\stackrel{*}{\rightharpoonup}} TD^2\psi\quad\text{in } \mathcal{M}(\Omega;\mathbb{R}^2), T D 2 ψ k ⇀ ∗ T D 2 ψ in M ( Ω ; R 2 ) ,
without normalizing the sequence. Lower semicontinuity therefore yields
J ‾ [ ψ ] = J [ ψ ] ≤ lim inf k → ∞ J [ ψ k ] = L . \overline{\mathcal{J}}[\psi]=\mathcal{J}[\psi]\leq\liminf_{k\to\infty}J[\psi_k]=L. J [ ψ ] = J [ ψ ] ≤ k → ∞ lim inf J [ ψ k ] = L .
Taking the infimum over all approximating sequences gives
J ‾ [ ψ ] ≤ J r e l [ ψ ] . \overline{\mathcal{J}}[\psi]\leq J_{\mathrm{rel}}[\psi]. J [ ψ ] ≤ J rel [ ψ ] .
If ψ ∉ B \psi\notin B ψ ∈ / B , this inequality already gives J r e l [ ψ ] = + ∞ J_{\mathrm{rel}}[\psi]=+\infty J rel [ ψ ] = + ∞ . If ψ ∈ B \psi\in B ψ ∈ B , Theorem 5.5 provides a sequence { ψ k } ⊂ A \{\psi_k\}\subset A { ψ k } ⊂ A such that
ψ k → ψ in W 1 , 1 ( Ω ) , J [ ψ k ] → J [ ψ ] . \psi_k\to\psi\quad\text{in } W^{1,1}(\Omega), \qquad J[\psi_k]\to\mathcal{J}[\psi]. ψ k → ψ in W 1 , 1 ( Ω ) , J [ ψ k ] → J [ ψ ] .
Hence
J r e l [ ψ ] ≤ J [ ψ ] = J ‾ [ ψ ] . J_{\mathrm{rel}}[\psi]\leq\mathcal{J}[\psi]=\overline{\mathcal{J}}[\psi]. J rel [ ψ ] ≤ J [ ψ ] = J [ ψ ] .
This proves (5.13).
Since A ⊂ B A\subset B A ⊂ B and J = J J=\mathcal{J} J = J on A A A ,
min ψ ∈ B J [ ψ ] ≤ inf ψ ∈ A J [ ψ ] . \min_{\psi\in B}\mathcal{J}[\psi]\leq\inf_{\psi\in A}J[\psi]. ψ ∈ B min J [ ψ ] ≤ ψ ∈ A inf J [ ψ ] .
Let ψ ∗ ∈ B \psi_*\in B ψ ∗ ∈ B be a minimizer provided by Theorem 5.4. Applying Theorem 5.5 to ψ ∗ \psi_* ψ ∗ gives a sequence { ψ k } ⊂ A \{\psi_k\}\subset A { ψ k } ⊂ A with J [ ψ k ] → J [ ψ ∗ ] J[\psi_k]\to\mathcal{J}[\psi_*] J [ ψ k ] → J [ ψ ∗ ] . Therefore
inf ψ ∈ A J [ ψ ] ≤ J [ ψ ∗ ] = min ψ ∈ B J [ ψ ] . \inf_{\psi\in A}J[\psi]\leq\mathcal{J}[\psi_*]=\min_{\psi\in B}\mathcal{J}[\psi]. ψ ∈ A inf J [ ψ ] ≤ J [ ψ ∗ ] = ψ ∈ B min J [ ψ ] .
Combining the two inequalities proves (5.14). □
Duality and optimality conditions
We conclude the variational analysis with optimality conditions that will be used in Section 6 to prove the optimality of explicit flows. The Euclidean norm in the energy is convex but not differentiable at zero, so we formulate these conditions through convex duality. We first derive a dual representation of the minimum and characterize minimizers in A A A . We then obtain a sufficient condition for minimization in the relaxed class B B B when the dual field is smooth.
Since the boundary trace is prescribed, the admissible variations form the space
V 0 : = { h ∈ W 0 1 , 1 ( Ω ) : T D 2 h ∈ L 1 ( Ω ; R 2 ) } , V_0:=\{h\in W^{1,1}_0(\Omega):TD^2h\in L^1(\Omega;\mathbb{R}^2)\}, V 0 := { h ∈ W 0 1 , 1 ( Ω ) : T D 2 h ∈ L 1 ( Ω ; R 2 )} ,
so that A = ψ b + V 0 A=\psi_b+V_0 A = ψ b + V 0 . We define Z \mathcal{Z} Z as the set of fields Z ∈ L ∞ ( Ω ; R 2 ) Z\in L^\infty(\Omega;\mathbb{R}^2) Z ∈ L ∞ ( Ω ; R 2 ) such that
∣ Z ∣ ≤ 1 a.e. in Ω , ∫ Ω Z ⋅ T D 2 h d x d y = 0 for every h ∈ V 0 . |Z|\leq1 \quad\text{a.e. in } \Omega, \qquad\int_\Omega Z\cdot TD^2h\,dx\,dy=0 \quad\text{for every } h\in V_0. ∣ Z ∣ ≤ 1 a.e. in Ω , ∫ Ω Z ⋅ T D 2 h d x d y = 0 for every h ∈ V 0 .
These fields provide lower bounds for the collision energy and lead to a dual characterization of the minimum and the minimizers in A A A .
Proposition 5.7 (Duality and optimality). Under Assumptions 2 and 3,
min ψ ∈ B J [ ψ ] = inf ψ ∈ A J [ ψ ] = 1 π max Z ∈ Z ∫ Ω Z ⋅ T D 2 ψ b d x d y . \min_{\psi\in\mathcal{B}} J[\psi]=\inf_{\psi\in\mathcal{A}}J[\psi]=\frac{1}{\pi}\max_{Z\in\mathcal{Z}}\int_{\Omega} Z\cdot TD^2\psi_b\,dx\,dy. ψ ∈ B min J [ ψ ] = ψ ∈ A inf J [ ψ ] = π 1 Z ∈ Z max ∫ Ω Z ⋅ T D 2 ψ b d x d y . The dual maximum is attained. A function ψ ∈ A \psi\in\mathcal{A} ψ ∈ A minimizes J J J if and only if there is a field Z ∈ L ∞ ( Ω ; R 2 ) Z\in L^\infty(\Omega;\mathbb{R}^2) Z ∈ L ∞ ( Ω ; R 2 ) such that
∣ Z ∣ ≤ 1 , Z ⋅ T D 2 ψ = ∣ T D 2 ψ ∣ a.e. in Ω , |Z|\leq1,\qquad Z\cdot TD^2\psi=|TD^2\psi|\quad\text{a.e. in }\Omega, ∣ Z ∣ ≤ 1 , Z ⋅ T D 2 ψ = ∣ T D 2 ψ ∣ a.e. in Ω , ∫ Ω Z ⋅ T D 2 h d x d y = 0 for every h ∈ V 0 . (34) \int_{\Omega}Z\cdot TD^2h\,dx\,dy=0\qquad\text{for every }h\in\mathcal{V}_0. \tag*{(34)} ∫ Ω Z ⋅ T D 2 h d x d y = 0 for every h ∈ V 0 . (34) Every dual maximizer satisfies the pointwise equality in (34) at every minimizer in A \mathcal{A} A .
Proof. Set g = T D 2 ψ b g=TD^2\psi_b g = T D 2 ψ b , and let
M : = { T D 2 h : h ∈ V 0 } ‾ L 1 ( Ω ; R 2 ) . M:=\overline{\{TD^2h:h\in\mathcal{V}_0\}}^{L^1(\Omega;\mathbb{R}^2)}. M := { T D 2 h : h ∈ V 0 } L 1 ( Ω ; R 2 ) .
Since M M M is a closed linear subspace,
π inf ψ ∈ A J [ ψ ] = inf h ∈ V 0 ∥ g + T D 2 h ∥ L 1 ( Ω ) = dist L 1 ( g , M ) . \pi\inf_{\psi\in\mathcal{A}}J[\psi]=\inf_{h\in\mathcal{V}_0}\|g+TD^2h\|_{L^1(\Omega)}=\operatorname{dist}_{L^1}(g,M). π ψ ∈ A inf J [ ψ ] = h ∈ V 0 inf ∥ g + T D 2 h ∥ L 1 ( Ω ) = dist L 1 ( g , M ) .
If g ∉ M g\notin M g ∈ / M , the Hahn–Banach theorem applied to L 1 ( Ω ; R 2 ) / M L^1(\Omega;\mathbb{R}^2)/M L 1 ( Ω ; R 2 ) / M gives a bounded linear functional on L 1 ( Ω ; R 2 ) L^1(\Omega;\mathbb{R}^2) L 1 ( Ω ; R 2 ) with norm one, vanishing on M M M , and taking the value dist L 1 ( g , M ) \operatorname{dist}_{L^1}(g,M) dist L 1 ( g , M ) at g g g . Representing this functional by a field in L ∞ ( Ω ; R 2 ) L^\infty(\Omega;\mathbb{R}^2) L ∞ ( Ω ; R 2 ) gives Z ∈ Z Z\in\mathcal{Z} Z ∈ Z with
∫ Ω Z ⋅ g d x d y = dist L 1 ( g , M ) . \int_{\Omega}Z\cdot g\,dx\,dy=\operatorname{dist}_{L^1}(g,M). ∫ Ω Z ⋅ g d x d y = dist L 1 ( g , M ) .
If g ∈ M g\in M g ∈ M , the same equality holds with Z = 0 Z=0 Z = 0 . Conversely, for every Z ∈ Z Z\in\mathcal{Z} Z ∈ Z and h ∈ V 0 h\in\mathcal{V}_0 h ∈ V 0 ,
∫ Ω Z ⋅ g d x d y = ∫ Ω Z ⋅ ( g + T D 2 h ) d x d y ≤ ∥ g + T D 2 h ∥ L 1 ( Ω ) . \int_{\Omega}Z\cdot g\,dx\,dy=\int_{\Omega}Z\cdot(g+TD^2h)\,dx\,dy\leq\|g+TD^2h\|_{L^1(\Omega)}. ∫ Ω Z ⋅ g d x d y = ∫ Ω Z ⋅ ( g + T D 2 h ) d x d y ≤ ∥ g + T D 2 h ∥ L 1 ( Ω ) .
Taking the infimum over h ∈ V 0 h\in\mathcal{V}_0 h ∈ V 0 gives
∫ Ω Z ⋅ g d x d y ≤ dist L 1 ( g , M ) for every Z ∈ Z . \int_{\Omega}Z\cdot g\,dx\,dy\leq\operatorname{dist}_{L^1}(g,M)\qquad\text{for every }Z\in\mathcal{Z}. ∫ Ω Z ⋅ g d x d y ≤ dist L 1 ( g , M ) for every Z ∈ Z .
The field constructed above achieves equality, so the dual maximum is attained and
max Z ∈ Z ∫ Ω Z ⋅ g d x d y = dist L 1 ( g , M ) = π inf ψ ∈ A J [ ψ ] . \max_{Z\in\mathcal{Z}}\int_{\Omega}Z\cdot g\,dx\,dy=\operatorname{dist}_{L^1}(g,M)=\pi\inf_{\psi\in\mathcal{A}}J[\psi]. Z ∈ Z max ∫ Ω Z ⋅ g d x d y = dist L 1 ( g , M ) = π ψ ∈ A inf J [ ψ ] .
The equality with the relaxed minimum follows from Theorem 5.6 .
For every ψ ∈ A \psi\in\mathcal{A} ψ ∈ A and Z ∈ Z Z\in\mathcal{Z} Z ∈ Z , the difference ψ − ψ b \psi-\psi_b ψ − ψ b belongs to V 0 \mathcal{V}_0 V 0 . Hence
J [ ψ ] − 1 π ∫ Ω Z ⋅ T D 2 ψ b d x d y = 1 π ∫ Ω ( ∣ T D 2 ψ ∣ − Z ⋅ T D 2 ψ ) d x d y ≥ 0. (35) J[\psi]-\frac{1}{\pi}\int_{\Omega}Z\cdot TD^2\psi_b\,dx\,dy
=\frac{1}{\pi}\int_{\Omega}\left(|TD^2\psi|-Z\cdot TD^2\psi\right)\,dx\,dy\geq0. \tag*{(35)} J [ ψ ] − π 1 ∫ Ω Z ⋅ T D 2 ψ b d x d y = π 1 ∫ Ω ( ∣ T D 2 ψ ∣ − Z ⋅ T D 2 ψ ) d x d y ≥ 0. (35)
If ψ \psi ψ is a minimizer and Z Z Z is a dual maximizer, the left-hand side vanishes. The nonnegative integrand therefore vanishes almost everywhere, proving necessity and the assertion concerning every dual maximizer. Conversely, if (34) holds, then for every φ ∈ A \varphi\in\mathcal{A} φ ∈ A , convexity gives
J [ φ ] − J [ ψ ] ≥ 1 π ∫ Ω Z ⋅ T D 2 ( φ − ψ ) d x d y = 0. J[\varphi]-J[\psi]\geq\frac{1}{\pi}\int_{\Omega}Z\cdot TD^2(\varphi-\psi)\,dx\,dy=0. J [ φ ] − J [ ψ ] ≥ π 1 ∫ Ω Z ⋅ T D 2 ( φ − ψ ) d x d y = 0.
Thus ψ \psi ψ is a minimizer. Equation (35) also shows that the corresponding Z Z Z is a dual maximizer. □
The dual objective is independent of the chosen lifting: two admissible liftings with the same trace differ by an element of V 0 V_0 V 0 . The pointwise conditions in (34) give
Z = T D 2 ψ ∣ T D 2 ψ ∣ a.e. on { T D 2 ψ ≠ 0 } . Z=\frac{T D^2\psi}{\lvert T D^2\psi\rvert}\quad\text{a.e. on }\{T D^2\psi\ne0\}. Z = ∣ T D 2 ψ ∣ T D 2 ψ a.e. on { T D 2 ψ = 0 } .
On { T D 2 ψ = 0 } \{T D^2\psi=0\} { T D 2 ψ = 0 } , they require only ∣ Z ∣ ≤ 1 \lvert Z\rvert\le1 ∣ Z ∣ ≤ 1 . For smooth fields, the weak condition follows from an interior equation and a natural boundary condition. The resulting criterion also applies to measure-valued derivatives in the relaxed class.
Proposition 5.8 (Smooth sufficient condition). Under Assumptions 2 and 3, let Z Z Z be a C 2 C^2 C 2 field on a neighborhood of Ω ‾ \overline{\Omega} Ω , with values in R 2 \mathbb{R}^2 R 2 , such that ∣ Z ∣ ≤ 1 \lvert Z\rvert\le1 ∣ Z ∣ ≤ 1 on Ω \Omega Ω and
2 ∂ x y Z 1 + ( ∂ x x − ∂ y y ) Z 2 = 0 in Ω , 2\partial_{xy}Z_1+(\partial_{xx}-\partial_{yy})Z_2=0\quad\text{in }\Omega, 2 ∂ x y Z 1 + ( ∂ xx − ∂ y y ) Z 2 = 0 in Ω , 2 Z 1 ν 1 ν 2 + Z 2 ( ν 1 2 − ν 2 2 ) = 0 a.e. on Γ . (36) 2Z_1\nu_1\nu_2+Z_2(\nu_1^2-\nu_2^2)=0\quad\text{a.e. on }\Gamma. \tag*{(36)} 2 Z 1 ν 1 ν 2 + Z 2 ( ν 1 2 − ν 2 2 ) = 0 a.e. on Γ. (36) Then
∫ Ω Z ⋅ d ( T D 2 h ) = 0 \int_{\Omega} Z\cdot\mathrm{d}(T D^2h)=0 ∫ Ω Z ⋅ d ( T D 2 h ) = 0 for every h ∈ W 0 1 , 1 ( Ω ) h\in W_0^{1,1}(\Omega) h ∈ W 0 1 , 1 ( Ω ) with T D 2 h ∈ M ( Ω ; R 2 ) T D^2h\in\mathcal{M}(\Omega;\mathbb{R}^2) T D 2 h ∈ M ( Ω ; R 2 ) . If ψ ∗ ∈ B \psi_*\in\mathcal{B} ψ ∗ ∈ B satisfies
∫ Ω Z ⋅ d ( T D 2 ψ ∗ ) = ∣ T D 2 ψ ∗ ∣ ( Ω ) , \int_{\Omega} Z\cdot\mathrm{d}(T D^2\psi_*)=\lvert T D^2\psi_*\rvert(\Omega), ∫ Ω Z ⋅ d ( T D 2 ψ ∗ ) = ∣ T D 2 ψ ∗ ∣ ( Ω ) , then ψ ∗ \psi_* ψ ∗ minimizes J \mathcal{J} J on B \mathcal{B} B . If ψ ∗ ∈ A \psi_*\in\mathcal{A} ψ ∗ ∈ A , it also minimizes J \mathcal{J} J on A \mathcal{A} A .
Proof. Let h ∈ W 0 1 , 1 ( Ω ) h\in W_0^{1,1}(\Omega) h ∈ W 0 1 , 1 ( Ω ) have T D 2 h ∈ M ( Ω ; R 2 ) T D^2h\in\mathcal{M}(\Omega;\mathbb{R}^2) T D 2 h ∈ M ( Ω ; R 2 ) , and set w = D ⊥ h w=D^\perp h w = D ⊥ h . By (15) , w ∈ B D ( Ω ) w\in BD(\Omega) w ∈ B D ( Ω ) , and div w = 0 \operatorname{div}w=0 div w = 0 . On Γ \Gamma Γ , w w w denotes its B D BD B D trace; see [2 Theorem 2.3] for the trace theorem and Gauss–Green formula. The Gauss–Green formula gives, for every η ∈ C ∞ ( Ω ‾ ) \eta\in C^\infty(\overline{\Omega}) η ∈ C ∞ ( Ω ) ,
∫ Γ η w ⋅ ν d s = ∫ Ω w ⋅ D η d x d y = ∫ Ω ( h y η x − h x η y ) d x d y = 0 , \int_{\Gamma}\eta\,w\cdot\nu\,\mathrm{d}s=\int_{\Omega}w\cdot D\eta\,\mathrm{d}x\,\mathrm{d}y=\int_{\Omega}(h_y\eta_x-h_x\eta_y)\,\mathrm{d}x\,\mathrm{d}y=0, ∫ Γ η w ⋅ ν d s = ∫ Ω w ⋅ D η d x d y = ∫ Ω ( h y η x − h x η y ) d x d y = 0 ,
where the last equality follows by first-order integration by parts and the zero trace of h h h . Therefore
w ⋅ ν = 0 a.e. on Γ . (37) w\cdot\nu=0\quad\text{a.e. on }\Gamma. \tag*{(37)} w ⋅ ν = 0 a.e. on Γ. (37)
Set
S : = ( Z 1 − Z 2 − Z 2 − Z 1 ) . S:=\begin{pmatrix} Z_1 & -Z_2 \\ -Z_2 & -Z_1 \end{pmatrix}. S := ( Z 1 − Z 2 − Z 2 − Z 1 ) .
By the definition of S S S ,
∫ Ω Z ⋅ d ( T D 2 h ) = ∫ Ω tr ( S d E w ) . \int_{\Omega} Z\cdot\mathrm{d}(T D^2h)=\int_{\Omega}\operatorname{tr}(S\,\mathrm{d}Ew). ∫ Ω Z ⋅ d ( T D 2 h ) = ∫ Ω tr ( S d E w ) .
Applying the Gauss–Green formula and then integrating by parts in h h h gives
∫ Ω Z ⋅ d ( T D 2 h ) = ∫ Γ ( S ν ) ⋅ w d s − ∫ Ω div S ⋅ D ⊥ h d x d y = ∫ Γ ( S ν ) ⋅ w d s + ∫ Ω h ( 2 ∂ x y Z 1 + ( ∂ x x − ∂ y y ) Z 2 ) d x d y . \begin{aligned}
\int_{\Omega} Z\cdot\mathrm{d}(T D^2h)
&=\int_{\Gamma}(S\nu)\cdot w\,\mathrm{d}s-\int_{\Omega}\operatorname{div}S\cdot D^\perp h\,\mathrm{d}x\,\mathrm{d}y\\
&=\int_{\Gamma}(S\nu)\cdot w\,\mathrm{d}s+\int_{\Omega}h\left(2\partial_{xy}Z_1+(\partial_{xx}-\partial_{yy})Z_2\right)\,\mathrm{d}x\,\mathrm{d}y.
\end{aligned} ∫ Ω Z ⋅ d ( T D 2 h ) = ∫ Γ ( S ν ) ⋅ w d s − ∫ Ω div S ⋅ D ⊥ h d x d y = ∫ Γ ( S ν ) ⋅ w d s + ∫ Ω h ( 2 ∂ x y Z 1 + ( ∂ xx − ∂ y y ) Z 2 ) d x d y .
The interior equation in (36) makes the last integral vanish. Since
( S ν ) ⋅ τ = − 2 Z 1 ν 1 ν 2 − Z 2 ( ν 1 2 − ν 2 2 ) = 0 a.e. on Γ , (S\nu)\cdot\tau=-2Z_1\nu_1\nu_2-Z_2(\nu_1^2-\nu_2^2)=0\quad\text{a.e. on }\Gamma, ( S ν ) ⋅ τ = − 2 Z 1 ν 1 ν 2 − Z 2 ( ν 1 2 − ν 2 2 ) = 0 a.e. on Γ ,
the vector S ν S\nu S ν is normal to Γ \Gamma Γ . Equation (37) therefore makes the boundary integral vanish as well.
For ψ ∈ B \psi\in B ψ ∈ B , apply this identity to h = ψ − ψ ∗ h=\psi-\psi_* h = ψ − ψ ∗ . The bound ∣ Z ∣ ≤ 1 |Z| \le1 ∣ Z ∣ ≤ 1 gives
π J [ ψ ] ≥ ∫ Ω Z ⋅ d ( T D 2 ψ ) = ∫ Ω Z ⋅ d ( T D 2 ψ ∗ ) = π J [ ψ ∗ ] . \pi\mathcal{J}[\psi] \ge\int_{\Omega} Z \cdot\mathrm{d}(T D^2\psi) = \int_{\Omega} Z \cdot\mathrm{d}(T D^2\psi_*) = \pi\mathcal{J}[\psi_*]. π J [ ψ ] ≥ ∫ Ω Z ⋅ d ( T D 2 ψ ) = ∫ Ω Z ⋅ d ( T D 2 ψ ∗ ) = π J [ ψ ∗ ] .
Thus ψ ∗ \psi_* ψ ∗ minimizes the relaxed energy. If ψ ∗ ∈ A \psi_* \in A ψ ∗ ∈ A , the equality J = J \mathcal{J}=J J = J on A A A gives the final assertion. □
The natural boundary condition in (36) reflects the freedom to prescribe the normal derivative of a smooth zero-trace variation on each smooth boundary segment.
Examples
The following examples use the optimality conditions in § 5.5 . We first study a radial flow, illustrating how radial compression generates collisions at constant density and proving that the radial profile is an exact minimizer. We then characterize the zero-energy minimizers for boundary traces generated by rotations and translations, and prove uniqueness on domains other than disks. The square example shows that uniqueness may fail when the minimum is positive: distinct minimizers can have different spatial distributions of local collision intensity. Finally, the square–disk comparison shows how domain geometry and the natural boundary condition determine whether a given flow is optimal.
Radial flow in an annular sector
We consider radial flow through an annular sector with prescribed inflow on the inner arc and outflow on the outer arc. As the flow spreads outward, incompressibility forces the radial speed to decrease with the radius. This deceleration creates radial compression alongside tangential expansion, producing a positive collision intensity even at constant density. We derive the radial velocity from incompressibility and the boundary flux, compute its collision energy, and prove that this radial flow minimizes the energy over the full admissible classes, including nonradial competitors.
Let 0 < ρ < R 0 < \rho< R 0 < ρ < R , 0 < α < 2 π 0 < \alpha< 2\pi 0 < α < 2 π , let e ( θ ) e(\theta) e ( θ ) be as in (1) , and consider
Ω = { r e ( θ ) : ρ < r < R , 0 < θ < α } . \Omega= \{re(\theta) : \rho< r < R,\ 0 < \theta< \alpha\}. Ω = { r e ( θ ) : ρ < r < R , 0 < θ < α } .
This is a bounded, simply connected, nonconvex Lipschitz domain whose boundary consists of two circular arcs and two radial sides. For α = π / 2 \alpha= \pi/2 α = π /2 , it is a quarter-annular passage.
Consider a radial velocity
v ( r e ( θ ) ) = u ( r ) e ( θ ) , u ∈ C 1 ( [ ρ , R ] ) . v(re(\theta)) = u(r)e(\theta), \qquad u \in C^1([\rho,R]). v ( r e ( θ )) = u ( r ) e ( θ ) , u ∈ C 1 ([ ρ , R ]) .
Its divergence is
div v = u ′ ( r ) + u ( r ) r = 1 r ( r u ( r ) ) ′ . \operatorname{div} v = u'(r) + \frac{u(r)}{r} = \frac{1}{r}(ru(r))'. div v = u ′ ( r ) + r u ( r ) = r 1 ( r u ( r ) ) ′ .
Incompressibility therefore requires r u ( r ) ru(r) r u ( r ) to be constant. For an outward flow, we write
u ( r ) = κ r , u ′ ( r ) = − κ r 2 < 0 , κ > 0. (38) u(r) = \frac{\kappa}{r}, \qquad u'(r) = -\frac{\kappa}{r^2} < 0, \qquad\kappa> 0. \tag*{(38)} u ( r ) = r κ , u ′ ( r ) = − r 2 κ < 0 , κ > 0. (38)
Thus the speed decreases from κ / ρ \kappa/\rho κ / ρ at the inner arc to κ / R \kappa/R κ / R at the outer arc. Figure 1 shows this decrease for ρ = 1 \rho= 1 ρ = 1 , R = 2 R = 2 R = 2 , and κ = 1 \kappa= 1 κ = 1 .
Figure 1. Radial velocity u ( r ) = 1 / r u(r)=1/r u ( r ) = 1/ r for 1 ≤ r ≤ 2 1\le r\le2 1 ≤ r ≤ 2 , with κ = 1 \kappa=1 κ = 1 . The outward speed decreases from u ( 1 ) = 1 u(1)=1 u ( 1 ) = 1 to u ( 2 ) = 1 / 2 u(2)=1/2 u ( 2 ) = 1/2 .
For the radial velocity field v = u ( r ) e ( θ ) v = u(r)e(\theta) v = u ( r ) e ( θ ) , the matrix of E v Ev E v in the orthonormal polar frame ( e ( θ ) , e ′ ( θ ) ) (e(\theta),e'(\theta)) ( e ( θ ) , e ′ ( θ )) is
E v = ( u ′ ( r ) 0 0 u ( r ) / r ) . Ev =
\begin{pmatrix}
u'(r) & 0 \\
0 & u(r)/r
\end{pmatrix}. E v = ( u ′ ( r ) 0 0 u ( r ) / r ) .
Its eigenvalues u ′ ( r ) u'(r) u ′ ( r ) and u ( r ) / r u(r)/r u ( r ) / r describe local compression or expansion in the radial and tangential directions, respectively. By Proposition 3.2 , the collision functional therefore has the radial representation
C Ω = α π ∫ ρ R r I ( diag ( u ′ ( r ) , u ( r ) r ) ) d r . (39) C_{\Omega}=\frac{\alpha}{\pi}\int_{\rho}^{R}r I\left(\operatorname{diag}\left(u'(r),\frac{u(r)}{r}\right)\right)\,dr. \tag*{(39)} C Ω = π α ∫ ρ R r I ( diag ( u ′ ( r ) , r u ( r ) ) ) d r . (39)
We prescribe the boundary flux generated by this field:
j = − κ ρ on the inner arc, j = κ R on the outer arc, j = 0 on the radial sides. \begin{aligned}
j&=-\frac{\kappa}{\rho} &&\text{on the inner arc,}\\
j&=\frac{\kappa}{R} &&\text{on the outer arc,}\\
j&=0 &&\text{on the radial sides.}
\end{aligned} j j j = − ρ κ = R κ = 0 on the inner arc, on the outer arc, on the radial sides.
The total inflow and outflow both equal κ α \kappa\alpha κ α .
A boundary primitive is j ˉ = κ θ \bar{j}=\kappa\theta j ˉ = κ θ , and we take the lifting
ψ b ( r e ( θ ) ) = κ θ . \psi_b(re(\theta))=\kappa\theta. ψ b ( r e ( θ )) = κ θ .
Since α < 2 π \alpha<2\pi α < 2 π and ρ > 0 \rho>0 ρ > 0 , the function θ \theta θ admits a smooth branch on a neighborhood of Ω ‾ \overline{\Omega} Ω . Thus Assumption 3 holds.
Among radial velocity fields, incompressibility and the prescribed boundary flux determine the flow uniquely. The following proposition shows that this flow also minimizes the collision energy over the full admissible classes.
Proposition 6.1 (Radial flow through a curved passage). Let 0 < ρ < R 0<\rho<R 0 < ρ < R , 0 < α < 2 π 0<\alpha<2\pi 0 < α < 2 π , and κ > 0 \kappa>0 κ > 0 . On
Ω = { r e ( θ ) : ρ < r < R , 0 < θ < α } , \Omega=\{re(\theta):\rho<r<R,\ 0<\theta<\alpha\}, Ω = { r e ( θ ) : ρ < r < R , 0 < θ < α } , let A \mathcal{A} A and B \mathcal{B} B be the admissible classes (10) and (21) with lifting ψ b ( r e ( θ ) ) = κ θ \psi_b(re(\theta))=\kappa\theta ψ b ( r e ( θ )) = κ θ . Then ψ ∗ ( r e ( θ ) ) = κ θ \psi_*(re(\theta))=\kappa\theta ψ ∗ ( r e ( θ )) = κ θ attains both minima, and
min ψ ∈ B J [ ψ ] = min ψ ∈ A J [ ψ ] = 2 κ α π log R ρ . \min_{\psi\in\mathcal{B}}J[\psi]=\min_{\psi\in\mathcal{A}}J[\psi]=\frac{2\kappa\alpha}{\pi}\log\frac{R}{\rho}. ψ ∈ B min J [ ψ ] = ψ ∈ A min J [ ψ ] = π 2 κ α log ρ R .
Proof. The function ψ ∗ \psi_* ψ ∗ is smooth on a neighborhood of Ω ‾ \overline{\Omega} Ω and has the prescribed trace, so ψ ∗ ∈ A \psi_* \in\mathcal{A} ψ ∗ ∈ A . Its velocity is
D ⊥ ψ ∗ ( r e ( θ ) ) = κ r e ( θ ) . D^\perp\psi_*(r e(\theta))=\frac{\kappa}{r}e(\theta). D ⊥ ψ ∗ ( r e ( θ )) = r κ e ( θ ) .
For this radial flow, the eigenvalues of E v E v E v are
u ′ ( r ) = − κ r 2 , u ( r ) r = κ r 2 . u'(r)=-\frac{\kappa}{r^2},\qquad\frac{u(r)}{r}=\frac{\kappa}{r^2}. u ′ ( r ) = − r 2 κ , r u ( r ) = r 2 κ .
Radial compression and tangential expansion balance, preserving local area. By Remark 3.3, the local collision intensity is
C ∗ ( r e ( θ ) ) = 1 π I ( E v ) = 2 π ∣ u ′ ( r ) ∣ = 2 κ π r 2 . \mathcal{C}_*(r e(\theta))=\frac{1}{\pi}I(Ev)=\frac{2}{\pi}|u'(r)|=\frac{2\kappa}{\pi r^2}. C ∗ ( r e ( θ )) = π 1 I ( E v ) = π 2 ∣ u ′ ( r ) ∣ = π r 2 2 κ .
Integration in polar coordinates gives
J [ ψ ∗ ] = 2 α π ∫ ρ R r ∣ u ′ ( r ) ∣ d r = 2 κ α π log R ρ . (40) J[\psi_*]=\frac{2\alpha}{\pi}\int_\rho^R r|u'(r)|\,dr=\frac{2\kappa\alpha}{\pi}\log\frac{R}{\rho}. \tag*{(40)} J [ ψ ∗ ] = π 2 α ∫ ρ R r ∣ u ′ ( r ) ∣ d r = π 2 κ α log ρ R . (40)
It remains to prove minimality over the full admissible classes. Define
Z ( r e ( θ ) ) = ( − cos ( 2 θ ) , sin ( 2 θ ) ) . Z(r e(\theta))=(-\cos(2\theta),\sin(2\theta)). Z ( r e ( θ )) = ( − cos ( 2 θ ) , sin ( 2 θ )) .
Then ∣ Z ∣ = 1 |Z|=1 ∣ Z ∣ = 1 , and direct differentiation gives
T D 2 ψ ∗ = 2 κ r 2 Z . T D^2\psi_*=\frac{2\kappa}{r^2}Z. T D 2 ψ ∗ = r 2 2 κ Z .
Since ρ > 0 \rho>0 ρ > 0 , Z Z Z is smooth on a neighborhood of Ω ‾ \overline{\Omega} Ω . Direct differentiation gives
2 ∂ x y Z 1 + ( ∂ x x − ∂ y y ) Z 2 = 0. 2\partial_{xy}Z_1+(\partial_{xx}-\partial_{yy})Z_2=0. 2 ∂ x y Z 1 + ( ∂ xx − ∂ y y ) Z 2 = 0.
The unit normal is ν = ± e ( θ ) \nu=\pm e(\theta) ν = ± e ( θ ) on the circular arcs and ν = ± ( − sin θ , cos θ ) \nu=\pm(-\sin\theta,\cos\theta) ν = ± ( − sin θ , cos θ ) on the radial sides. In both cases,
2 Z 1 ν 1 ν 2 + Z 2 ( ν 1 2 − ν 2 2 ) = 0. 2Z_1\nu_1\nu_2+Z_2(\nu_1^2-\nu_2^2)=0. 2 Z 1 ν 1 ν 2 + Z 2 ( ν 1 2 − ν 2 2 ) = 0.
Thus Z Z Z satisfies (36) . Since Z ⋅ T D 2 ψ ∗ = ∣ T D 2 ψ ∗ ∣ Z\cdot T D^2\psi_*=|T D^2\psi_*| Z ⋅ T D 2 ψ ∗ = ∣ T D 2 ψ ∗ ∣ , Proposition 5.8 shows that ψ ∗ \psi_* ψ ∗ minimizes both J J J over A \mathcal{A} A and J \mathcal{J} J over B \mathcal{B} B .
For κ < 0 \kappa<0 κ < 0 , the same conclusion holds with minimum
2 ∣ κ ∣ α π log R ρ . \frac{2|\kappa|\alpha}{\pi}\log\frac{R}{\rho}. π 2∣ κ ∣ α log ρ R .
The flow is inward, with inflow through the outer arc and outflow through the inner arc. Indeed, negating the stream function reverses the prescribed boundary flux and leaves both J J J and J \mathcal{J} J unchanged.
For α = π / 2 \alpha=\pi/2 α = π /2 , the minimum is ∣ κ ∣ log ( R / ρ ) |\kappa|\log(R/\rho) ∣ κ ∣ log ( R / ρ ) . Figure 2 shows this minimizer for ρ = 1 \rho=1 ρ = 1 , R = 2 R=2 R = 2 , and κ = 1 \kappa=1 κ = 1 . Since this domain is not a disk, Corollary 5.3 applies without normalization.
Figure 2. Radial flow through the quarter-annular passage 1 < r < 2 1 < r < 2 1 < r < 2 , 0 < θ < π / 2 0 < \theta< \pi/2 0 < θ < π /2 , with κ = 1 \kappa= 1 κ = 1 : streamlines of D ⊥ ψ ∗ D^\perp\psi_* D ⊥ ψ ∗ , where ψ ∗ = θ \psi_* = \theta ψ ∗ = θ (left), and local collision intensity (right). The flow points outward and its total collision energy is J [ ψ ∗ ] = log 2 J[\psi_*] = \log2 J [ ψ ∗ ] = log 2 .
Zero collision minimizers
Rigid rotations and translations preserve distances between particles and produce no local compression. Their symmetric velocity gradients vanish, so their collision energy is zero on any admissible domain. Since the energy is nonnegative, each such flow is a minimizer for the boundary data it generates.
By Lemma 4.1, every zero-energy stream function has the form
q ( x , y ) = α 2 ( x 2 + y 2 ) + β x + γ y + δ , α , β , γ , δ ∈ R . q(x,y)=\frac{\alpha}{2}(x^2+y^2)+\beta x+\gamma y+\delta,\qquad\alpha,\beta,\gamma,\delta\in\mathbb{R}. q ( x , y ) = 2 α ( x 2 + y 2 ) + β x + γ y + δ , α , β , γ , δ ∈ R .
Its velocity
D ⊥ q = ( α y + γ , − α x − β ) D^\perp q=(\alpha y+\gamma,-\alpha x-\beta) D ⊥ q = ( α y + γ , − α x − β )
is a rigid rotation combined with a constant translation. The remaining question is whether the prescribed trace allows more than one such flow. For the prescribed trace q ∣ Γ q|_\Gamma q ∣ Γ , every minimizer has zero energy, so uniqueness is determined by the zero-trace kernel K 0 K_0 K 0 .
Corollary 6.2 (Zero-energy minimizers). Let Ω ⊂ R 2 \Omega\subset\mathbb{R}^2 Ω ⊂ R 2 be a bounded simply connected Lipschitz domain, and let q q q be a polynomial of the form (12) . For the admissible classes A \mathcal{A} A and B \mathcal{B} B with lifting ψ b = q \psi_b = q ψ b = q ,
min ψ ∈ A J [ ψ ] = min ψ ∈ B J [ ψ ] = 0. \min_{\psi\in\mathcal{A}} J[\psi] = \min_{\psi\in\mathcal{B}} \mathcal{J}[\psi] = 0. ψ ∈ A min J [ ψ ] = ψ ∈ B min J [ ψ ] = 0. The minimizers in both classes are exactly the functions in q + K 0 q + K_0 q + K 0 . Thus q q q is the unique minimizer when Ω \Omega Ω is not a disk. On a disk, all minimizers have the form q + t q 0 q + tq_0 q + t q 0 , t ∈ R t \in\mathbb{R} t ∈ R , where q 0 q_0 q 0 is given by (14) .
Proof. Since T D 2 q = 0 T D^2q = 0 T D 2 q = 0 and q ∈ A q \in\mathcal{A} q ∈ A , both minima are zero. A function ψ ∈ B \psi\in\mathcal{B} ψ ∈ B is a minimizer if and only if T D 2 ψ = 0 T D^2\psi= 0 T D 2 ψ = 0 . Since ψ − q \psi- q ψ − q has zero trace, this is equivalent to ψ − q ∈ K 0 \psi- q \in K_0 ψ − q ∈ K 0 . Each such function also belongs to A \mathcal{A} A . The conclusions for disks and other domains follow from Proposition 4.2.
On a disk, the additional rotation generated by q 0 q_0 q 0 is tangent to the boundary and therefore preserves the prescribed flux. If Ω \Omega Ω is not a disk, then J [ ψ ] > 0 J[\psi] > 0 J [ ψ ] > 0 for every ψ ∈ B ∖ { q } \psi\in\mathcal{B} \setminus\{q\} ψ ∈ B ∖ { q } .
Minimizers on the square
The preceding corollary gives uniqueness on domains other than disks when the prescribed trace admits zero energy. Uniqueness need not hold when the minimum is positive. On the square, the saddle flow belongs to a family of distinct minimizers with the same prescribed boundary flux and total energy, but different local collision intensities.
For κ > 0 \kappa> 0 κ > 0 , the saddle stream function ψ = κ x y \psi= \kappa xy ψ = κ x y on Ω = ( 0 , 1 ) 2 \Omega= (0,1)^2 Ω = ( 0 , 1 ) 2 gives
D ⊥ ψ = ( κ x , − κ y ) , T D 2 ψ = ( 2 κ , 0 ) , C ≡ 2 κ π . D^\perp\psi= (\kappa x,-\kappa y), \qquad T D^2\psi= (2\kappa,0), \qquad C \equiv\frac{2\kappa}{\pi}. D ⊥ ψ = ( κ x , − κ y ) , T D 2 ψ = ( 2 κ , 0 ) , C ≡ π 2 κ .
Its total collision energy is J [ κ x y ] = 2 κ / π J[\kappa xy] = 2\kappa/\pi J [ κ x y ] = 2 κ / π .
For κ = 1 \kappa= 1 κ = 1 , the following proposition applies the optimality conditions of § 5.5 to a family of distinct minimizers with trace x y xy x y .
Proposition 6.3 (Non-uniqueness on the square). Let Ω = ( 0 , 1 ) 2 \Omega=(0,1)^2 Ω = ( 0 , 1 ) 2 , and let A x y \mathcal{A}_{xy} A x y and B x y \mathcal{B}_{xy} B x y be the admissible classes (4.7) and (5.2) with lifting ψ b = x y \psi_b=xy ψ b = x y . Then
min ψ ∈ A x y J [ ψ ] = min ψ ∈ B x y J [ ψ ] = 2 π . \min_{\psi\in\mathcal{A}_{xy}}\mathcal{J}[\psi]=\min_{\psi\in\mathcal{B}_{xy}}\mathcal{J}[\psi]=\frac{2}{\pi}. ψ ∈ A x y min J [ ψ ] = ψ ∈ B x y min J [ ψ ] = π 2 . Both minima are attained by every member of the family
ψ a ( x , y ) = x y + a sin ( π x ) sin ( π y ) , 0 ≤ a ≤ π − 2 . (41) \psi_a(x,y)=xy+a\sin(\pi x)\sin(\pi y),\qquad0\leq a\leq\pi^{-2}. \tag*{(41)} ψ a ( x , y ) = x y + a sin ( π x ) sin ( π y ) , 0 ≤ a ≤ π − 2 . (41) The stream functions and their local collision intensities are pairwise distinct as a a a varies.
Proof. Each ψ a \psi_a ψ a is smooth and has trace x y xy x y , so ψ a ∈ A x y \psi_a\in\mathcal{A}_{xy} ψ a ∈ A x y . Direct differentiation gives
T D 2 ψ a = ( 2 + 2 a π 2 cos ( π x ) cos ( π y ) , 0 ) . T D^2\psi_a=\left(2+2a\pi^2\cos(\pi x)\cos(\pi y),0\right). T D 2 ψ a = ( 2 + 2 a π 2 cos ( π x ) cos ( π y ) , 0 ) . The first component is strictly positive in the open square for 0 ≤ a ≤ π − 2 0\leq a\leq\pi^{-2} 0 ≤ a ≤ π − 2 . Hence every member has the same field Z = ( 1 , 0 ) Z=(1,0) Z = ( 1 , 0 ) , which satisfies the interior equation in (5.17). On each side of the square, the natural boundary condition reduces to Z 2 = 0 Z_2=0 Z 2 = 0 , which also holds.
Since Z ⋅ T D 2 ψ a = ∣ T D 2 ψ a ∣ Z\cdot T D^2\psi_a=\lvert T D^2\psi_a\rvert Z ⋅ T D 2 ψ a = ∣ T D 2 ψ a ∣ , Proposition 5.8 shows that every ψ a \psi_a ψ a minimizes both J \mathcal{J} J over A x y \mathcal{A}_{xy} A x y and J \mathcal{J} J over B x y \mathcal{B}_{xy} B x y . Direct integration gives
J [ ψ a ] = J [ ψ a ] = 2 π ∫ Ω ( 1 + a π 2 cos ( π x ) cos ( π y ) ) d x d y = 2 π . \mathcal{J}[\psi_a]=\mathcal{J}[\psi_a]=\frac{2}{\pi}\int_{\Omega}\left(1+a\pi^2\cos(\pi x)\cos(\pi y)\right)\,dx\,dy=\frac{2}{\pi}. J [ ψ a ] = J [ ψ a ] = π 2 ∫ Ω ( 1 + a π 2 cos ( π x ) cos ( π y ) ) d x d y = π 2 . The local collision intensities satisfy
C a ( x , y ) = 2 π ( 1 + a π 2 cos ( π x ) cos ( π y ) ) , C_a(x,y)=\frac{2}{\pi}\left(1+a\pi^2\cos(\pi x)\cos(\pi y)\right), C a ( x , y ) = π 2 ( 1 + a π 2 cos ( π x ) cos ( π y ) ) , ∥ C a ∥ L ∞ ( Ω ) = 2 π ( 1 + a π 2 ) . \lVert C_a\rVert_{L^\infty(\Omega)}=\frac{2}{\pi}(1+a\pi^2). ∥ C a ∥ L ∞ ( Ω ) = π 2 ( 1 + a π 2 ) . Thus distinct values of a a a give distinct stream functions and local collision intensities.
By positive homogeneity, multiplying these stream functions by κ > 0 \kappa>0 κ > 0 gives minimizers for the prescribed trace κ x y \kappa xy κ x y . By Proposition 4.2, K 0 = { 0 } \mathcal{K}_0=\{0\} K 0 = { 0 } on the square. Thus a trivial zero-trace kernel does not imply uniqueness of the minimizer. Figure 3 compares the endpoints a = 0 a=0 a = 0 and a = π − 2 a=\pi^{-2} a = π − 2 of the minimizing family. This comparison shows that a minimizer of total collision intensity need not minimize its spatial supremum.
Figure 3. Two minimizers from (41) with trace x y xy x y on ( 0 , 1 ) 2 (0,1)^2 ( 0 , 1 ) 2 : a = 0 a=0 a = 0 (upper row) and a = π − 2 a=\pi^{-2} a = π − 2 (lower row). Streamline colors (left) use the common speed scale [ 0 , 2 ] [0,\sqrt{2}] [ 0 , 2 ] , and collision intensities (right) use [ 0 , 4 / π ] [0,4/\pi] [ 0 , 4/ π ] . Both have energy 2 / π ≃ 0.636620 2/\pi\simeq0.636620 2/ π ≃ 0.636620 , with intensity suprema 2 / π 2/\pi 2/ π and 4 / π 4/\pi 4/ π , respectively.
The saddle flow on a disk
The optimality of a flow depends on the domain and the prescribed boundary data. The saddle flow in § 6.3 minimizes the collision energy on the square. On a disk, we construct a perturbation that preserves the prescribed boundary trace and strictly reduces the total collision energy. This difference in optimality is reflected in the natural boundary condition.
Let
Ω = { ( x , y ) ∈ R 2 : x 2 + y 2 < R 2 } , R > 0 , \Omega=\left\{(x,y)\in\mathbb{R}^2:x^2+y^2<R^2\right\},\qquad R>0, Ω = { ( x , y ) ∈ R 2 : x 2 + y 2 < R 2 } , R > 0 ,
and prescribe j ˉ = κ x y \bar{j}=\kappa xy j ˉ = κ x y on Γ \Gamma Γ , with κ > 0 \kappa>0 κ > 0 and lifting ψ b = κ x y \psi_b=\kappa xy ψ b = κ x y . The signed boundary flux is
j = ∂ τ j ˉ = κ R cos ( 2 θ ) . j=\partial_\tau\bar{j}=\kappa R\cos(2\theta). j = ∂ τ j ˉ = κ R cos ( 2 θ ) .
For the saddle stream function, the pointwise conditions in (5.15) force Z = ( 1 , 0 ) Z=(1,0) Z = ( 1 , 0 ) . This field satisfies the interior equation in (5.17). On the disk, however, ν = e ( θ ) \nu=e(\theta) ν = e ( θ ) , so the left-hand side of the natural boundary condition in (5.17) equals sin ( 2 θ ) \sin(2\theta) sin ( 2 θ ) . Thus the natural boundary condition fails. The following perturbation gives an explicit decrease of the energy.
For t ≥ 0 t \ge0 t ≥ 0 , define
ψ t ( x , y ) ≔ κ x y + t x y ( R 2 − x 2 − y 2 ) . \psi_t(x,y) \coloneqq\kappa xy+txy\left(R^2-x^2-y^2\right). ψ t ( x , y ) : = κ x y + t x y ( R 2 − x 2 − y 2 ) .
Every ψ t \psi_t ψ t is smooth and has trace j ˉ \bar{j} j ˉ , so these functions belong to the same admissible class and have the same prescribed boundary flux. Direct differentiation gives
T D 2 ψ t = ( 2 κ + 2 t ( R 2 − 3 x 2 − 3 y 2 ) , 0 ) . T D^2\psi_t=\left(2\kappa+2t\left(R^2-3x^2-3y^2\right),0\right). T D 2 ψ t = ( 2 κ + 2 t ( R 2 − 3 x 2 − 3 y 2 ) , 0 ) .
The first component is nonnegative when 0 ≤ t ≤ κ / ( 2 R 2 ) 0 \le t \le\kappa/(2R^2) 0 ≤ t ≤ κ / ( 2 R 2 ) , but changes sign for larger t t t . Retaining the absolute value, polar integration gives
J [ ψ t ] = 4 ∫ 0 R ∣ κ + t R 2 − 3 t r 2 ∣ r d r . J[\psi_t]=4\int_0^R\left|\kappa+tR^2-3tr^2\right|r\,dr. J [ ψ t ] = 4 ∫ 0 R κ + t R 2 − 3 t r 2 r d r .
In particular, J [ ψ 0 ] = 2 κ R 2 J[\psi_0]=2\kappa R^2 J [ ψ 0 ] = 2 κ R 2 .
Choose t c = 3 κ / ( 4 R 2 ) t_c=3\kappa/(4R^2) t c = 3 κ / ( 4 R 2 ) , giving a simple explicit competitor with a sign change in the first component of T D 2 ψ t c TD^2\psi_{t_c} T D 2 ψ t c . Then
C t c ( r ) = κ 2 π ∣ 7 − 9 r 2 R 2 ∣ . C_{t_c}(r)=\frac{\kappa}{2\pi}\left|7-9\frac{r^2}{R^2}\right|. C t c ( r ) = 2 π κ 7 − 9 R 2 r 2 .
which vanishes on the circle r = 7 R / 3 r=\sqrt{7}R/3 r = 7 R /3 . With u = r 2 / R 2 u=r^2/R^2 u = r 2 / R 2 , we obtain
J [ ψ t c ] = κ R 2 2 ∫ 0 1 ∣ 7 − 9 u ∣ d u = 53 36 κ R 2 = 53 72 J [ ψ 0 ] . \begin{aligned}
J[\psi_{t_c}]=\frac{\kappa R^2}{2}\int_0^1|7-9u|\,du \\
&=\frac{53}{36}\kappa R^2=\frac{53}{72}J[\psi_0].
\end{aligned} J [ ψ t c ] = 2 κ R 2 ∫ 0 1 ∣7 − 9 u ∣ d u = 36 53 κ R 2 = 72 53 J [ ψ 0 ] .
Thus ψ 0 = κ x y \psi_0=\kappa xy ψ 0 = κ x y is not a minimizer for this prescribed boundary trace on the disk. The admissible perturbation reduces the energy by approximately 26.39%, while increasing the peak collision intensity by 75%, since
∥ C t ∥ L ∞ ( Ω ) = 7 κ 2 π , ∥ C 0 ∥ L ∞ ( Ω ) = 2 κ π . \lVert\mathcal{C}_t\rVert_{L^\infty(\Omega)}=\frac{7\kappa}{2\pi}, \qquad\lVert\mathcal{C}_0\rVert_{L^\infty(\Omega)}=\frac{2\kappa}{\pi}. ∥ C t ∥ L ∞ ( Ω ) = 2 π 7 κ , ∥ C 0 ∥ L ∞ ( Ω ) = π 2 κ .
By Theorem 5.6 , this calculation gives
min ψ ∈ B J [ ψ ] = inf ψ ∈ A J [ ψ ] ≤ 53 36 κ R 2 . \min_{\psi\in\mathcal{B}}\mathcal{J}[\psi]=\inf_{\psi\in\mathcal{A}}\mathcal{J}[\psi]\leq\frac{53}{36}\kappa R^2. ψ ∈ B min J [ ψ ] = ψ ∈ A inf J [ ψ ] ≤ 36 53 κ R 2 .
The exact value is not determined by this comparison.
Note that all these competitors already lie in the normalized subspace ker P 0 \ker P_0 ker P 0 defined by (16) : since the kernel generator
q 0 ( x , y ) = x 2 + y 2 − R 2 2 q_0(x,y)=\frac{x^2+y^2-R^2}{2} q 0 ( x , y ) = 2 x 2 + y 2 − R 2
is radial and ψ t \psi_t ψ t is odd in x x x , we have
∫ Ω ψ t q 0 d x d y = 0 , P 0 ψ t = 0. \int_{\Omega}\psi_tq_0\,\mathrm{d}x\,\mathrm{d}y=0, \qquad P_0\psi_t=0. ∫ Ω ψ t q 0 d x d y = 0 , P 0 ψ t = 0.
Thus the energy reduction comes from a genuine interior deformation orthogonal to the rotational kernel K 0 K_0 K 0 .
Figure 4 compares t = 0 t=0 t = 0 and t = 3 / 4 = 0.75 t=3/4=0.75 t = 3/4 = 0.75 for R = κ = 1 R=\kappa=1 R = κ = 1 .
Figure 4. Saddle flow (t = 0 t=0 t = 0 , upper row) and an admissible competitor (t = 3 / 4 t=3/4 t = 3/4 , lower row) on the unit disk, with R = κ = 1 R=\kappa=1 R = κ = 1 . Streamlines (left) are colored by speed, with respective ranges [ 0 , 1 ] [0,1] [ 0 , 1 ] and [ 0 , 7 7 / 18 ] [0,7\sqrt{7}/18] [ 0 , 7 7 /18 ] . Local collision intensities (right) share the scale [ 0 , 7 / ( 2 π ) ] [0,7/(2\pi)] [ 0 , 7/ ( 2 π )] . At the same prescribed boundary flux, the energy decreases from 2 to 53 / 36 ≃ 1.4722 53/36\simeq1.4722 53/36 ≃ 1.4722 , a reduction of approximately 26.39%.
Conclusions and future directions
We introduced a collision functional for stationary flows and studied its minimization in the dense incompressible regime m ≡ 1 m \equiv1 m ≡ 1 , under prescribed boundary flux. On bounded simply connected Lipschitz domains, we proved existence of relaxed minimizers and exact relaxation in the strong W 1 , 1 ( Ω ) W^{1,1}(\Omega) W 1 , 1 ( Ω ) topology. Coercivity and compactness follow after removing the zero-trace kernel, with normalization required only on disks. The existence theorem allows concentration of the controlled derivatives and does not establish smoothness of minimizing velocities. For such relaxed solutions, collision intensity is represented by a measure.
The examples show that a trivial zero-trace kernel does not imply uniqueness. Distinct minimizers may have the same prescribed boundary flux and total energy while producing different spatial distributions of local collision intensity. They also illustrate how the geometry of the domain and the natural boundary condition affect optimality.
Natural extensions include the study of variable-density flows in addition to the dense incompressible setting considered here. Another direction is the optimization of domain geometry, inflow and outflow portions, and boundary flux distributions, for both constant and variable densities, to reduce the minimum total collision intensity while keeping the total inflow fixed. Shape optimization will also require geometric constraints that prevent domain degeneration. A mean-field-game extension would require specifying individual costs and comparing the resulting equilibria with the collision-minimizing flows studied here. CEMSE Division, King Abdullah University of Science and Technology (KAUST), Thuwal 23955–6900, Saudi Arabia
Email address: yuri.ashrafyan@kaust.edu.sa
CEMSE Division, King Abdullah University of Science and Technology (KAUST), Thuwal 23955–6900, Saudi Arabia
Email address: diogo.gomes@kaust.edu.sa