In this section we will carry out this procedure onT2 viewed as C
2πZ+2πiZ. Assume the number of particles is n=N2 for some integer N. Define the complex vector fields
Lk =iei(k1x+k2y)(k2
∂ ∂x −k1
∂
∂y), k1, k2 = 1, . . . , N.
EachLk corresponds to two real vector fields by taking the real and imaginary parts.
We find that
Additionally the vector fields {Lk} form an orthogonal basis for square integrable
divergence-free vector fields (see [Zei91])6. Now let X1, . . . , Xn be the sequence
L1,0 , L2,0 , . . . , LN,0
L1,1 , L2,1 , . . . , LN,1
..
. ,... , ,... L1,N , L2,N , . . . , LN,N.
We define the reconstruction mapping R(q,q) =˙ Pn
j=1cj(q,q)X˙ j, where the coeffi-
cients cj(q,q)˙ ∈ C are the solution to the inverse problem Pnj=1cj(q,q)X˙ j(qi) = ˙qi.
In matrix form, this is written as
˙
q= [w]·c,
where wij = Xj(qi), c = (c1, . . . , cn). This reconstruction mapping defines the hori-
zontal space above ϕ∈π−1(q) to be
H(ϕ) := span(X1◦ϕ, . . . , Xn◦ϕ)
and the horizontal lift to be
˙ q↑ϕ = n X j=1 cj(q,q)X˙ j ◦ϕ.
The kinetic energy of the particles is given by:
Lpart(q,q) =˙
n
X
j=1
kXjk2kcj(q,q)k˙ 2.
6Our integrator is different from [Zei91] because we choose a different closure method. In [Zei91]
the spatial velocity field was constrained to span(L1,1, . . . , LN,N) by equatingLk1+N,k2 withLk1,k2 (and similarly for Lk1,k2+N). Here we constrain spatial velocity to span(L1,1, . . . , LN,N) through holonomic constraints.
The magnitudes for each Xj can be obtained from the magnitude kLkk2 = 4π2kkk2.
This choice of horizontal space induces the horizontal projection
hor(ϕ, δϕ) = [δϕ()]↑ϕ
and the principal connection
A(ϕ, δϕ) = (δϕ−hor(ϕ, δϕ))◦ϕ−1.
Finally the reduced curvature form is ˜
B( ˙q, δq) =ϕ∗A [ ˙q↑ϕ, δq↑ϕ],
where we define the bracket onT SDiff(M) as the pullback of the standard Lie bracket on the Lie algebra. That is, [ ˙ϕ, δϕ] := [ ˙ϕϕ−1, δϕ◦ϕ−1]◦ϕ. Additionally, because
H(ϕ) is spanned by the complex vector fields Lk◦ϕ, it is useful to calculate the curvature
in this basis. We find:
B(Lk◦ϕ, Lj◦ϕ) = i(k1j2−j1k2)ϕ∗A(Lk+j ◦ϕ) if k1+j1 > N or k2+j2 > N 0 else .
To discretize time and calculate trajectories we can invoke the framework of “dis- crete Lagrangian mechanics” [MW01,MV91] by choosing the discrete Lagrangian:
Ld(q, q+) = Lpart q+q+ 2 , q+−q h .
The integrator is equal to the discrete Euler-Lagrange equations
D2Ld(q−, q) +D1Ld(q, q+) = 0,
which are then solved with a root-finding algorithm, such as Newton’s method. This produces a symplectic variational integrator with approximate conservation of energy
over large times and exact conservation of Noetherian momenta. This would be method A.
To implement method B, one can approximate the curvature force, iq˙Bµ, in dis-
crete time by substituting ˙q with 21h(q+−q−) to get a covector Fd, and then solving
the forced discrete Euler-Lagrange equations
D2Ld(q−, q) +D1(q, q+) =Fd.
Infinitesimal time error analysis The complex vector fields, {Lk}, serve as a
basis for square integrable vector fields on M. Given the initial condition u0 ∈
Xdiv(M) we set ˙q =u0() as the initial velocity for our particle method.7 Using the
fact thatR(q,q)(˙ i)−u0(i) = 0, we find that the reconstructed vector fieldR(q,q)˙
satisfies the error bound
kR(q,q)˙ −u0k∞ ≤ k∇u0k∞k∆xk,
where ∆x is the largest distance between neighboring particles.
Method B: We can get an error bound which is second order in time. By using the remeshing method of §3.4 to guarantee that ∆x remains below some some threshold ∆xmax, we notice that translating q to qnew does not alter the reconstructed vector
field. This is because we set:
˙
qnew :=R(q,q)(q˙ new)
and then we findR(qnew,q˙new)(qnew) = ˙qnew =R(q,q)(q˙ new). This last equation places
the same constraints on the space spanned by X1, . . . , Xn and so R(qnew,q˙new) =
7This would make the estimated spatial velocity field,R(q,q˙), less then optimal (in the Euclidean
2-norm) because a better approximation would be to orthogonally project the desired initial velocity, u0, onto the horizontal space. However, the “improved” approximation would induce an initial
condition which would lead to first order in time accuracy in predicting particle velocities. It is imperative to get particle velocities correct at timet= 0 in order to say anything meaningful about error for particle methods over infinitesimal times.
R(q,q). Therefore, remeshing leaves the estimated spatial velocity unaltered and can˙ be used at each time step to keep the particles within a distance ∆xmax without
penalty. This generalizes the idea of semi-Lagrangian methods by relieving the con- straint that we drag the data carried by the particles to predetermined nodes; instead, we may drag the data to whatever nodes we consider convenient. Finally, the error of method B will come from neglecting the curvature term, iq˙Bµ. If u is the exact
solution and Cu is a bound on kukover the time interval [0,∆t], then the magnitude
of the force satisfieskiq˙Bµk ≤Cu2Bmax, whereBmax is the supremum of the expression
“kivBµk” on unit vectorsv ∈T Qpart. If there exists a bound,C∇u, onk∇ukover the
time interval [0,∆t], then the bound, Bmax, on the missing curvature term produces
the second-order error bound
kR(q,q)˙ −uk∞≤(C∇u∆xmax)· Cu2Bmax
∆t2
for the reconstructed velocity field at time ∆t.
Method A: If the vertical component, ξq =u− R(q,q) = 0 at time˙ t= 0, then we
can implement method A by including the curvature term, iq˙Bµ( ˙q,0), to get an extra
order of accuracy. If we estimate the vertical component to be 0 for all time (as a closure method), then by equation (3.12) the error of this estimate would satisfy the bound k∆ξk< Cξkq˙0k∆t for some constant Cξ. The important thing to note is that
this error bound is first order in time. The error ∆ξ would introduce an error of size kiq˙B∆ξ( ˙q,0)k ∼O(∆t) in our estimate of ¨q. This would produce the error bound
kR(q,q)˙ −uk∞≤(C∇u∆xmax)· Cu2CξBmax
∆t3.
This makes method A third-order accurate in time.
A finite time error bound A second advantage to the geometric framework is that we can consider finite-time error bounds. That is to say, we can construct conservative bounds on the error over times of order 1. We are able to consider
this possibility because we know exactly what we are missing: the dynamics on E. We know that if ξ = 0 at time t = 0, then the exact equations of motion would satisfy ˙ξ = R(q,q)˙ · ∇(R(q,q))˙ − ∇p. However, in the context of methods A and B we are consistently neglecting ξ and effectively setting it to 0 for all time. Thus, ξ stores the error of our reconstructed vector field. Forming the quantity δe=kR(q,q)˙ · ∇(R(q,q))˙ − ∇pat each time step, we may construct an error bound at timeT given by
emax =
Z T
0
δedt.
Assuming e can be calculated or approximated within some tolerance, we can use it as a stopping criterion. Such stopping criteria are important when accuracy is desired over long times.