Hamiltonian Dynamics of Semiclassical Gaussian Wave Packets in

Electromagnetic Potentials
Nolan King and Tomoki Ohsawa

Department of Mathematical Sciences

arXiv:1908.11454v1 [math-ph] 29 Aug 2019

The University of Texas at Dallas

800 W Campbell Rd, Richardson, TX 75080-3021, USA
[email protected], [email protected]

Abstract
We extend our previous work on symplectic semiclassical Gaussian wave packet dynamics to incorpo-
rate electromagnetic interactions by including a vector potential. The main advantage of our formulation
is that the equations of motion derived are naturally Hamiltonian. We obtain an asymptotic expansion
of our equations in terms of ~ and show that our equations have O(~) corrections to those presented
by Zhou, whereas ours also recover the equations of Zhou in the case of a linear vector potential and
quadratic scalar potential. One and two dimensional examples of a particle in a magnetic field are given
and numerical solutions are presented and compared with the classical solutions and the expectation val-
ues of the corresponding observables as calculated by the Egorov or Initial Value Representation (IVR)
method. We numerically demonstrate that the O(~) correction terms improve the accuracy of the classi-
cal or Zhou’s equations for short times in the sense that our solutions converge to the expectation values
calculated using the Egorov/IVR method faster than the classical solutions or those of Zhou as ~ → 0.

1 Introduction
1.1 Motivation
Gaussian wave packets have historically been used to solve the time-dependent semiclassical Schrödinger
equation [3, 6–9, 13]. While the Schrödinger equation is computationally non-trivial to solve in the semi-
classical regime, those methods using the Gaussian wave packets provide an alternative set of differential
equations that may be solved for the time-dependent parameters of the Gaussian wave packet. The Gaussian
wave packet is an ansatz for an exact solution in the case of linear vector potentials with quadratic scalar
potentials (see Hagedorn [6]), and gives a good short time approximation of the solution for other potentials
in the semiclassical regime as shown by Zhou [26].
However, the set of differential equations of Zhou for the parameters is not a Hamiltonian system in
general. Given that the equations of motion for a classical particle is a Hamiltonian system and also that
the Schrödinger equation is a (infinite-dimensional) Hamiltonian system as we will explain in a moment, it
is rather natural to seek a Hamiltonian formulation of the dynamics of the Gaussian wave packet. This was
the main motivation of our previous work [24] on the symplectic/Hamiltonian formulation of the Gaussian
wave packet dynamics.
In this paper, we utilize the same symplectic-geometric framework to derive a Hamiltonian system of
equations governing the evolution the wave packet parameters under the influence of electromagnetic fields
by taking into account a vector potential. Semiclassical dynamics under the influence of electromagnetic fields
has been of great interest recently because of its significance in quantum control and solid state physics.

1
1.2 Hamiltonian Formulation of Classical Dynamics
It is well known that the equations of motion of a classical particle in Rd is a Hamiltonian system. From the
symplectic-geometric point of view, one takes the cotangent bundle T ∗ Rd ∼ = R2d = {(q, p) | q, p ∈ Rd } as the
phase space and define the classical symplectic form Ω0 := dqi ∧ dpi (the Einstein summation convention is
assumed). This renders T ∗ Rd a symplectic manifold. We also define a Hamiltonian function H0 : T ∗ Rd → R
as
1
H0 (q, p) := (p − A(q))2 + V (q), (1)
2m
where V : Rd → R and A : Rd → Rd are scalar and vector potentials respectively, and we set the charge to
be 1 for simplicity.
∂
Now let XH0 = q̇i ∂q∂ i + ṗi ∂p i
be the vector field on T ∗ Rd defined by iXH0 Ω0 = dH0 where i stands for the
contraction. Then the equation yields the equations of motion of the classical particle in the electromagnetic
field:
1 1
∇q |A(q)|2 − 2A(q) · p − ∇V (q),


q̇ = (p − A(q)), ṗ = − (2)
m 2m
where ∇q stands for the gradient with respect to the variable q, and | · | stands for the Euclidean distance
in Rd .

1.3 Hamiltonian Formulation of the Schrödinger Equation

We may generalize the notion of Hamiltonian system as follows: Let P be a symplectic manifold, i.e., a
manifold equipped with a closed non-degenerate 2-form Ω, and let H : P → R be a smooth function. Then
we define the Hamiltonian vector field XH on P corresponding to the Hamiltonian function H by setting
iXH Ω = dH. The vector field XH then defines the evolution equation on P . We may take it as the definition
of a generalized Hamiltonian system; see, e.g., Marsden and Ratiu [15] for more details.
We may now formulate the time-dependent Schrödinger equation as a Hamiltonian system as follows:
Let H := L2 (Rd ) with the standard (right-linear) inner product h · , · i, and equip it with the symplectic form
ΩH (ψ1 , ψ2 ) := 2~ Im hψ1 , ψ2 i, and take the expectation value hHi : H → R of the Hamiltonian operator as
the Hamiltonian function. Then the Hamiltonian vector field XhHi on H defined by iXhHi ΩH = dhHi yields
the usual time-dependent Schrödinger equation
∂
i~ ψ = Ĥψ, (3)
∂t

where Ĥ is the Hamiltonian operator defined below.

1.4 Geometry of Reduced Models

Given that the basic equations of classical and quantum dynamics are both Hamiltonian systems, it is natural
to expect that the basic equations of semiclassical dynamics—or more generally approximation/reduced
models of quantum dynamics—are Hamiltonian as well. Is there a way to exploit the above symplectic
structure ΩH on H = L2 (Rd ) to find a Hamiltonian formulation of reduced models?
Lubich [14] (see also Kramer and Saraceno [11]) came up with a general prescription to achieve this by
geometrically interpreting approximation models of quantum dynamics. Suppose that we have an ansatz
ϕ : M → H := L2 (Rd ); y 7→ ϕ(y; · ) for the solution of the Schrödinger equation, where M is a finite-
dimensional manifold (where the parameters for the ansatz live). The parameters y evolve in time according
to the dynamics in M to be determined, and the time evolution t 7→ ϕ(y(t); · ) gives an approximation
to the solution of the Schrödinger equation (3). Lubich [14] (see also Ohsawa and Leok [24]) showed that
one can achieve the best approximation in M as follows: Consider the embedding ι : M → H defined by
the ansatz ϕ as ι(y) := ϕ(y; · ). Then we can pull back the symplectic form ΩH to M to obtain a 2-form
Ω := ι∗ ΩH on M . Under a certain technical condition (see [14] and [24, Proposition 2.1] for details), Ω

2
defines a symplectic form
D on M . One E may also define the pull-back H := hHi ◦ ι of the Hamiltonian
function, i.e., H(y) := ϕ(y; · ), Ĥϕ(y; · ) . Then we may define the Hamiltonian vector field XH on M by
setting iXH Ω = dH. Lubich [14] showed that this gives the least squares approximation of the vector field
XhHi in the following sense: For any y ∈ M and any Vy ∈ Ty M ,

kXhHi (ι(y)) − Ty ι(XH (y))k ≤ kXhHi (ι(y)) − Ty ι(Vy )k

in terms of the L2 norm in H = L2 (Rd ); see Fig. 1.

Figure 1: The Hamiltonian Vector Field XH gives the best approximation on M of the vector field XhHi .

2 Hamiltonian Dynamics of Gaussian Wave Packets in Electro-

magnetic Potentials
2.1 Gaussian Wave Packet in Electromagnetic Potentials
Our ansatz or approximation/reduced model is the Gaussian wave packet of Heller [7, 8, 9] and Hagedorn
[3, 6] (see also Littlejohn [13]):
  
i 1
χM (q, p, A, B, φ, δ; x) := exp (x − q)T (A + iB)(x − q) + p · (x − q) + φ + iδ , (4)
~ 2
where (q, p) ∈ T ∗ Rd ∼
= R2d is the phase space center, φ ∈ S1 is the phase factor, δ ∈ R controls the norm,
and A + iB ∈ Σd := {A + iB ∈ Cd×d | A, B ∈ Symd (R), B − positive definite}. Note that the above Gaussian
is not normalized: r
(π~)d
 
2 2δ
N (B, δ) := kχ(y; · )k = exp − , (5)
det B ~
where we set y = (q, p, A, B, φ, δ). We will address this issue later.
Following the geometric picture of Lubich [14] described above, we define M to be the space of the above
parameters:
M := T ∗ Rd × Σd × S1 × R,
and consider the embedding ι : M → H defined as ι(y) := χM (y; · ). Then one can show that the pull-back
ΩM := ι∗ ΩH is in fact a symplectic form on M ; see Ohsawa and Leok [24, Section 3].
In this paper, we would like to incorporate the effect of electromagnetic fields to the dynamics of the
Gaussian wave packet. So we take Hamiltonian operator Ĥ to be
 2
1
Ĥ := − i~∇ − A(x) + V (x),
2m

3
where we assume that the scalar and vector potentials V : Rd → R and A : Rd → Rd are smooth functions; we
write the i-th component of Ai as opposed to the more conventional Ai in order to make it more conspicuous
and to avoid possible confusions with the components of A.
One may then evaluate the pull-back HM := hHi ◦ ι : M → R of the Hamiltonian hHi by evaluating the
expectation value of the Hamiltonian operator as follows:
D E
HM (y) := hHi ◦ ι(y) = χM (y; · ), ĤχM (y; · )
 2
p ~ 1
Tr B −1 (A2 + B 2 ) −


= N (B, δ) + hA(x) · pi (6)
2m 4m m

~ D  E 1

Tr DAT (x)AB −1 + |A(x)|2 + hV (x)i ,


+
2m 2m

where A is regarded as a column vector, DA(x) is the d × d matrix whose (i, j)-component is ∂A ∂xj (x), and
i

h · i stands for the expectation value of an observable with respect to the normalized Gaussian χM /kχM k:
For any smooth function U : Rd → R satisfying a certain growth condition (see Section 3.1),
1
hU(x)i := hχM (y; · ), U( · )χM (y; · )i
N (B, δ)
Z   (7)
1 1
= U(x) exp − (x − q)T B(x − q) dx.
N (B, δ) Rd ~

2.2 Hamiltonian Formulation of Gaussian Wave Packet Dynamics

One may now certainly formulate a Hamiltonian system on M using the above pull-backs of the symplectic
form and the Hamiltonian. However, the pull-back of the symplectic form turns out to be very cumbersome;
neither does it provide much insight into its relationship with the symplectic form Ω0 of classical dynamics;
see Ohsawa and Leok [24, Eq. (10)].
Fortunately, there is a way around it to obtain a simpler and more appealing formulation by exploiting
the inherent symmetry of the system [24, Section 4]. First observe that the Hamiltonian (6) does not depend
on the phase factor variable φ; that is, the Hamiltonian is invariant under the following S1 -action on the
manifold M :
S1 × M → M ; (θ, (q, p, A, B, φ, δ)) 7→ (a, p, A, B, φ + ~ θ, δ).
This action turns out to be symplectic and the corresponding momentum map (Noether conserved quantity)
is
N : M → R; N(y) := −~ N (B, δ).
It is natural to look at the level set N−1 (−~) because, in view of the definition (5) of N , this level set
corresponds to the choice of the parameter δ so that the Gaussian χM is normalized, i.e., kχ(y; · )k = 1.
Furthermore, one may eliminate the variables (φ, δ) from the formulation, because now we may apply the
Marsden–Weinstein reduction [16] (see also Marsden et al. [17, Sections 1.1 and 1.2]) to obtain the reduced
symplectic manifold
M := N−1 (−~)/S1 = T ∗ Rd × Σd .
See [24, Section 4] for the details of this reduction.
As a result, the symplectic form ΩM on M gives rise to the following symplectic form Ω~ on M :
~ −1 −1
Ω~ = dqi ∧ dpi + B B dAij ∧ dBkl
4 ik lj (8)
~ −1
= dqi ∧ dpi + dBij ∧ dAij .
4

4
Notice that the symplectic form is very simple and also appealing because it has an additional O(~) correction
term compared to the classical symplectic form Ω0 . It is also clear from the above expressions that B −1 and
A are conjugate variables. The corresponding Poisson Bracket is then
!
∂F ∂G ∂G ∂F 4 ∂F ∂G ∂G ∂F
{F, G}~ := − + −1 ∂A − −1 ∂A .
∂qi ∂pi ∂qi ∂pi ~ ∂Bjk jk ∂Bjk jk

Since we are now looking at the normalized Gaussian, we have N (B, δ) = 1, and thus the reduced
Hamiltonian H : M → R becomes
p2 ~ 1
Tr B −1 (A2 + B 2 ) −


H(q, p, A, B) = + hA(x) · pi
2m 4m m (9)
~ D  E 1

Tr DAT (x)AB −1 + |A(x)|2 + hV (x)i .


+
2m 2m
The Hamiltonian vector field XH on M defined by the Hamiltonian system iXH Ω~ = dH or equivalently
ȳ˙ = {ȳ, H}~ with ȳ = (q, p, A, B) gives the following set of ordinary differential equations:

1
q̇i = (pi − hAi (x)i) ,
m
1
~
−1
Di |A(x)|2 − 2 hDi Aj (x)pj i −


ṗi = − Ak (x)Akj Bji − hDi V (x)i ,
2m  2m 
1 2 2 1 2 1
2 2

Ȧij = − (A − B )ij + Dij Ak (x)pk − Dk Ai (x)Akj − Aik Dj Ak (x) − Dij |A(x)| , (10)
m m 2
~
2 −1

2 
− Dij (Dl Ak (x)Alm Bmk ) − Dij V (x) ,
2m
1 1
Ḃij = − (AB + BA)ij + hBik Dj Ak (x) + Dk Ai (x)Bkj i ,
m m
∂Ai 2 ∂Ak
where (DA)ij = Dj Ai = ∂xj , and Dij Ak = ∂xi ∂xj .

3 Asymptotic Expansion
3.1 Asymptotic Expansion of Hamiltonian
While the above set (10) of equations is Hamiltonian by construction, it has the drawback that it is not in a
closed form: The potential terms—involving either the scalar potential V or the vector potential A—appear
as expectation values (with respect to the normalized Gaussian). Unfortunately, it is impossible to explicitly
evaluate these expectation values unless V and A are polynomials.
So we apply Laplace’s method to obtain an asymptotic expansions of the integrals as ~ → 0 (see, e.g.,
Miller [18, Chapter 3]). Assuming that the Gaussian is normalized, i.e., N (B, δ) = 1, each potential term is
of the form (see (7)): Z  
1
hUi (q, B) = U(x) exp − (x − q)T B(x − q) dx,
Rd ~
As is proved in Ohsawa and Leok [24, Proposition 7.1] (see also Miller [18, Section 3.7]), if U satisfies a
certain growth condition as |x| → ∞, then hUi has the following asymptotic expansion:

~
Tr B −1 D2 U(q) + O(~2 )


hUi (q, B) = U(q) + as ~ → 0, (11)
4
where D2 U(q) is the Hessian matrix of U(x) evaluated at x = q. We note in passing that this asymptotic
expansion is exact (i.e., the O(~2 ) term vanishes) if U is quadratic.

5
 As a result, we have the following asymptotic expansion for the Hamiltonian H from (6):
H = H~ + O(~2 ) as ~→0
with
1
H~ (q, p, A, B) := (p − A(q))2
2m   
~ −1 2 2 T 2 1 2 2
+ Tr B A + B − DA (q)A − ADA(q) − D (A(q) · p) + D |A(q)| (12)
4m 2
~
+ V (q) + Tr B −1 D2 V (q) .


4
Notice that, just as for the symplectic form Ω in (8), this semiclassical Hamiltonian H has an additional
O(~) correction term compared to the classical Hamiltonian H0 from (1).
One may now replace the Hamiltonian H by H~ to define the Hamiltonian vector field XH~ as iXH~ Ω~ =
dH~ . Then the vector field XH~ yields
pi Ai (q) ~ −1 2
q̇i = − − B D Ai (q),
m m 4m jk kj 
1 ~ −1 2 ~ −1 2
ṗi = − Di |A(q)|2 + (Bjk Dkj |A(q)|2 ) − 2Ak (q)pk − Bjk Dkj Al (q)pl
2m 4 2
 
~ −1 2
− Di V (q) + (Bjk Dkj V (q)) , (13)
4
 
1 1 1 2
Ȧij = − (A2 − B 2 )ij + Di Ak (q)Akj + Aik Dj Ak (q) + Dij 2
Ak (q)pk − Dij |A(q)|2 − Dij
2
V (q),
m m 2
1 1
Ḃij = − (AB + BA)ij + (Di Ak (q)Bkj + Bik Dj Ak (q)).
m m
If we define those terms involving scalar and vector potentials with O(~) corrections as
~
V~ (q, B) := V (q) + Tr B −1 D2 V (q) ,


4
~ −1 2 ~
|A|2~ (q, B) := |A(q)|2 + Tr B −1 D2 |A(q)|2 ,



A~,i (q, B) := Ai (q) + Tr B D Ai (q) ,
4 4
we can rewrite the first two of the above set of equations in a slightly more succinct form:
1
q̇ = (p − A~ (q, B)) ,
m
1
∇q |A|2~ (q, B) − 2A~ (q, B) · p − ∇q V~ (q, B).


ṗ = −
2m
Notice also its similarity to the classical equations (2).

3.2 Linear Vector Potential with Quadratic Scalar Potential

As mentioned in the Introduction, when the vector potential A is linear (A(x) = Ax, where A is a constant
d × d matrix) and the scalar potential V is quadratic, the Gaussian wave packet (4) gives an exact solution
to the Schrödinger equation if the parameters satisfy the following set of equations (along with additional
equations for φ and δ):
1
q̇i = (pi − Ai (q)),
m
1
Di |A(q)|2 − 2Aj (q)pj − Di V (q),


ṗi = −
2m (14)
1 1 2 1
Ȧij = (Di Ak (q)Akj + Aik Dj Ak (q)) − Dij |A(q)|2 − (A2 − B 2 )ij − Dij
2
V (q),
m 2m m
1 1
Ḃij = (Bik Dj Ak (q) + Di Ak (q)Bkj ) − (AB + BA)ij .
m m

6
This result is a special case of the more general result of Hagedorn [6] on quadratic Hamiltonians, and
also is equivalent to the set of equations given by Zhou [26]. We note that both Hagedorn and Zhou use,
instead of (A, B), parameters (Q, P ) that are d × d complex matrices satisfying QT P − P T Q = 0 and
Q∗ P − P ∗ Q = 2iId ; more precisely, Hagedorn uses parameters A, B ∈ Cd×d , which are related to Q and P as
A = Q and B = −iP . In fact, these two sets of parameters are related by A + iB = P Q−1 ; see Ohsawa [23]
for the geometric interpretation of these two different parametrizations. It is straightforward calculations
using this relation to check that Zhou’s equations imply the above set of equations.
Our set of equations, either (10) or (13), recovers the above set of equations under the above assumptions
on the potentials. In fact, as mentioned above, the asymptotic expansion (11) is exact if U is quadratic. This
implies that the Hamiltonian (9) and its O(~2 ) approximation (12) coincide, and thus so do the equations (10)
and (13). Now, if the vector potential A is linear and the scalar potential V is quadratic, many of the terms
in (13) involving the Hessians of the potentials vanish, hence recovering (14).

4 Semiclassical Angular Momentum in Electromagnetic Poten-

tials
One advantage of the Hamiltonian formulation using the language of symplectic geometry is that it is
amenable to the geometric treatment of symmetry. Specifically, if the Hamiltonian function of the system is
invariant under some Lie group action, it is desirable to investigate any conserved quantities in the system
via Noether’s Theorem. Particularly, in this section, we show that the semiclassical angular momentum
found in our previous work [22] is conserved if the electromagnetic potentials possess a rotational symmetry.

4.1 Symmetry in Electromagnetic Potentials

Suppose that the scalar and vector potentials V and A possess the rotational symmetry in the following
sense: For any R ∈ SO(d) and any x ∈ Rd ,

V (Rx) = V (x) and A(Rx) = RA(x); (15)

that is, V is SO(d)-invariant whereas A is SO(d)-equivariant. We note that the latter condition implies
DA(Rx) = RA(x)RT for any R ∈ SO(d) and any x ∈ Rd .
As is done in [22], we define the action of the rotation group SO(d) on the symplectic manifold M =
∗ d
T R × Σd as follows:

(q, p, A, B) 7→ ΓR (q, p, A, B) := Rq, Rp, RART , RBRT .

Γ : SO(d) × M → M ;

It is easy to check that Γ is symplectic, i.e., Γ∗R Ω = Ω for any R ∈ SO(d). Then our Hamiltonian, either (9)
or (12), is invariant under this action, i.e., for any R ∈ SO(d), H ◦ ΓR = H and H~ ◦ ΓR = H~ . In fact, for
the Hamiltonian (12), it follows from a straightforward calculation using the above symmetry assumptions
on V and A. For the Hamiltonian (9), note that the expectation values of the potentials maintain the same
symmetry, i.e.,

hV i (Rq, RBRT ) = hV i (q, B), and hAi (Rq, RBRT ) = R hAi (q, B).

4.2 Semiclassical Angular Momentum

The momentum map J~ : M → so(d)∗ corresponding to the action Γ defined above is given by (see Ohsawa
[22, Section 3] for the derivation)

~ −1
J~ (q, p, A, B) = q  p − [B , A], (16)
2

7
where (q  p)ij = qj pi − qi pj (see Holm [10, Remark 6.3.3]), and we identified so(d)∗ with so(d) via an inner
product. Setting ~ = 0 reduces the above to the classical angular momentum, hence we call the above
the semiclassical angular momentum. Interestingly, this semiclassical angular momentum coincides with the
expectation value of the angular momentum with respect to the normalized Gaussian, i.e., for d = 3,

hx̂ × p̂i = J~ (q, p, A, B).

Now, assuming the symmetry (15) in the potentials, by Noether’s Theorem (see, e.g., Marsden and Ratiu
[15, Theorem 11.4.1]), we conclude that the semiclassical angular momentum (16) is a conserved quantity of
our semiclassical equation (10) or (13).

5 Numerical Examples
Given that our set of equations (13) differs from that of Zhou by O(~) correction terms, a natural question
is whether these correction terms improve the accuracy of approximation. Specifically, we are interested
in comparing the time evolution t 7→ z(t) = (q(t), p(t)) of the phase space variables of our semiclassical
equations with that of the expectation values hẑi of the position and momentum operators ẑ = (x̂, p̂), i.e.,
t 7→ hẑi(t) = hψ(t, · ), ẑψ(t, · )i, where t 7→ ψ(t, · ) is a solution of the Schrödinger equation (3).
In the following, we compare numerical solutions of the classical equations (2), the semiclassical equa-
tions (13), as well as the time-dependent expectation values of observables as calculated by the Egorov
[1, 2, 12] or Initial Value Representation [19–21, 25] (Egorov/IVR) method. Note that the time evolution of
(q, p) of Zhou’s equations (14) is identical to that of the classical equations (2).
We employ the Egorov/IVR method because it is computationally prohibitive to solve for the highly
oscillatory wave functions numerically in the semiclassical regime. It is also suited for our purposes because
we are interested in the time evolution of expectation values. In the Egorov/IVR method, the Wigner
function of the initial wave function is calculated. An observable is evaluated along the solutions of the
classical system for each sampled point in phase space where the phase space is sampled according to the
Wigner function. This gives √ an O(~2 ) approximation of the expectation value of that observable, with an
error proportional to 1/ N , where N is the number of samples.
In all of the following, we solved our equations and the classical equations by the explicit Runge–Kutta
method with a time step of 0.01. For the Egorov/IVR computations, we used 106 samples for each value of
~, with the exception of ~ = 0.01 for which we used 107 samples.

5.1 1D Example
Here we let d = 1, m = 1, V (x) = 1 − 21 cos2 (x), A(x) = cos(x), subject to the initial conditions q(0) = 0.5,
p(0) = −1, A(0) = 0, B(0) = 1; the scalar and vector potentials are taken from Zhou [26, Example 1]. In
order to see how the error converges as ~ → 0, we ran the computations for ~ = 0.5, 0.3, 0.1, 0.05, 0.03, 0.01.
Figure 2 shows the solutions on the classical phase space T ∗ R = R2 from t = 0 to t = 3 as well as the
error |hẑi(t) − z(t)| at t = 1.6 in terms of the Euclidean norm on the classical phase space. As can be seen,
our solutions are closer to the Egorov/IVR than the classical solutions. Furthermore, as ~ → 0, our solutions
converge to the Egorov/IVR solutions faster than the classical equations.
Figure 3 shows the time evolutions of the Hamiltonians for the classical, semiclassical, and Egorov/IVR
solutions. Note that the Hamiltonians for all these three cases are different: It is H0 in (1) for the classical
case and H~ in (12) for the semiclassical case, whereas for the Egorov/IVR case, it is the expectation value
hĤi of the Hamiltonian operator Ĥ. In each of these cases, the corresponding Hamiltonian is a conserved
quantity. Notice that the semiclassical Hamiltonian gives a better approximation to the expectation value
of the Hamiltonian.

8
 (a) ~ = 0.5 (b) ~ = 0.1

(c) ~ = 0.05 (d) ~ = 0.01

(e) Convergence of errors as ~ → 0

Figure 2: Results of 1D computations with m = 1, V (x) = 1 − 21 cos2 (x), A(x) = cos(x). (a)–(d): Parametric
plots of t 7→ q(t) = (q1 (t), q2 (t)) in the classical phase space T ∗ R ∼
= R2 for ~ = 0.5, 0.1, 0.05, 0.01 from t = 0
to t = 3. Our solutions are closer to the Egorov/IVR than the classical solutions. (e): The error |hẑi(t)−z(t)|
for several values of ~ at t = 1.6. As ~ → 0, our solutions converge to the Egorov/IVR solutions faster than
the classical equations. The equation of the best fit line for the semiclassical error is exp(−0.190) ∗ ~1.864 ,
and exp(−0.125) ∗ ~0.894 for the classical.

9
 (a) ~ = 0.5 (b) ~ = 0.1

(c) ~ = 0.05 (d) ~ = 0.01

Figure 3: Time evolution of the Hamiltonian for the above 1D system solutions for ~ = 0.5, 0.1, 0.05, 0.01.
The semiclassical Hamiltonian (12) more closely approximates the Egorov/IVR expectation value hĤi of the
Hamiltonian operator than the classical Hamiltonian (1).

10
5.2 2D Example
Here we let d = 2, V (x) = 12 |x|2 + 14 |x|4 , A(x) = (−x2 , x1 ), subject to the initial conditions q(0) = (1, 0),
   
−3 −6 1 1/2
p(0) = (0, 1), A(0) = , B(0) = .
−6 −6 1/2 1
Figure 4 shows the solutions on the classical configuration space R2 = {(q1 , q2 )} from t = 0 to t = 10
as well as the error |hẑi(t) − z(t)| at t = 2 in terms of the Euclidean norm on the classical phase space
T ∗ R2 ∼
= R4 . Figure 5 shows the time evolutions of the Hamiltonians for the classical, semiclassical, and
Egorov/IVR solutions just as in the 1D case. The same observations as above apply to these 2D results as
well.
For this 2D example, the scalar and vector potentials chosen above satisfy the symmetry condition (15).
Therefore, based on the result of Section 4, the semiclassical angular momentum (16) is also a conserved
quantity of the semiclassical system (13) as well. Figure 6 shows the time evolutions of the classical angular
momentum along the classical solutions, the semiclassical angular momentum (16) along the semiclassical
solutions, and the expectation value of the angular momentum operator along the Egorov/IVR solutions.
We see that the semiclassical angular momentum gives a better approximation to the expectation value of
the angular momentum than the classical one does.

6 Conclusion and Future Work

We extended our earlier work on the Hamiltonian formulation of Gaussian wave packets to incorporate
electromagnetic fields. Many of the results are extensions of our previous works to incorporate the electro-
magnetic effects. These results greatly expand the range of applications of semiclassical dynamics because
of its importance in quantum control and solid state physics.
As seen in the above numerical results, our solutions converge to the the expectation value of the operator
z = (q, p) along the Egorov/IVR solution faster than the classical solution. Since the equations for q and p
given by Zhou [26] are identical to the classical equations, our solutions also converge faster than those of
Zhou. These results demonstrate that the O(~) correction terms in our semiclassical equations (13) indeed
improve the accuracy of the approximations of expectation values.
Our preliminary studies (under certain technical assumptions and without electromagnetic fields) indicate
that the errors in the observables of the classical solution is O(~) whereas O(~3/2 ) for the semiclassical
solution, despite the well-known fact that the Gaussian wave packet dynamics gives O(~1/2 ) approximation
in terms of the wave functions in L2 -norm established by Hagedorn [3, 4, 5, 6]. Our numerical results seem
to support these claims. A proof of this error estimate remains for a future work.

References
[1] M. Combescure and D. Robert. Coherent States and Applications in Mathematical Physics. Springer,
2012.

[2] Y. V. Egorov. The canonical transformations of pseudodifferential operators. Uspekhi Mat. Nauk, 24
(5(149)):235–236, 1969.

[3] G. A. Hagedorn. Semiclassical quantum mechanics. I. The ~ → 0 limit for coherent states. Communi-
cations in Mathematical Physics, 71(1):77–93, 1980.

[4] G. A. Hagedorn. Semiclassical quantum mechanics. III. the large order asymptotics and more general
states. Annals of Physics, 135(1):58–70, 1981.

11
 [5] G. A. Hagedorn. Semiclassical quantum mechanics, IV: large order asymptotics and more general
states in more than one dimension. Annales de l’institut Henri Poincaré (A) Physique théorique, 42(4):
363–374, 1985.

[6] G. A. Hagedorn. Raising and lowering operators for semiclassical wave packets. Annals of Physics, 269
(1):77–104, 1998.

[7] E. J. Heller. Time-dependent approach to semiclassical dynamics. Journal of Chemical Physics, 62(4):
1544–1555, 1975.

[8] E. J. Heller. Classical S-matrix limit of wave packet dynamics. Journal of Chemical Physics, 65(11):
4979–4989, 1976.

[9] E. J. Heller. Frozen Gaussians: A very simple semiclassical approximation. Journal of Chemical Physics,
75(6):2923–2931, 1981.

[10] D. D. Holm. Geometric Mechanics, Part II: Rotating, Translating and Rolling. Imperial College Press,
2nd edition, 2011.

[11] P. Kramer and M. Saraceno. Geometry of the time-dependent variational principle in quantum mechan-
ics. Lecture notes in physics. Springer-Verlag, 1981.

[12] C. Lasser and S. Röblitz. Computing expectation values for molecular quantum dynamics. SIAM
Journal on Scientific Computing, 32(3):1465–1483, 2010.

[13] R. G. Littlejohn. The semiclassical evolution of wave packets. Physics Reports, 138(4-5):193–291, 1986.

[14] C. Lubich. From quantum to classical molecular dynamics: reduced models and numerical analysis.
European Mathematical Society, Zürich, Switzerland, 2008.

[15] J. E. Marsden and T. S. Ratiu. Introduction to Mechanics and Symmetry. Springer, 1999.

[16] J. E. Marsden and A. Weinstein. Reduction of symplectic manifolds with symmetry. Reports on
Mathematical Physics, 5(1):121–130, 1974.

[17] J. E. Marsden, G. Misiolek, J. P. Ortega, M. Perlmutter, and T. S. Ratiu. Hamiltonian Reduction by

Stages. Springer, 2007.

[18] P. D. Miller. Applied Asymptotic Analysis. American Mathematical Society, Providence, R.I., 2006.

[19] W. H. Miller. Classical S matrix: Numerical application to inelastic collisions. The Journal of Chemical
Physics, 53(9):3578–3587, 1970.

[20] W. H. Miller. Quantum mechanical transition state theory and a new semiclassical model for reaction
rate constants. The Journal of Chemical Physics, 61(5):1823–1834, 1974.

[21] W. H. Miller. The semiclassical initial value representation: A potentially practical way for adding
quantum effects to classical molecular dynamics simulations. The Journal of Physical Chemistry A, 105
(13):2942–2955, 2001.

[22] T. Ohsawa. Symmetry and conservation laws in semiclassical wave packet dynamics. Journal of Math-
ematical Physics, 56(3):032103, 2015.

[23] T. Ohsawa. The Siegel upper half space is a Marsden–Weinstein quotient: Symplectic reduction and
Gaussian wave packets. Letters in Mathematical Physics, 105(9):1301–1320, 2015.

12
[24] T. Ohsawa and M. Leok. Symplectic semiclassical wave packet dynamics. Journal of Physics A:
Mathematical and Theoretical, 46(40):405201, 2013.

[25] H. Wang, X. Sun, and W. H. Miller. Semiclassical approximations for the calculation of thermal rate
constants for chemical reactions in complex molecular systems. The Journal of Chemical Physics, 108
(23):9726–9736, 1998.

[26] Z. Zhou. Numerical approximation of the Schrödinger equation with the electromagnetic field by the
Hagedorn wave packets. Journal of Computational Physics, 272:386–407, 2014.

13
 (a) ~ = 0.5 (b) ~ = 0.1

(c) ~ = 0.05 (d) ~ = 0.01

(e) Convergence of errors as ~ → 0

Figure 4: Results of 2D computations with d = 2, V (x) = 21 |x|2 + 14 |x|4 , A(x) = (−x2 , x1 ). (a)–(d):
Parametric plots of t 7→ q(t) = (q1 (t), q2 (t)) in the classical configuration space R2 for ~ = 0.5, 0.1, 0.05, 0.01
from t = 0 to t = 3. (e): The error |hẑi(t) − z(t)| for several values of ~ at t = 2. Again, as ~ → 0, our
solutions converge to the Egorov/IVR solutions faster than the classical equations. The equation of the best
fit line for the semiclassical error is exp(1.544)~1.3612 , and exp(1.41845)~0.752 for the classical.

14
 (a) ~ = 0.5 (b) ~ = 0.1

(c) ~ = 0.05 (d) ~ = 0.01

Figure 5: Time evolution of the Hamiltonian for the above 2D system solutions for ~ = 0.5, 0.1, 0.05, 0.01.
The semiclassical Hamiltonian (12) more closely approximates the Egorov/IVR expectation value hĤi of the
Hamiltonian operator than the classical Hamiltonian (1).

15
 (a) ~ = 0.5 (b) ~ = 0.1

(c) ~ = 0.05 (d) ~ = 0.01

Figure 6: Time evolution of the classical angular momentum along the classical solution, the semiclassical
angular momentum along the semiclassical solution, and the expectation value of the angular momentum
operator for the 2D system. The semiclassical angular momentum is in far closer agreement than the angular
momentum along the classical solutions.

16