Quadrature integration for orthogonal wavelet systems - Rice...
Transcript of Quadrature integration for orthogonal wavelet systems - Rice...
Quadrature integration for orthogonalwavelet systems
Bruce R. Johnson,a Jason P. Modisette,b
Peter J. Nordlander,b and James L. Kinsey a
a Department of Chemistry and Rice Quantum InstituteRice University
Houston, TX 77251-1892
b Department of Physics and Rice Quantum InstituteRice University
Houston, TX 77251-1892
2
Abstract
Wavelet systems can be used as bases in quantum mechanical applications where
localization and scale are both important. General quadrature formulae are developed
for accurate evaluation of integrals involving compact support wavelet families, and their
use is demonstrated in examples of spectral analysis and integrals over anharmonic
potentials. In contrast to usual expectations for these uniformly-spaced basis functions, it
is shown that nonuniform spacings of sample points are readily allowed. Adaptive
wavelet quadrature schemes are also presented for the purpose of meeting specific
accuracy criteria without excessive oversampling.
3
I. Introduction
There is now an extensive wavelet literature within a variety of applications, e.g.,
digital signal processing, image processing, fingerprint classification, remote sensing,
target recognition, etc. 1-7 There has been some exploration of the use of wavelets in
chemical applications,8-20 but of only limited extent to date. Nevertheless, wavelet
methods promise a distinct advantage over Fourier methods for problems requiring
different levels of resolution in different locations, e.g., spectral analysis and
compression, molecular calculations of electronic or nuclear motion, etc. Regarded as
basis functions for calculation or analysis of quantum wave functions, wavelet systems
offer certain unusual properties. In fact, only with the introduction of the orthonormal
compact support wavelets by Daubechies1,3 did it become clear that there exist basis sets
that are capable of being simultaneously localized, multiscale, multicenter, and
orthonormal. The purpose of this paper is to provide efficient and accurate quadrature
methods for required integrals over compact support wavelet families and their
generalizations.
Such functions do not have simple closed forms, but are instead defined by
recursion between different scales. The original Daubechies bases are generated by two
related functions, the scaling function φ(x) and the wavelet ψ(x). Both are nonzero only
for 0 < x < L–1, where L is an even integer. Figure 1 shows these functions for L = 8.
From these functions one obtains others that are just copies widened by factors 2j and
translated by steps k 2j, where j and k are integers,
φkj x( ) = 2− j /2 φ 2− j x−k( ) , (1)
ψ kj x( ) = 2− j /2 ψ 2− j x− k( ) , (2)
4
with φ00 = φ , ψ 0
0 = ψ . By construction, the set of scaling functions at a particular
resolution level j, φkj{ }, form a basis which is orthogonal though not complete. A more
complete set φkj −1{ } is also orthogonal but is twice as dense. The difference in detail
between these two bases is spanned by the set of wavelets at level j, so that
φkj −1{ } = φ k
j{ }⊕ ψ kj{ } . This is expressed in terms of the recursion relations,
φkj x( ) = h ′ k φ2 k+ ′ k
j−1 x( )′ k = 0
L −1
∑ , (3)
ψ kj x( ) = g ′ k φ2 k + ′ k
j −1 x( )′ k =0
L −1
∑ , (4)
where the constants are related by gk = − 1( )k +1hL− 1− k .2 Ultimately, every quantity that
must be calculated boils down to some function of the hk, so that the latter take on a
central role in calculations. An orthonormal multiresolution basis can be comprised of
scaling functions φkJ{ } on a coarsest level j = J and wavelet functions ψ k
j{ } from this
coarsest level down to a finest scale j = 0 which provide the various additional octaves
of detail. The approximate expansion of a function f in this basis then takes the form
f x( ) ≈ ψ kj f ψ k
j x( )k
∑j = 0
J
∑ + φkJ f φ k
J x( )k
∑ . (5)
Detail functions ψ kj{ } at each scale which are unimportant (have small expansion
coefficients ψ kj f ) may simply be dropped, leaving a finite basis with customized
resolution. This immediately generalizes to multidimensional Cartesian cases by using
products of such basis functions.
For quantum applications, the needs arise to calculate the projection integrals
φ kj f , ψ k
j f for the vector basis and the matrix elements φ ′ k ′ j f φ k
j , φ ′ k ′ j f ψ k
j ,
5
ψ ′ k ′ j f ψ k
j , where f may be a general function of x . Use of the recursion relations in
Eqs. (3) and (4) allows all of these integrals to be expressed in terms of integrals with a
scaling function (or functions) on a single finest scale, but accurate methods are still
required for calculation of the remaining set of integrals. One method mentioned by
Daubechies3 (p. 166) for calculation of the vector projection integrals uses Fourier
convolution. While this method is not widely used, Wei and Chou17 have adapted it to
their density functional calculations. Alternatively, a quadrature method with uniformly-
spaced sample points was derived for the vector elements by Sweldens and Piessens.21,22
The latter method was used by Modisette, et al.,15 for wavelet expansion of a semi-
singular double-well potential; by expanding the wave functions in the same basis, matrix
elements could then be evaluated in terms of exactly-calculable "connection coefficients,"
or integrals of products of three functions from the wavelet basis.7,23,24 The portability of
this approach is complicated by the large number of connection coefficients generally
required, however. A recent adaptive wavelet collocation method developed for O(N)
solution of PDE's 25 directly addresses calculation of matrix elements, though a linear
system of equations depending on the sample grid must generally be solved.
In the present paper, parallel treatments are given of polynomial-based quadrature
formulae for direct calculation of φ kj f and φ ′ k
j f φ kj . An unusual aspect is that there
is no requirement that sample points lie on a uniformly spaced grid. For a given uniform
or irregular grid, the quadrature weights are explicitly expressed in terms of Lagrange
interpolating polynomials and polynomial moments of the scaling functions. The latter
are independent of the grid choice and need only be calculated once. Furthermore, use of
the scaling function recursion in Eq. (3) leads naturally to adaptive evaluation whereby a
6
chosen accuracy goal can be reached without excessive over-sampling. Straightforward
generalizations of these developments also hold for biorthogonal compact support
wavelet systems.26
II. Projection Integral Quadrature
For a given j and k, a change of variables transforms the general vector integral
for scaling functions into a standardized form,
φ kj f = 2− j /2 dx φ 2− j x−k( )∫ f x( ) = 2 j / 2 dx φ x( )∫ f 2 j x + k( )[ ]. (6)
Thus it is only necessary to consider the j = k = 0 case, for which the integral is
approximated by quadrature,
dx φ x( )∫ f x( ) ≈ ω q f xq( )q =1
r
∑ . (7)
The only restriction imposed on the sample points is that the xq are assumed to be
ordered, x1 < x2 … < xr. One can specify theωq in terms of the xq by requiring that Eq.
(6) be exact for f equal to a polynomial of order r–1 or less:
ωq xqp
q= 1
r
∑ = mp ≡ dx φ x( )∫ x p , 0 ≤ p ≤ r −1. (8)
The general solution to these equations (see Appendix A) is
ωq =mp
p!p= 0
r− 1
∑ Lr,qp( ) 0( ) , (9)
using the Lagrange interpolating polynomials
Lr , q x( ) =x − x ′ q
xq − x ′ q ′ q ≠ q
r
∏ (10)
7
and their derivatives Lr , qp( ) at x = 0 . The Lagrange polynomials are the coefficients for
interpolation of smooth functions,27
Lr ,q x( ) f xq( )q= 1
r
∑ ≈ f x( ), (11)
with equality holding for f(x) a polynomial in x of order less than r. Their connection
with numerical quadrature is well known,28,29 although the simple result Eq. (9) (which is
not really restricted to wavelet systems) appears to be new. Lagrange polynomials are
also familiar in wavelet contexts from their use in interpolatory refinement schemes, e.g.,
the symmetric iterative interpolation process of Deslauriers and Dubuq.30 The latter
scheme uses a "fundamental solution" that was subsequently shown31 to coincide with
the autocorrelation function of the compact support scaling functions (see Appendix B).
The Lagrange polynomials29,32 and their derivatives are easily evaluated once the
quadrature nodes are specified. As for the moments mp, their calculation may be
accomplished by the simple recursive equation derived by Gopinath and Burrus,33
mp = 1− 2− p( )−1 p
′ p
m ′ p µp− ′ p
′ p = 0
p −1
∑ , (12)
where the µp are discrete moments of the scaling coefficients of Eq. (3),
µp =1
2hk k p
k =0
L−1
∑ . (13)
The leading moment m0 = 1 as a matter of normalization, providing the starting point for
the recursion. (Not all of the resulting moments mp are independent of each other, as is
discussed in Appendix B.) The moments are thus determined completely by the scaling
coefficients hk, whereas the quadrature coefficients are determined in terms of both the hk
8
and the nodes xq. In calculations for a given compact support basis, the moments need
only be calculated once.
Sweldens and Piessens21 considered the important case of uniformly-spaced xq.
Such an arrangement allows the sample points for scaling functions of adjacent k to
overlap, requires only a single set of quadrature coefficients for all k, minimizes the
number of function evaluations needed, and yields a higher order error of O ∆x r( ), where
∆x is the spacing. This is a Newton-Cotes type of quadrature28 for which is allowed a
single shift of the entire set of sample points with respect to the origin; this degree of
freedom can be used to make the quadrature exact for polynomials up to order r instead
of r – 1. One may achieve even greater accuracy by defining a Gauss quadrature for
which all r of the xq are chosen (non-uniformly) so as to exactly include polynomials up
to order 2r – 1; in this case there is no overlap between the sample points for adjacent
scaling functions. In the present work, the disposition of the xq is left open for purposes
of adaptability.
For compact support wavelets, it is known21,33 that m2 = m12, although higher
moments are not simple powers of m1 (Appendix B). Introducing the differences
∆mp = mp −m1p , (14)
Eq. (8) can be rearranged,
ωq =m1
p
p!p= 0
r − 1
∑ Lr, qp( ) 0( )+
∆mp
p!p= 3
r − 1
∑ L r,qp( ) 0( )
= Lr ,q m1( ) +∆mp
p!p= 3
r − 1
∑ Lr, qp( ) 0( ) , (15)
9
where the first series has been recognized as the Taylor expansion of the Lagrange
interpolation polynomial evaluated at m1. To the extent that the second series can be
ignored, Eqs. (7) and (11) show that
dx φ x( )∫ f x( ) ≈ Lr ,q m1( ) f xq( )q= 1
r
∑ ≈ f m1( ) (16)
This directly reflects the fact determined by Gopinath and Burrus that a one-point
quadrature using m1 as the sample point has O ∆x3( ) error. Higher order error generally
requires a multi-point quadrature. An exception is found in the Coifman wavelet
family,34,35 where m1 = m2 = …= mM -1 = 0 for chosen integer M, leading to a one-point
quadrature error of O ∆x M( ) .
For many applications, it is worthwhile to use a more flexible framework than
orthogonal wavelet families. In biorthogonal families,26 the functions φ and ψ are
orthogonal to dual functions ˜ φ and ˜ ψ rather than to themselves. The scaled and
translated versions of these functions, ˜ φ kj and ˜ ψ k
j , satisfy their own scaling relations, cf.,
Eqs. (3) and (4),
˜ φ kj x( ) = ˜ h ′ k
˜ φ 2k+ ′ k j−1 x( )
′ k ∑ , (17)
˜ ψ kj x( ) = ˜ g ′ k
˜ φ 2k+ ′ k j−1 x( )
′ k ∑ , (18)
The extra freedom within biorthogonal families allows, for instance, φ and ψ to possess
definite symmetry, an impossibility for the orthogonal compact support families.1,3 It
also provides the framework for the Lifting scheme introduced by Sweldens36,37 for
10
purposes of adaptable multiresolution analyses. The general expansion of a function in a
biorthogonal basis may be expressed as
f x( ) ≈ ˜ ψ kj f ψ k
j x( )k
∑j = 0
J
∑ + ˜ φ kJ f φ k
J x( )k
∑ (19)
rather than Eq. (5). Similarly the analogs of Eqs. (7) and (8) are
dx ˜ φ x( )∫ f x( ) ≈ ˜ ω q f xq( )q =1
r
∑ , (20)
˜ ω q xqp
q =1
r
∑ = ˜ m p ≡ dx ˜ φ x( )∫ x p , 0 ≤ p ≤ r −1. (21)
The structure of the equations is otherwise unchanged so that the quadrature coefficent
expressions immediately adapt to the biorthogonal wavelet case.
III. Wavelet Transform of UV Ozone Absorption Spectrum
One of the potential uses of wavelet analysis is to molecular spectra with complex
features corresponding to action on different scales. The ozone molecule provides a
particular example in its Hartley UV absorption band (see Johnson, et al.,38-40 and
references therein), as shown in Fig. 2. Photoexcitation to a steeply–repulsive
dissociative region of the potential energy surface for the ˜ D 1B2 state leads to a single
intense band which is continuous across several thousand cm–1. In addition, smaller
features appear across much of the band as minor irregular structure with average spacing
250 cm–1. Under Fourier transformation, the time domain autocorrelation function
exhibits completely–resolved features at shorter (< 10 fs) and longer (40–150 fs) times
corresponding, respectively, to direct and indirect dissociation contributions. Dynamical
analysis of the indirect contribution, corresponding to Franck-Condon-region exit and
11
return, has progressed in recent years with the aid of both classical40 and quantum41-43
propagations. More definitive information is currently being obtained by Le, et al.,44, in
the form of detailed Raman excitation profiles for the different accessible final
vibrational states; these exhibit structure that reflects upon the same excited-state
dynamics, but in a vibration-specific manner.
One limitation of the Fourier transform (FT) is that it completely defocuses
information about the old variable ν. The tremendous breadth of the total band implies
that there are many different energetic regimes mixed together in the autocorrelation
function, and that a joint time-frequency analysis with partial resolution in both variables
is to be preferred. The first such application to ozone used a sliding window on the
spectral data prior to the FT operation,39 thereby allowing the following of time features
between partially-localized regions of frequency. Gabor45 and Vibrogram/Husimi46
transform techniques have also proven to be valuable alternatives for time-frequency
analysis of classical and quantum dynamics in other systems, though they have not been
applied to ozone. In the present work, an alternative separation of the distinct dynamical
regimes is obtained through wavelet multiresolution analysis of the absorption spectrum.
Using one of the orthonormal Daubechies wavelet families (L = 8), it is possible
to calculate projection integrals of translated and scaled scaling functions on a level j = 0,
dν1
∆νφ
ν − ν0
∆ν− k
∫ σ ν( )
= ∆ν dx φ x( )∫ σ ν0 + k∆ν + x∆ν( )
≈ ∆ν ωqσ ν0 + k∆ν + xq∆ν( )q=1
r
∑ . (22)
12
An absolutely typical initial complication arises from the fact that the absorption cross-
section σ was measured47 in uniform wavelength steps of 0.1 Å, i.e., the frequency
variable ν has non-equidistant spacings. One may perform interpolation of different
types in order to obtain uniform steps in ν, but this is unnecessary if the quadrature
described above is used. A total of r irregularly–spaced experimental points (with
average spacing close to ∆ν) were selected from within the support of each φk0 and the
associated projections then calculated by the rules described above. With the average
spacing chosen to be 3 cm–1, comparison of integrals for r = 7 and r = 8 generally showed
convergence to 5 significant figures. This reflects, of course, a combination of both the
inherent quadrature errors and any point-to-point experimental uncertainties.
Using the recursions in Eqs. (3) and (4), the projections onto an orthonormal basis
of wavelets ψ kj{ }, j ≤ J, and scaling functions φk
J{ } on level J may then be calculated in
the normal fashion. In this way, the spectrum is decomposed simultaneously into
different scales and positions. The smoothed version of the absorption cross-section
shown in Fig. 2 was obtained by summing over all k the projections onto the φkJ for J = 8.
This partial reconstruction process requires an algorithm that accurately calculates the
values of the scaling functions at particular points (see Burrus, et al.6). Figure 3 shows
the corresponding summed ψ kj projections, representing the differences in detail between
different scales j. It is seen that the spectral structure is separated from the smooth
component and is concentrated primarily in three octaves of scale change spanned by the
wavelets for j = 5–7. One obtains localized frequency information even from these
summed quantities, e.g., it is clear from j = 5 that higher-frequency components of the
structure are more prominent on the low-frequency side of band center.
13
The corresponding wavelet coefficients shown in Fig. 4 of course contain much
more detailed local frequency information. All of the information so obtained is
preserved under Fourier transformation of the wavelets and scaling functions to the time
domain, so that one can determine how the longer-time features in the autocorrelation
function change with both scale and general frequency region, subject only to the
constraints of the Uncertainty Principle. (Visual analysis, e.g., for comparison with the
windowed FT analysis of Johnson and Kinsey,39 would be simpler using the Continuous
Wavelet Transform3 although the latter is redundant and less efficient to implement.) A
more important advantage of this type of decomposition should be realized in the analysis
of the Raman experiments in ozone. Transformation of a Raman excitation profile to the
time domain is not a simple linear operation since the phase is not measured and must be
recovered using Hilbert transform techniques.48-52 The wavelet decomposition should
prove useful here for separation of the early and later time domain features; from these
results the shapes of the excited state potential and transition dipole in the Franck-Condon
geometry region and beyond can be determined53,54 (see also Shapiro55).
It merits emphasis that the irregular sampling is accommodated easily in this
application. This type of irregularity stands in contrast to that in the Lifting scheme,36,37
where the biorthogonal basis itself is tailored. In the present case, the underlying wavelet
basis is regularly-spaced as usual, but the sample points are not tied to that arrangement.
The numerical quadrature performed for the expansion coefficients on the finest scale
does not change the O N( ) nature of the wavelet recursion, but does increase the work
needed for initialization by a factor proportional to the quadrature order r. In the
language of filter banks used in engineering contexts, this corresponds to a "prefiltering"
14
operation on the data,5,56,57 and the quadrature scheme used here represents a polynomial
prefilter for irregularly-spaced samples.
IV. Morse Potential Expansion and Adaptive Quadrature
In wavelet applications, easy compression is one of the usual goals. If a particular
wavelet coefficient is calculated to be small, that element of the basis may be eliminated.
On the other hand, efficiency and/or storage in numerical applications may require that
these coefficients should not even be calculated if they are known to be small, for
instance, if overlapping elements from coarser scales have already been excluded. This
argues for a refinement approach3,30 where finer scales are used only as needed by the
problem under investigation. In this section, the results of Section II are embedded in an
adaptive wavelet quadrature.
A prototypical anharmonic molecular potential function is given by the Morse
potential,
V x( ) = De exp − 2a x − x e( )[ ]−2De exp − a x − xe( )[ ], (23)
characterized by a steeply repulsive inner wall, a more slowly-varying neighborhood of
the minimum, and a flat outer region. Parameters appropriate to molecular hydrogen are
used, a = 1.9425 Å–1, xe = 0.7416 Å, De = 4.7446 eV. Projections are calculated of this
function on the scaling functions from the L = 10 Daubechies system,
φkJ x( ) = 2 J λ( )−1/2
φ x / 2J λ − k( ) , (24)
with spacing 2J λ = 0.4 Å (λ and J are not independent parameters here).
15
For the sake of simplicity (but not necessity), the uniformly-spaced quadrature
associated with the basis functions on each level is used, xq = k + q −1( ) 2 j λ . The same
quadrature weights are therefore used for each functionφkj , and r – 1 of the points for
neighboring values of k overlap. In Fig. 5 the absolute values of the differences ∆
between integrals for orders r and r + 1 are displayed; it is very clear that significant lack
of convergence occurs for low r when only the coarsest quadrature is used. The highly
inaccurate case r = 1 is the usual basis of the Fast Wavelet Transform. Given the
O ∆x r( )error, proceeding to finer scales unfortunately provides only slow linear
convergence with respect to spacing (or cubic in the special case that the sampling is
moved to the position of the first moment). Errors are exponentially reduced for higher r
in the range covered, r ≤ 10. (Even higher orders r are allowed but, as is generally the
case with Newton-Cotes quadratures, can eventually lead to divergence.) The asymptotic
rates of convergence (slopes on the right-hand side of the semi-log plot) all conform. The
absolute errors, however, differ systematically for different k. As is evident from the top
of Fig. 6, the source of the differences is the strongly varying behavior of the anharmonic
Morse potential over the region spanned by the different φkj 's.
Greater accuracy for level J can be obtained by evaluation of the scaling function
integrals on levels J – 1, J – 2, …, until the desired level of accuracy is achieved,
followed by use of Eqs. (3) and (4) to recur back to level J . The best procedure is clearly
to use a value of r large enough to obtain satisfactory convergence rates and thus to
avoid needlessly high numbers of recursions, but coupled with selectivity over which
regions are subjected to refinement. Convergence may be measured either by (i)
comparison of the numerical results using quadrature at level j with those of quadrature at
16
level j – 1 followed by recursion back to j, or (ii) direct quadrature at level j using orders
r and r + 1 as done for the highest level in Fig. 5. The latter (simpler) choice was found
to be adequate in the calculations here, though this need not always be the case.
In the bottom section of Fig. 6 are shown the results for a demanding coarsest-
level convergence tolerance of 10–8 eV. For each level of refinement, the tolerance is
correspondingly reduced by a factor of 2. The markers in each row correspond to the
left-hand sides of included scaling functions; these functions are shown explicitly for the
coarsest level j = 4. On the finest level required, j = 0, the adaptive procedure has
reduced the 199 potentially-required integrals to only 20. One may obtain greater
compression for a smaller choice of L, although this comes at the price of reduced
smoothness in the basis functions.
Efficiency in the calculations can be gained by utilizing the fact that only
approximately half of the sample points required on each finer level are new.
Furthermore, each φkj for which the integral is recalculated using refinement requires the
integrals for φ2kj −1 , φ2k +1
j −1 , …, φ2k +L − 1j −1 , but there is a multiplexing advantage sinceφk ±1
j
require L – 2 of the same integrals, etc. Finally, when performing calculations with a
basis consisting of functions from different levels of resolution, many of the intermediate
integrals calculated will be of use for basis function integrals at more than one scale j .
Thus, the intermediate results can be stored and re-used to speed execution time. If
storage becomes an issue (e.g., in multidimensional wavelet calculations), however, it
may be more desirable to allow for a certain amount of recalculation of integrals to trade
speed for storage savings.
17
V. Matrix Element Quadrature
The considerations above have only considered integrals linear in the vector basis.
Bilinear integrals are standard in quantum mechanics, where one typically evaluates
matrix elements of kinetic and potential energy operators between two basis functions.
While kinetic energy operators in a compact support basis may be calculated by
straightforward methods,23,52 the same is not true for matrix elements of a potential
which is some general function of the coordinate (or, in multiple dimensions,
coordinates). Nevertheless, the quadrature methods described above translate
straightforwardly to the case of such bilinear integrals.
As before, the problem is first simplified by virtue of the recursions in Eqs. (3)
and (4). For a particular finite multiresolution basis of wavelets and scaling functions, all
necessary matrix elements can be reduced to integrals between scaling functions on the
same scale. It is therefore sufficient to consider the j = 0 case,
φ ′ k 0 f φ k
0 = dx∫ φ x − ′ k ( ) f x( )φ x − k( ). (25)
This matrix is automatically banded. Since the scaling functions vanish for arguments
outside (0, L –1), functions of index k and k' spatially overlap only for ′ k − k ≤ L − 2 .
Due to symmetry, there are L –1 distinct cases, exemplified by the specific choices k' = 0,
k = 0, 1, 2, …, L–2.
A quadrature strategy is adopted which allows for distinct sets of quadrature
points xq,k and weights ωq,k for each of these L – 1 values of k,
dx∫ φ x( ) f x( )φ x − k( ) ≈ ωq ,k f xq, k( )q= 1
rk
∑ . (26)
18
The xq,k may be chosen to overlap for different values of k, reducing the number of
samples required. Insisting that the quadrature be exact for powers 0 through rk – 1 leads
to
ωq, k xq ,kp
q =1
rk
∑ = mp ,k ≡ dx φ x( )∫ xp φ x − k( ), 0 ≤ p ≤ rk − 1. (27)
While the moments are different from before, the structure of the equations is identical.
Therefore the weights are given by
ωq, k =mp,k
p!i =0
rk −1
∑ Lrk ,q ,kp( ) 0( ). (28)
The particular moments in Eq. (27) may be obtained in the manner used by Beylkin52 to
obtain matrix representations of operators. This method uses change of scales for both
scaling functions in Eq. (27) and change of integration variables to obtain linear
equations that the moments must satisfy,
mp,k = 2− pp
′ p
′ p = 0
p
∑ m ′ p ,2 k + ′ k ap− ′ p , ′ k ′ k =1− L
L−1
∑ . (29)
The ap,k are the easily–calculated discrete sums
ap, k = hn +k hn n p
n∑ , (30)
where the scaling function coefficients hn are zero unless 0 ≤ n ≤ L–1. Also, from Eq.
(27) the moments for k < 0 can be shown to be related to those for k > 0 by
mp,− k =p
′ p
′ p =0
p
∑ − k( )p − ′ p m ′ p ,k . (31)
Thus Eqs. (29) need only be solved for the moments with non-negative k (using standard
methods), and this is required only once.
19
The polynomial-based quadrature above can be adapted straightforwardly to
matrix elements between the dual sets of functions in a biorthogonal basis. The latter
type of integral can arise, for instance, in solution of differential equations using the
biorthogonal systems in the Wavelet Galerkin method.58-61 In quantum contexts,
biorthogonal matrices may have application to problems in which non-Hermitian
Hamiltonian-type operators arise, e.g., in the use of complex rotation methods for L2 basis
calculations of resonance eigenvalues.62,63 One loses the advantages of symmetry of the
matrices in these cases, but gains in flexibility.
Following the Morse potential example of the last section, an adaptive quadrature
may be implemented to evaluate in tandem the potential matrix elements for the L = 10
scaling function basis. The same uniformly-spaced quadrature is used, but this time the
starting point of the grid for each matrix element must be chosen for two scaling
functions which partially overlap. A reasonable choice is to start at the left-hand side of
their region of intersection, i.e., at 2j λ max(k', k ). The multiplexing advantage stemming
from overlapping grids still applies despite the fact that a matrix is evaluated in this case.
The number of samples required at the first stage is only linearly proportional to the
distance spanned by the scaling function basis at level j.
Convergence may again be monitored by comparing results for adjacent orders rk
and rk + 1. All matrix elements with the same starting samples, i.e., the same max(k', k ),
are regarded jointly; if any of them fail to converge, then the entire subset is earmarked
for more accurate evaluation by recursion of the relevant integrals from the next finer
level. This starts a recursive refinement effectively focusing on the local behavior of the
potential sample points. The procedure is continued to finer scales, tightening the
20
convergence threshold by a factor of 2 each time, until the desired convergence is
reached and the results may be backtracked to the original scale. The alternative
convergence criterion mentioned in the vector case applies to the matrix case as well: one
may monitor convergence by agreement of quadrature at level j and of quadrature at level
j – 1 with recursion to level j. This would require one extra scale as well as a greater
amount of computation at intermediate steps.
While the matrix integrals are thus capable of being calculated in close analogy to
the vector integrals, there is one chief difference. The recursion between scales must be
implemented for both the bra and ket scaling functions at the same time, thereby
increasing the computational effort. In this regard, it is in principle possible to streamline
things by combining the two separate matrix operations into a single compound
operation.
For the level–j basis, 0 ≤ k ≤ 4, the required matrix elements were sought with a
convergence threshold of 10–5 eV. Quadrature orders 6 and 7 were used for checking of
convergence for all matrix elements. Table 2 shows the rapid convergence as a function
of scale for all of the matrix elements needed to evaluate the top–level integrals. The
number of distinct function evaluations is kept minimal by the current procedure. The
required matrix elements φ ′ k j V φ k
j are indicated explicitly in Fig. 7, with the starting
points of the two scaling functions mapped onto separate axes. Thus one can see
immediately the spatial distribution of the matrix elements required and, in particular, the
adaptive increase in density for the region near the steep inner wall of the potential.
21
VI. Conclusion
Integration by polynomial-based quadrature for orthonormal (or bi-orthonormal)
compact support wavelet families has been investigated for the purpose of use in
chemical and physical applications. In the above, a quadrature formula has been
developed in terms of easily–calculated derivatives of Lagrange interpolation
polynomials for the chosen coordinate grid and particular moments which are
grid–independent. Easy application to irregular grids was demonstrated in the wavelet
decomposition of the Hartley absorption spectrum of ozone. Adaptive quadrature was
developed for integrals of products of reasonably smooth functions and one or two
scaling functions, and applied to integrals of the anharmonic Morse potential of the types
that will arise in numerical quantum mechanical calculations using wavelet bases.
Acknowledgments
This work was supported by the National Science Foundation and the Robert A.
Welch Foundation. The authors also wish to thank R. O. Wells, C. S. Burrus, J. Tian and
W. Sweldens for conversations.
22
Appendix A
Derivation of Eq. (9) for the quadrature weights proceeds directly from the
interpolation formula in Eq. (11). The equality sign holds for the latter if f x( ) is a power
of x less than r,
Lr ,q x( ) xqn
q=1
r
∑ = xn . (A1)
Differentiating this with respect to x p times and setting x equal to zero yields
Lr ,qp( ) 0( ) xq
n
q= 1
r
∑ = p!δp, n, (A2)
or, multiplying both sides by mp / p!,
mp
p!Lr ,q
p( ) 0( )
xq
n
q =1
r
∑ = mp δp ,n . (A3)
If this is summed over p up to r – 1, the Kronecker delta is unity for p = n and zero
otherwise. Thus,
mp
p!Lr ,q
p( ) 0( )p = 0
r −1
∑
xqn
q=1
r
∑ = mn . (A4)
Since this is true for all n < r, the quantity in brackets is the solution of Eq. (8).
Appendix B
Nonlinear relations exist between various moments of the scaling function in
orthogonal compact support wavelet systems. This is most easily seen by considering the
autocorrelation function (the fundamental function of the Deslauriers-Dubuq scheme30)
Φ y( ) = dx∫ φ x( )φ x + y( ) . (B1)
23
The moments of Φ y( ) may be determined after a little algebra in terms of the moments
mp of the scaling function defined in Eq. (8),64
Mn = dy∫ yn Φ y( ) = −1( )n − p n
p
mn − p mp
p =0
n
∑ (B2)
On the other hand, Beylkin52 and Tian and Wells35 have shown that Mn = 0 for n = 0, 1,
…, L–1. For odd n this yields no new information; for even n, however, one finds
m2 = m12 , L ≥ 4, (B3)
m4 = 4 m3 m1 − 3m14 , L ≥ 6, (B4)
m6 = 6m5 m1 +10m32 − 60m3 m1
3 + 45m16, L ≥ 8, (B5)
m8 = 8m7 m1 +56m5 m3 −168m5 m13 +2520m3 m1
5 − 840m32 m1
2 −1575m18,
L ≥ 10, (B6)
etc. The first of these relations was derived by Gopinath and Burrus33 for the Daubechies
scaling functions and later shown to be more general by Sweldens and Piessens.21 To the
knowledge of the authors, the higher-order relations do not appear to have been noted
before. They are valid for all of the compact support wavelet families since the latter all
share the same autocorrelation function.
24
Bibliography
1 I. Daubechies, Comm. Pure Appl. Math. 41, 909 (1988).
2 S. G. Mallat, IEEE Trans. Pattern Anal. Machine Intell. 11, 674 (1989).
3 I. Daubechies, Ten Lectures on Wavelets (SIAM Publications, Philadelphia,
1992).
4 Wavelets: Mathematics and Applications, Vol. , edited by J. J. Benedetto and M.
W. Frazier (CRC Press, Boca Raton, 1994).
5 G. Strang and T. Nguyen, Wavelets and Filter Banks (Wellesley-Cambridge
Press, Wellesley, 1996).
6 C. S. Burrus, R. A. Gopinath, and H. Guo, Introduction to Wavelets and Wavelet
Transforms (Prentice-Hall, 1996).
7 H. L. Resnikoff and R. O. Wells, Jr., Wavelet Analysis: The Scalable Structure of
Information (Springer Verlag, 1998).
8 T. Paul, in Wavelets: Time-Frequency Methods and Phase Space, edited by J. M.
Combes, A. Grossman, and P. Tchamitchian (Springer Verlag, Berlin, 1989), pp. 315.
9 D. Permann, Phys. Rev. Lett. 69, 2607 (1992).
10 Z. Li, A. Borrmann, and C. C. Martens, Chem. Phys. Lett. 214, 362 (1993).
11 K. Cho, T. A. Arias, J. D. Joannopolous, and P. K. Lam, Phys. Rev. Lett. 71, 1808
(1993).
12 P. Fischer and M. Defranceschi, Int. J. Quant. Chem. 45, 619 (1993).
13 D. Permann and I. Hamilton, J. Chem. Phys. 100, 379 (1994).
14 L. Gagnon and J. M. Lina, J. Phys. A 27, 8207 (1994).
25
15 J. P. Modisette, P. Nordlander, J. L. Kinsey, and B. R. Johnson, Chem. Phys. Lett.
250, 485 (1996).
16 A. Askar, A. E. Cetin, and H. Rabitz, J. Phys. Chem. 100, 19165 (1996).
17 S. Wei and M. Y. Chou, Phys. Rev. Lett. 76, 2650 (1996).
18 P. Fischer and M. Defranceschi, SIAM J. Numer. Anal. 35, 1 (1996).
19 S. Goedecker and O. V. Ivanov, Sol. State Comm. 105, 665 (1998).
20 D. J. Kouri, (to be published).
21 W. Sweldens and R. Piessens, SIAM J. Numer. Anal. 31, 1240 (1994).
22 W. Sweldens and R. Piessens, Numerische Mathematik 68, 377 (1994).
23 A. Latto, H. L. Resnikoff, and E. Tenenbaum, Compactly Supported Wavelets,
AD91078, Aware, Inc. (1991).
24 W. Dahmen and C. A. Micchelli, SIAM J. Numer. Anal. 30, 507 (1993).
25 O. V. Vasilyev and S. Paolucci, J. Comp. Phys. 125, 498 (1996).
26 A. Cohen, I. Daubechies, and J. C. Feauveau, Comm. Pure Appl. Math. 55, 485
(1992).
27 M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions with
Formulas, Graphs, and Mathematical Tables (U. S. Government Printing Office,
Washington, D.C., 1972).
28 A. Ralston and P. Rabinowitz, A First Course in Numerical Analysis, Second ed.
(McGraw-Hill, New York, 1978).
29 H. R. Schwarz and J. Waldvogel, Numerical Analysis: A Comprehensive
Introduction (Wiley, Chichester, 1989).
26
30 G. Deslauriers and S. Dubuc, Constr. Approx. 5, 49 (1989).
31 M. J. Shensa, IEEE Trans. Signal Processing 40, 2464 (1992).
32 H. Rutishauser, Lectures on Numerical Mathematics (Birkhaüser, Boston, 1990).
33 R. A. Gopinath and C. S. Burris, in Proceedings of the ISCAS-92 (San Diego,
1992).
34 G. Beylkin, R. Coifman, and V. Rokhlin, Comm. Pure and Appl. Math 44, 141
(1991).
35 J. Tian and R. O. Wells, Jr., Vanishing Moments and Wavelet Approximation,
Computer Mathematics Laboratory Report CML9501, Rice University (1995).
36 W. Sweldens, in Wavelet Applications in Signal and Image Processing III, Proc.
SPIE 2569, edited by A. F. Laine and M. Unser (1995), pp. 68.
37 W. Sweldens, SIAM J. Math. Anal. 29, 511 (1998).
38 B. R. Johnson and J. L. Kinsey, Phys. Rev. Lett. 62, 1607 (1989).
39 B. R. Johnson and J. L. Kinsey, J. Chem. Phys. 91, 7638 (1989).
40 B. R. Johnson, B.-Y. Chang, C.-W. Hsiao, L. Le, and J. L. Kinsey, J. Chem. Phys.
108, 7670 (1998).
41 C. Leforestier, F. LeQuéré, K. Yamashita, and K. Morokuma, J. Chem. Phys. 101,
3806 (1994).
42 N. Balakrishnan and G. D. Billing, J. Chem. Phys. 101, 2968 (1994).
43 C. Leforestier, (to be published).
44 L. D. Le, C.-W. Hsiao, E. S. Lotfi, C. Kittrell, B. R. Johnson, and J. L. Kinsey,
(in progress).
27
45 D. C. Rouben and G. S. Ezra, J. Chem. Phys. 103, 1375 (1995).
46 K. Hirai, E. J. Heller, and P. Gaspard, J. Chem. Phys. 103, 5970 (1995).
47 J. Brion, A. Chakir, D. Daumont, J. Malicet, and C. Parisse, Chem. Phys. Lett.
213, 610 (1993).
48 F. Remacle, R. D. Levine, and J. L. Kinsey, Chem. Phys. Lett. 205, 267 (1993).
49 F. Remacle and R. D. Levine, J. Chem. Phys. 99, 4908 (1993).
50 F. Remacle and R. D. Levine, J. Phys. Chem. 97, 12553 (1993).
51 S.-Y. Lee, Z. W. Feng, and R. C. K. Yeo, J. Raman Spec. 28, 411 (1997).
52 G. Beylkin, SIAM J. Numer. Anal. 6, 1716 (1992).
53 B. R. Johnson and J. L. Kinsey, J. Chem. Phys. 99, 7267 (1993).
54 B. R. Johnson and J. L. Kinsey, in Femtosecond Chemistry, edited by J. Manz and
L. Wöste (Verlag Chemie, New York, 1995), pp. 353.
55 M. Shapiro, Chem. Phys. Lett. 242, 548 (1995).
56 X.-P. Peng, L.-S. Tian, and Y.-N. Peng, IEEE Trans. Signal Processing 44, 129
(1996).
57 P. Abry and P. Flandrin, IEEE Sig. Proc. Lett. 1, 32 (1994).
58 R. Glowinski, W. M. Lawton, M. Ravachol, and E. Tenenbaum, Wavelet solutions
of linear and nonlinear elliptic, parabolic and hyperbolic problems in one space
dimension, AD890527, Aware, Inc. (1989).
59 J. Ko, A. J. Kurdila, R. O. Wells, and X. Zhou, Comm. Num. Methods Eng. 12,
281 (1996).
60 J. R. Williams and K. Amaratunga, J. Comp. Phys. 122, 30 (1995).
28
61 W. Dahmen, A. Kunoth, and K. Urban, Computing 56, 259 (1996).
62 W. P. Reinhardt, Int. J. Quant. Chem. 21, 133 (1982).
63 K. McCormick and P. J. Nordlander, (unpublished).
64 M. Unser, IEEE Trans. Signal Proc. 44, 519 (1996).
29
Tables
Table 1. Scaling coefficients and lowest moments for the L = 10 Daubechies scaling
functions.
k hk mk
0 0.160102397974192914 1.000000000000000001 0.603829269797189671 1.193908018023822952 0.724308528437772928 1.425416355501573153 0.138428145901320732 1.580259805599972984 –0.242294887066382032 1.451304050272140505 –0.322448695846383746e–1 0.8137105292496663066 0.775714938400457135e–1 –0.2295042676756414267 –0.624149021279827427e–2 –0.9549759730451576788 –0.125807519990819995e–1 –1.672306166280536229 0.333572528547377128e–2 9.13925141629235832
10 4.06900434121189231e211 5.86543569280676173e3
Table 2. Convergence of Morse potential matrix element calculations in adaptive
quadrature. For each scale, Max |∆|, the maximum discrepancy between r = 6 and r = 7
evaluations of the matrix elements, occurs for k' = k = 0. Only approximately half of the
samples needed at each scale refinement require new evaluations.
No. of No. of NewScale Max |∆| Tolerance Samples Samples
4 6.76e-2 1.00e-5 11 113 7.19e-3 5.00e-6 24 132 3.25e-4 2.50e-6 34 171 8.86e-6 1.25e-6 42 210 1.84e-7 6.25e-7 38 19
30
Figures
1.0
0.5
0.0
-0.5
76543210x
(x) (x)
Figure 1. Scaling function φ (x ) and wavelet ψ(x ) for Daubechies family with L = 8.
1200
1000
800
600
400
200
0
44x103 40383634Frequency (cm–1)
Figure 2. Hartley absorption band of ozone measured at 218 K by Brion, et al.,47 withsmoothed reconstruction obtained from expansion in scaling functions at scale 3.28 cm–1.
31
-50
0
50
44x103 40383634Frequency (cm–1)
j = 87
6
54
Figure 3. The significant detail components of the Hartley absorption cross-section atscales 3 . 2j obtained by expansion in wavelets ψ k
j and summation over all k contributionsfor each j. Vertical offsets are given to traces other than that for j = 6.
44x103 40383634Frequency (cm–1)
j = 87
6
54
Figure 4. The wavelet coefficients corresponding to the components of the Hartleyabsorption cross-section shown in Fig. 3. For each j and k, the position on the frequencyscale marks the left hand side of the corresponding wavelet. These positions are markedby dots for the less congested scales. Vertical offsets are given to coefficients other thanthose for j = 6.
32
-4
-2
0
2
1086420Quadrature order r
01234
k
Figure 5. Absolute values of the differences for order r and order r + 1 quadratureevaluation of projection integrals φ k
j V for the Morse potential. The L = 10
Daubechies scaling functions of spacing ∆x = 0.4 Å are used. The sample spacing here is
the same as the basis function spacing.
3210x (Å)
V(x)
φkj
jj – 1j – 2j – 3j – 4
Figure 6. Morse potential and L = 10 Daubechies scaling functions for which projectionintegrals are calculated. For each scale j, the + signs represent the left hand endpoints ofscaling functions included in the adaptive quadrature. The greatest detail is required inthe vicinity of the steep repulsive wall.
33
4
3
2
1
0
43210k 2j λ (Å)
Figure 7. Required integrals φ ′ k j V φ k
j from adaptive calculation of the Morse potential
matrix elements. Each symbol represents a matrix element involving the product φ ′ k j φk
j ;the positions along the left and bottom axes represent the initial points of the k' and kscaling functions, respectively. The circles correspond to the target integrals on level j =4, while the black dots correspond to the matrix elements required on finer levels.Symmetry reduces the number of independent integrals required by approximately afactor of 2.