The Generalized Quadrature Method of Moments

Rodney O. Fox, Frédérique Laurent, Alberto Passalacqua

To cite this version:

Rodney O. Fox, Frédérique Laurent, Alberto Passalacqua. The Generalized Quadrature Method of
Moments. 2022. �hal-03762976v2�

HAL Id: hal-03762976

https://hal.science/hal-03762976v2
Preprint submitted on 1 Sep 2022 (v2), last revised 6 Oct 2022 (v3)

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 The Generalized Quadrature Method of Moments

Rodney O. Foxa,∗, Frédérique Laurentb , Alberto Passalacquac

a Department of Chemical and Biological Engineering, Iowa State University, 618 Bissell Road, Ames, IA 50011-1098, USA
b LaboratoireEM2C & Fédération de Mathématiques de CentraleSupélec, CNRS, CentraleSupélec, Université Paris-Saclay, 3
rue Joliot-Curie, 91192 Gif-sur-Yvette, France
c Department of Mechanical Engineering, Iowa State University, 62529 Union Drive, Ames, IA 50011-2030, USA

Abstract
The quadrature method of moments (QMOM) for a one-dimensional (1-D) population balance equation
was introduced by R. McGraw (Aerosol Science and Technology, 27, 255-265, 1997) to close the moment
source terms. QMOM is defined based on the properties of the monic orthogonal polynomials Qi of degrees
i = 0, 1, . . . , n that are uniquely defined by the set of 2n moments up to order 2n − 1. The moment of order
2n is fixed to the boundary of moment space such that the distribution function is approximated by a sum
of n Dirac delta functions. Using the recursion coefficients of the orthogonal polynomials for i > n ≥ 1,
the generalized quadrature method of moments (GQMOM) extends the quadrature representation to a sum
of N > n terms using the same moments as QMOM. In doing so, the known moments are preserved and
higher-order moments correspond to a distribution function in the interior of moment space. Here, GQMOM
closures for distributions on R, R+ , and (0, 1) are defined and analyzed. Generally speaking, GQMOM
provides a more accurate moment closure than QMOM without increasing the number of moments and at
nearly the same computational cost.
Keywords: population balance equation, quadrature-based moment methods, moment closures

1. Introduction

The quadrature method of moments (QMOM) (McGraw, 1997) is arguably the most successful and
widely used closure for finding the lower-order moments of the number density function (NDF) found from a
one-dimensional (1-D) population balance equation (PBE). There are numerous publications in the scientific
literature demonstrating its accuracy for treating particle growth, aggregation and breakage processes. In
many applications, only a relatively small number of moments are needed to attain sufficient accuracy. This
is because realistic growth, aggregation and breakage kernels are relatively smooth functions of the particle
size, which is a necessary condition for the Gaussian-quadrature approximation of integrals with respect to
the NDF to be accurate (Grosch et al., 2007). For example, most aerosol dynamics problems can be treated
accurately with between three to five quadrature nodes using QMOM (Marchisio and Fox, 2013; McGraw,
1997), which corresponds to six to ten moments. In this work, we use the ‘standard’ moments written as
integer powers of the phase-space variable. It is also possible to use generalized moments (Grosch et al.,
2007; Lage, 2011); however, the limitations associated with the QMOM that we wish to address remain the
same (Grosch et al., 2007).
Despite its many successes, the QMOM has an obvious shortcoming, i.e., in order to increase the number
of quadrature nodes, one must solve for a larger number of moments. In comparison, closures based on
reconstructing the NDF from a fixed set of moments do not suffer from this problem. For example, entropy

∗ Correspondingauthor
Email addresses: [email protected] (Rodney O. Fox), [email protected] (Frédérique Laurent),
[email protected] (Alberto Passalacqua)

Preprint submitted to Journal of Aerosol Science September 1, 2022

maximization (EM) (Jaynes, 1957; Mead and Papanicolaou, 1984) provides a closed form for the NDF that
can be integrated to arbitrary accuracy when evaluating the moment governing equations found from the
PBE. In other words, if a closed form for the NDF (e.g., exponential, lognormal, etc.) is available, then
the connection between the number of moments and the number of quadrature nodes is severed, and the
overall accuracy of the closure is then controlled by the number of moments. Nonetheless, applying EM
with a large number of moments is computationally prohibitive compared to QMOM. Therefore, alternative
methods with the computational simplicity of QMOM, but with the accuracy of EM are needed. In any case,
the number of moments used in QMOM cannot be increased indefinitely. In practice, the moment-inversion
algorithm suffers from ill-conditioning (Gautschi, 2004; Grosch et al., 2007; Wheeler, 1974), so that practical
applications cannot use more than fourteen to twenty moments.
As we describe in section 2, QMOM corresponds to a particular reconstruction of the NDF for the given
set of moments. Compared to the NDF from EM, the QMOM NDF produces the smallest possible value
for the next (unknown) even-order moment. Theoretically, there exists an infinite number of NDF with the
same (known) moments used in EM and QMOM. Thus, the problem reduces to choosing one such NDF in
a computationally efficient manner. For example, with an efficient algorithm, we should be able to find the
Gaussian quadrature that corresponds to the chosen NDF with the same algorithm employed with QMOM.
In prior work (Yuan et al., 2012), we introduced the extended quadrature method of moments (EQMOM) to
achieve this goal. The basic idea behind EQMOM is to reconstruct the NDF using a sum of kernel density
functions (KDF) centered around different sizes, but with the same standard deviation σ. One additional
(even-order) moment is needed to determine σ using an iterative, 1-D, root-finding procedure. Compared
to EM, the dual-quadrature algorithm for EQMOM is very efficient, and the quadrature nodes are easily
found for the known KDF (Gautschi, 2004; Lage, 2011). EQMOM has been widely used in the literature,
and various extensions have been proposed, e.g., using different definitions for the KDF (Madadi-Kandjani
and Passalacqua, 2015; Pigou et al., 2018). Nonetheless, it is not always possible to find a value for σ that
yields the additional moment and, even when such a σ exists, the iterative process needed for EQMOM is
not as fast as QMOM.
In this work, we propose a generalized version of QMOM (GQMOM) that overcomes all of the above-
mentioned shortcomings of QMOM and EQMOM. The key technical point employed in GQMOM is the
three-term recurrence relation for the family of orthogonal polynomials corresponding to the reconstructed
NDF. In short, instead of attempting to specify the unknown moments of the NDF (which is extremely
difficult, if not impossible), we choose an NDF with an allowable set of recurrence coefficients. In section 2.2,
we discuss how the recurrence coefficients are related uniquely to the moments, and in section 3 we define
closures for the GQMOM recurrence coefficients for NDF defined on R, R+ , and (0, 1) based on well-known
NDF (e.g., Gaussian, gamma, lognormal, beta). However, these closures are not unique and the reader can
easily compose choices with other desired properties. Once the recurrence coefficients are fixed, the Gaussian
quadrature can be computed immediately for an arbitrarily large number of nodes (Gautschi, 2004; Wheeler,
1974). In this manner, we do not need the explicit functional form for the NDF, but rather we determine the
NDF implicitly through knowledge of its recurrence coefficients. In practice, an approximation of the NDF
is available through the weights and abscissae of the Gaussian quadrature by setting their number to be
sufficiently large. In summary, given its facile implementation in PBE solvers based on QMOM, we expect
that GQMOM will become the method of choice for solving moment systems derived from a 1-D PBE.
The remainder of this work is arranged as follows. In section 2 we provide a brief overview of QMOM
in the context of a 1-D, time-dependent PBE with emphasis on its relationship to orthogonal polynomials
and realizable moments. In section 3 we define GQMOM for NDF defined on R, R+ , and (0, 1), and show
some simple examples of the resulting quadrature. By a change of variables, any NDF defined on a semi-
infinite or finite interval can be mapped to R+ or (0, 1), respectively. Therefore, these three definitions of
GQMOM cover almost all situations arising in applications. In general, GQMOM is equivalent to QMOM
for applications where QMOM works well. Thus, in section 4, we evaluate the performance of GQMOM
for some difficult test cases for which QMOM does not yield satisfactory results. Generally speaking, in
the context of a 1-D PBE, GQMOM provides the same advantages as EQMOM and EM, but at a lower
computational cost. Examples comparing the different methods for some difficult cases are also provided in
section 4.
2
2. QMOM for the resolution of a 1-D PBE

Here we provide a brief overview of QMOM for a spatially homogeneous, 1-D PBE. Application to
spatially inhomogeneous cases is described in detail in Marchisio and Fox (2013).

2.1. Principle of the method

The key principle of QMOM is to close moment equations derived from a PBE by using a Gauss quadra-
ture for the source terms (McGraw, 1997). Let f (t, ξ) be the NDF for the phase-space variable ξ ∈ B ⊆ R
at time t ∈ R+ with initial condition f (t, ξ) = f0 (ξ). For clarity, in this work the set B will be either R, R+ ,
or the interval (0, 1). To illustrate QMOM, hereinafter we assume that f (t, ξ) is governed by a PBE with
nonlinear source terms (see examples in section 4 ):

∂t f = S(t, ξ). (1)

In general, S(t, ξ) will be a functional of f (t, ξ) and its moments (Marchisio and Fox, 2013). The kth moment
of the NDF is
Z
Mk (t) = ξ k f (t, ξ) dξ (2)
B

for k ∈ N. Let M2n−1 := (M0 , M1 , . . . , M2n−1 ) be the moment vector of length 2n with n ∈ N.
The ordinary differential equation (ODE) system for the moment vector found from the PBE then has
the form
Z
dMk
= ξ k S(t, ξ) dξ = S k (t) (3)
dt B

where S k cannot generally be expressed exactly in terms of the moments of f contained in M2n−1 . QMOM
provides a closure for S k given M2n−1 (McGraw, 1997) of the form
n
X
S k (t) = wi Sk (t, ξi ) (4)
i=1

with non-negative weights wi (t) and abscissae ξi (t). QMOM has the property that if S(t, ξ) = f (t, ξ), then
Sk (t, ξi ) = ξik and S k (t) = Mk (t) for k = 0, 1, . . . , 2n−1; meaning that the weights and abscissae corresponds
to a Gauss quadrature. Computing these parameters from the moments is a well-known problem solved
thanks to the theory of orthogonal polynomials (Gautschi, 2004).

2.2. Monic orthogonal polynomials

Let us recall classical results that can be found in the literature (Dette and Studden, 1997; Gautschi,
2004; Schmüdgen, 2017). For a moment vector MN , one can define the linear functional h·i on the space
R[X]N of the real polynomial function of degree smaller than N by

hX k i = Mk , for k ∈ {0, 1, . . . , N }. (5)

Then, this linear functional defines a scalar product (P, Q) 7→ hP Qi on R[X]n , with n = bN/2c1 , as soon as
MN is strictly realizable, i.e., is associated with a positive NDF f ∈ L2 (B) through eq. (2). In such cases,
we have
Z
∀P ∈ R[X]N hP iMN = P (ξ)f (ξ)dξ, (6)
R

1 brc is the largest integer less than or equal to the real number r

3
and we can define a sequence of monic orthogonal polynomials Qi for i = 0, . . . , n with Qi of degree i. This
sequence satisfied a three-term recurrence relation:

Qi+1 = (X − ai )Qi − bi Qi−1 for i = 0, 1, . . . , n − 1 (7)

with Q0 = 1 and Q−1 = 0, and where

hXQ2i i hQ2i i
ai = , b i = . (8)
hQ2i i hQ2i−1 i

Note that an−1 depends on moments up to M2n−1 and bn−1 up to M2n−2 .

Given the moment vector M2n−1 , the Chebyshev algorithm (Gautschi, 2004; Wheeler, 1974) finds recur-
rence coefficients ai and bi for i = 0, 1, . . . , n − 1. However, it is important to note that adding an even-order
moment to get M2n allows computation of bn , while adding an odd-order moment to get M2n+1 allows
computation of an , etc.. This relation is one-to-one, thus the inverse is also true (i.e., the moments can
be found from the recurrence coefficients using the reverse Chebyshev algorithm (Fox and Laurent, 2022)).
These fundamental properties of orthogonal polynomials are used to define GQMOM in section 3.

2.3. Strict realizability characterization

The strict realizability of MN is equivalent to the positiveness of certain Hankel determinants (Dette
and Studden, 1997; Gautschi, 2004; Schmüdgen, 2017). Thus, let us define these determinants for d = 0, 1
and any non-negative k such that 2k + d ≤ N :

Md M1+d ... Mk+d

M1+d M2+d ... Mk+1+d
H 2k+d = .. .. .. .. , (9)
. . . .
Mk+d Mk+1+d ... M2k+d

and
M1−d − M2−d M2−d − M3−d ... Mk − Mk+1
M2−d − M3−d M3−d − M4−d ... Mk+1 − Mk+2
H 2k+d = .. .. .. .. . (10)
. . . .
Mk − Mk+1 Mk+1 − Mk+2 ... M2k+d−1 − M2k+d

The characterization of the strict realizability of MN , depending on the choice of B, is as follows:

• If B = R: H 2k > 0 for k = 0, . . . , b N2 c.

• If B = R+ : H k > 0 for k = 0, . . . , N .
• If B = (0, 1): H k > 0 and H k > 0 for k = 0, . . . , N .
On the boundary of moment space (i.e., for weakly realizable moments), one (or more) of the Hankel
determinants is null.
An equivalent characterization can be done using the coefficients of the three-term recurrence relation
of the orthogonal polynomials:
• If B = R: bi > 0 for i = 1, . . . , b N2 c.
• If B = R+ : existence of ζi > 0 for i = 1, . . . , N such that

i = 1, . . . , N 2−1
  
a0 = ζ1 , ai = ζ2i + ζ2i+1
(11)
bi = ζ2i−1 ζ2i i = 1, . . . , N2


4
 • If B = (0, 1): existence of ζi > 0 for i = 1, . . . , N satisfying eq. (11) and existence of pi ∈ (0, 1) such
that

ζ1 = p1 , ζi = pi (1 − pi−1 ) i = 2, . . . , N. (12)

Moreover, in the last case, the pi are referred to as the canonical moments (Dette and Studden, 1997).
Hereinafter, we assume that the known moment vector M2n is strictly realizable. In any case, this
assumption can be verified by applying the Chebyshev algorithm, which computes the recurrence coefficients
ai and bi from the known moments.

2.4. Moment space and realizability

The moment vector M2n lies in a convex subset of a linear vector space of dimension 2n + 1 (Dette and
Studden, 1997; Schmüdgen, 2017). We refer to this convex subset as moment space. The strict realizability
conditions discussed above define the boundary of moment space. If M2n lies in the interior of moment
space, then it is strictly realizable (all Hankel matrices are positive). A point on the boundary of moment
space has at least one of the Hankel matrices equal to zero. Hereinafter, we will use this terminology when
describing the properties of various moment closures.

2.5. QMOM closure

The weights wi and abscissae ξi of the Gauss quadrature are found using the recurrence coefficients.
(More details on this step are providedPin section 3.) Then, the QMOM
Pn closure corresponds to replacing
n
f (ξ) dξ by the finite positive measure i=1 wi δξi (ξ). Thus M2n = i=1 wi ξi2n , which is computed by the
quadrature, corresponds to bn = 0. This is also the lower bound of all possible M2n such that M2n is
realizable. In contrast, for the truncated Hamburger moment problem (Hamburger, 1944) (i.e. for B = R)
from a strictly realizable moment vector M2n , the odd-order moment M2n+1 can take on any finite real
value for M2n+1 being realizable. On the other hand, for B = R+ and B = (0, 1) the odd-order moments
are also bounded by the realizability constraints.

3. GQMOM for moment vector M2n

The basic idea behind GQMOM is to retain the properties of QMOM while using N > n ≥ 1 quadrature
points found from M2n with the Chebyshev algorithm. The two principal desired properties are that the
source-term closure has the form
N
X
S k (t) = wi Sk (t, ξi ), (13)
i=1

and the moment constraints for k = 0, 1, . . . , 2n are

N
X
Mk = wi ξik ; (14)
i=1

i.e., the same as with QMOM but with n replaced by N ≥ n ≥ 1. Roughly speaking, this can be accomplished
by choosing realizable moments Mk for k = 2n+1, 2n+2, . . . , 2N −1, and applying the Chebyshev algorithm
to M2N −1 . This is equivalent to selecting recurrence coefficients ai for i = n, n + 1, . . . , N − 1; and bi > 0
for i = n + 1, . . . , N − 1 that satisfy the realizability constraints in section 2.3. It is important to recall that
if the recurrence coefficients are realizable, then we are guaranteed that the quadrature nodes lie in B, and
their weights are positive. Expressed in terms of the monic orthogonal polynomials Qi (x), this selection is
done as follows.

5
3.1. Definition of GQMOM using recurrence coefficients
With GQMOM, the abscissae correspond to the roots of the monic orthogonal polynomial QN with
N > n ≥ 1. Thus, we must define the unknown recurrence coefficients (ai , bi ) for i = n, n + 1, . . . , N . First,
we start by defining the monic orthogonal polynomials.

3.1.1. Monic orthogonal polynomials

For i ≥ n ≥ 1, we extend the monic orthogonal polynomials in eq. (7) using
[M2n ] [M2n ]
Qi+1 = (x − ai )Qi − bi Qi−1 (15)

where the unknown recurrence coefficients depend on M2n and are formally defined by eq. (8). For example,
[M ] [M ]
bn 2n = bn is known, while an 2n depends on the closure for M 2n+1 (or vice versa). Hereinafter, an over-
[M ]
line is used to denote a closure for a higher-order moment given M2n . In general, the choice for bi 2n with
[M2n ]
i > n closes the moment M 2i . As long as bi > 0, this closure will be realizable for B = R (Hamburger,
1944), and similar constraints apply for B = R+ and B = (0, 1).
The unknown weight function is µ[M2n ] > 0, i.e., a distribution function on B with known moments
Z
Mk = ξ k dµ[M2n ] for k = 0, 1, . . . , 2n; (16)
B

and higher-order moments that depend on M2n :

Z
M 2n+k = ξ 2n+k dµ[M2n ] for k ≥ 1. (17)
B

The polynomials Qi are orthogonal:

Z
[M ]
Qi Qj dµ[M2n ] dξ = hQi Qj i = γi 2n δi,j (18)
B

where
[M2n ] H2i
γi = ≥0 (19)
H2i−2
[M ]
depends on moments up to M2i through the Hankel determinant H2i = H 2i . If, for some k, γk 2n = 0,
then bm = 0 for m ≥ k. This occurs when the Hankel determinant H2k = 0 and the moments M2n are on
[M ]
the boundary of moment space. For clarity, unless stated otherwise, we assume that γk 2n > 0, and thus
that M2n is in the interior of moment space for B = R. More generally, the GQMOM closure will depend
on the definition of the NDF, in particular on B.

3.1.2. Recurrence coefficients for B = R

[M ] [M ]
In general, choosing the definitions of ai 2n and bi 2n for i > n ≥ 1 corresponds to selecting a form
[M2n ]
for µ (among the infinite number of possible choices). Only when bn = 0 is the choice of µ[M2n ] unique
[M ]
because then M2n resides on the boundary of moment space. When this is the case, bi 2n = 0 for i > n. For
[M ]
this reason, expressing bi 2n as proportional to bn ensures the correct limiting behavior at the boundary of
moment space. Moreover, it may be desirable to capture a specific NDF, such as the Gaussian distribution.
The recurrence coefficients for the Hermite polynomials, corresponding to a Gaussian distribution, with
mean µ and variance σ 2 are

aH
i = µ, bH 2
i = iσ . (20)

These observations lead to the following definition.

6
 100

10-2

10-4

10-6

10-8

10-10
-8 -6 -4 -2 0 2 4 6 8

Figure 1: GQMOM quadrature weights versus abscissae for standardized Gaussian moments with n = 5, N = 101, and different
values of ν.

Definition 1 (GQMOM with moments M2n for n ≥ 1 from a NDF defined on R). The unknown re-
currence coefficients are selected as follows. For i = n in eq. (15), let
n−1
1X
a[M
n
2n ]
= ai , bn[M2n ] = bn . (21)
n i=0

For i > n, let

i−1 n−1  ν
[M ] 1X 1X [M2n ] [M2n ] i
ai 2n = ak = ak = an , bi = bn (22)
i n n
k=0 k=0

where the parameter ν ∈ R controls the tails of the distribution function.

[M ]
The choice for ai 2n is consistent because it uses the arithmetic average of the known lower-order
[M ]
coefficients. The choice for bi 2n is based on recovering Gaussian statistics when the known moments M2n
are Gaussian and ν = 1 (i.e., Hermite polynomials). In fig. 1, the weights are shown for different values of ν.
Setting ν = 2 gives exponential tails, while ν = 0 gives tails that decay very quickly in a bounded interval.2
Negative ν forces bi towards zero (i.e., the boundary of moment space), and eventually multiple modes are
observed. The NDF from EM for the first three moments (i.e., n = 1) corresponds to ν = 1. Thus, we
will refer to the choice ν = 1 as Gaussian–GQMOM. Otherwise, Definition 1 provides a family of GQMOM
closures parameterized by ν for moment vectors defined on R.

3.1.3. Recurrence coefficients for B = R+

[M ] [M ]
Due to the additional constraint of H 2n+1 > 0, different choices for ai 2n and bi 2n are needed for
strictly positive variables. In this case, it is convenient to introduce the sequence of positive ζi with a0 = ζ1
and

ai = ζ2i + ζ2i+1 , bi = ζ2i−1 ζ2i . (23)

[M ] [M ] [M ]
Thus, selecting ζi 2n for i > 2n is equivalent to choosing ai 2n and bi 2n . The Chebyshev algorithm
computes ai and bi from the moments M2n , and then eq. (23) is inverted to find ζi for 1 ≤ i ≤ 2n. For

q
2ν = 0 corresponds to Chebychev polynomials of the second kind for the distribution 1 − (ξ/ξmax )2 on the finite interval
√
(−ξmax , ξmax ) with ξmax = 2 b1 , which are the special case of the shifted Jacobi polynomials with α = β = 1/2.

7
 100

10-5

10-10

10-15
0 10 20 30 40 50 60 70

Figure 2: Gamma–GQMOM quadrature weights versus abscissae for gamma moments with α = 10, N = 101, and different
values of n.

the GQMOM closure, we can select the additional ζi to produce the generalized Laguerre polynomials for
moments from a gamma NDF where f (x) ∝ xα e−βx and β > 0. The ζi in this case are, for i ≥ 1:

L i+α L i
ζ2i−1 = , ζ2i = . (24)
β β
Note that β is a scaling parameter, while α changes the shape of the NDF and is dimensionless. These
two parameters are fixed given the mean and variance of the NDF. These properties lead to the following
definition.

Definition 2 (Gamma–GQMOM with moments M2n for n ≥ 1 from a NDF defined on R+ ). Let

M12
α= − 1.
M2 M0 − M12

Then α > −1 and the unknown recurrence coefficients are selected as follows. For i > n ≥ 1, let

[M2n ] i+α [M ] i
ζ2i−1 = ζ2n−1 , ζ2i 2n = ζ2n . (25)
n+α n
Thus, for i > n ≥ 1, in eq. (15) the recurrence coefficients are
[M2n ] [M2n ] [M 2n] [M2n ] [M
2n ] [M2n ]
ai = ζ2i + ζ2i+1 , bi = ζ2i−1 ζ2i . (26)

In fig. 2, the weights and abscissae are shown for gamma–PDF moments with α = 10 and β = 1 and three
different values of n. In all cases, the abscissas are positive and do not depend on n. The NDF from EM
with the same zero- and first-order moments corresponds to α = 0 (i.e., an exponential NDF).
Another possibility is to select the additional ζi to produce the Stieltjes–Wigert polynomials for moments
from a lognormal NDF:

(ln(ξ) − µ)2
 
1
f (ξ) ∝ √ exp − . (27)
ξσ 2π 2σ 2

The ζi in this case are, for i ≥ 1, denoting η = exp(σ 2 /2) > 1 (Madadi-Kandjani and Passalacqua, 2015;
Wilck, 2001):
W
ζ2i−1 = eµ η 4i−3 , W
ζ2i = eµ η 2i−1 (η 2i − 1). (28)

These properties lead to the following definition.

8
 100

10-2

10-4

10-6

10-8

10-10

10-12

10-14

10-16

10-18

10-20
100

Figure 3: Lognormal–GQMOM quadrature weights versus abscissae for lognormal moments with η = 1.01, N = 201, and
different values of n.

Definition 3 (Lognormal–GQMOM with moments M2n for n ≥ 1 from a NDF defined on R+ ).

Let
s
M 2 M0
η= . (29)
M12

With η > 1, the recurrence coefficients for i > n ≥ 1 are found from eq. (26) using
 2i 
[M2n ] 4(i−n) [M2n ] 2(i−n) η −1
ζ2i−1 = η ζ2n−1 , ζ2i =η ζ2n . (30)
η 2n − 1
In fig. 3, the weights and abscissae are shown for three different values of n with lognormal moments and
η = 1.01. Again, the abscissae are positive and do not depend on n, but many have weights below 10−15
due to the long tail of the lognormal NDF.

3.1.4. Recurrence coefficients for B = (0, 1)

For the finite interval (0, 1), we define GQMOM using the canonical moments pi ∈ (0, 1), which are
related to the ζi by

ζi = pi (1 − pi−1 ) (31)

for 1 ≤ i ≤ 2n with p0 = 0. For this purpose, the Chebyshev algorithm with input M2n can be easily
modified to compute the pi from the ζi found from eq. (23). Here, the choice for pi produces the Jacobi
polynomials for moments from a beta NDF where f (x) ∝ xβ (1 − x)α , but also that the quadrature nodes
are always in the interval (0, 1), which is the condition for the strict realizability. The canonical moments
in this case are, for i ≥ 1:
β+i i
pJ2i−1 = , pJ2i = . (32)
2i + α + β 2i + 1 + α + β
The two parameters α and β are fixed given the mean and variance of the NDF. These properties lead to
the following definition.

Definition 4 (Beta–GQMOM with moments M2n for n ≥ 1 from a NDF defined on (0, 1)). The
unknown recurrence coefficients are selected as follows. Let
1 − p1 − 2p2 + p1 p2 p1 − p2 − p1 p2
α= , β= (33)
p2 p2
9
 0.5 0.04

0.45
0.035

0.4
0.03
0.35

0.025
0.3

0.25 0.02

0.2
0.015

0.15
0.01
0.1

0.005
0.05

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 4: Beta–GQMOM quadrature weights versus abscissae for beta-NDF moments in (0, 1) with N = 51 and different values
of n.

and, for i > n ≥ 1,

 J
p2n−1 pJ2i−1 if p2n−1 ≤ pJ2n−1 or pJ2n−1 ≥ pJ2i−1 ,
[M2n ] p
p2i−1 = p2n−1 (1−p2n−1
J J J (34)
2i−1 )+p2i−1 −p2n−1

1−pJ
otherwise.
2n−1
 J
p2n pJ2i if p2n ≤ pJ2n or pJ2n ≥ pJ2i ,
[M2n ] p2n
p2i = p2n (1−pJ )+pJ −pJ (35)
 2i
1−pJ
2i 2n
otherwise.
2n

[M2n ] [M2n ] [M ] [M2n ] [M2n ]

These provide ζi = pi (1 − pi−12n ) for i > 2n, and hence, with eq. (26), ai and bi for i > n.

In fig. 4, the weights and abscissae are shown for moments from a beta NDF for different values of α, β > −1.
As can be observed, the abscissas remain bounded in (0, 1) and the quadrature points are the same for all
n. By a linear change of variables, beta–GQMOM on (0, 1) can be shifted to define a Gauss quadrature for
any finite interval (a, b).

3.2. Generalized QMOM for M2n−1

When working with a legacy code based on QMOM, it may be convenient to use M2n−1 instead of
M2n for n ≥ 1. Because M2n−1 contains moments only up to order 2n − 1 (and not 2n), bn is unknown.
However, for example, we can still use the procedure in eq. (22) to define bn given bn−1 . Thus, a sequence of
orthogonal polynomials can also be defined for this case. These polynomials can then be used to close even-
order moments starting at order 2n. For example, for n = 2 with M3 = (1, 0, 1, S3 ) and Gaussian–GQMOM,
we find b2 = 2 and hence the standardized fourth-order moment is 3 + S32 . As before, the closed moments
will be in the interior of moment space as long as bn−1 > 0. In summary, GQMOM can be used with
either M2n−1 or M2n . However, for consistency with EQMOM (Yuan et al., 2012), unless stated otherwise,
hereinafter we will use M2n .

10
3.3. Computation of weights and abscissas with GQMOM
As done with QMOM (McGraw, 1997; Wheeler, 1974), the N weights and N abscissae are found from
the eigenvectors and eigenvalues, respectively, of the following Jacobi matrix:
 √ 
√a0 b1 √
 b1 a1 b2 
 
JN = 
 .. .. .. .

(36)
 . . . 
p p
 bN −2 p aN −2 bN −1 
bN −1 aN −1

Using the definitions in eq. (21) and eq. (22), it is straightforward to modify the Chebyshev algorithm in Fox
and Laurent (2022) to construct JN given the moments M2n , and then to find the weights and abscissae.
By construction, the GQMOM quadrature satisfies eq. (14). The source terms in the moment equations
eq. (3) are closed using eq. (13). Unlike with QMOM, with GQMOM the value of N can be very large
as long as n is not too large (n ≤ 10). Notwithstanding, increasing N with fixed n does not increase our
knowledge of the unknown NDF governed by the PBE. Thus, as with the secondary quadrature in EQMOM
(Yuan et al., 2012), increasing N is only needed to reduce the quadrature error on S k . In other words, for
fixed n, the value of N can be increased until S k satisfies a convergence criterion for all k.

4. Numerical examples and validation

In this section we provide numerical examples for problems that are difficult to treat using QMOM. First,
to illustrate the differences between QMOM and GQMOM, we show examples of NDF composed of three
modes. Recall that with n = 1, GQMOM corresponds to a known NDF (e.g., Gaussian) while for QMOM,
it corresponds to a Dirac delta function. By increasing n, it is possible to approximate a multi-modal NDF.
It is important to understand that QMOM with M2n−1 represents the NDF by the sum of n Dirac delta
functions, and, given the additional moment M2n , GQMOM provides a continuous NDF. Heuristically, for
fixed M2n−1 GQMOM spreads out the n Dirac delta functions found with QMOM as M2n increases from
its minimum value. Thus, with sufficiently large M2n , the GQMOM NDF will become mono-modal, losing
all resemblance with the QMOM NDF.
Generally speaking, if the time evolution of the moments of a NDF are adequately captured with QMOM,
then GQMOM will provide equivalent accuracy. Thus, in the numerical examples starting in section 4.2, we
focus on systems for which QMOM is known to have difficulties. While such systems are relatively rare in
real applications, we shall see that GQMOM offers a viable alternative to QMOM for such cases.

4.1. Trimodal Gaussian

Recalling that the objective of GQMOM is to provide a more accurate N -node quadrature for fixed n, we
compute the moments to arbitrary order from a trimodal Gaussian NDF composed of three Gaussian NDF
with different means and standard deviations. Using QMOM, such a distribution requires at least n = 3 to
capture the locations of the three modes, and n ≥ 6 to capture their variances. Here, we take the weights of
the three Gaussian NDF to be equal, their means to be (−3, 0, 4) with standard deviations, respectively, of
(1/10, 1/20, 1/30). The quadratures found with Gaussian–GQMOM (N = 101) and QMOM are shown in
fig. 5 for n = 3, 6, 9. The GQMOM result with n = 3 consists of two delta functions at approximately ξ = −3
and 4, and a Gaussian distribution centered at ξ = 0 with significant variance. As n increases, the weights
take values that more closely correspond to the true NDF. Consistent with EQMOM, we observe that the
value of n mainly determines the quality of the reconstruction, while N provides “extra” refinement. The
latter is needed to converge the source-term closure in eq. (13) for a given n.
In most cases, N = 3n provides adequate refinement, but larger N may be needed for highly nonlinear,
localized, source terms. This point is illustrated in fig. 6 using a trimodal NDF on (0, 1) with beta–GQMOM.
In this figure, the weights are scaled to sum to N and n = 6 is held constant. With N = 6, the mean and
variance of each mode is roughly captured. Increasing to N = 3n = 18 provides several more quadrature
11
 0.35 0.35

n
0.3 0.3 3
6
9

0.25 0.25

0.2 0.2

0.15 0.15

0.1 0.1

0.05 0.05

0 0
-6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6

Figure 5: Trimodal Gaussian NDF on R with n = 3, 6, 9. Left: Gaussian-GQMOM with N = 101. Right: QMOM. The
quadrature points are connected by lines to ease interpretation.

4.5

3.5

2.5

1.5

0.5

0
0 0.2 0.4 0.6 0.8 1

Figure 6: Trimodal NDF on (0, 1) with n = 6 and different N . Weights are scaled to sum to N .

points for each peak, while N = 60 covers the full domain with roughly equal-spaced points. In all cases,
because GQMOM reproduces the same moments as QMOM, using N > n does not produce a poorer quality
quadrature for fixed n. In practice, a NDF with up to five modes can be approximated with GQMOM to
reasonable accuracy.

4.2. Application to aggregation and breakup problems

We consider in this section an application of GQMOM to a PBE involving aggregation and breakup.
Cases 5 and 8 studied in prior work (Madadi-Kandjani and Passalacqua, 2015; Vanni, 2000) are examined
here, since they allow for a direct comparison of the results obtained with GQMOM to those provided by
EQMOM in Madadi-Kandjani and Passalacqua (2015), and to the exact solution found in Vanni (2000). For
all other cases in Madadi-Kandjani and Passalacqua (2015), QMOM yields an accurate solution without the
additional quadrature nodes provided by GQMOM. For consistency, we maintain the same case numbers
used in Madadi-Kandjani and Passalacqua (2015).
The 1-D NDF describing the particle size ξ is f (t, ξ), and its temporal evolution is regulated by the
following PBE:
∂f (t, ξ)
= B̄ a (t, ξ) − D̄a (t, ξ) + B̄ b (t, ξ) − D̄b (t, ξ). (37)
∂t
Following Marchisio and Fox (2013); Marchisio et al. (2003); Randolph and Larson (1988), the source terms
for aggregation and breakage are expressed as
ξ 2 ξ β((ξ 3 − ξ 03 )1/3 , ξ 0 )
Z
B̄ a (t, ξ) = f (t, (ξ 3 − ξ 03 )1/3 )f (t, ξ 0 ) dξ 0 , (38)
2 0 (ξ 3 − ξ 03 )2/3
12
 Table 1: Cases examined for the aggregation and breakup process using M5 . M6 is reported in Case 8 for EQMOM.

Case β (ξ, ξ 0 ) a
((ξ) b (ξ|ξ 0 ) Mk (t = 0)
0 ξ=1
5 1 1, table 2 Mk = 1, k = 0, . . . , 5
0.02 ξ > 1


 M0 = 1




 M1 = 1.13
M2 = 1.294
( 


0 2 2 02 0 ξ=1
8 (ξ + ξ ) |ξ − ξ | 2, table 2 M3 = 1.5
0.01ξ 6 ξ>1 
M4 = 1.760





M5 = 2.087




M6 = 2.513

Z ∞
D̄a (t, ξ) = f (t, ξ) β(ξ, ξ 0 )f (t, ξ 0 ) dξ 0 , (39)
0

Z ∞
B̄ b (t, ξ) = a(ξ 0 )b(ξ|ξ 0 )f (t, ξ 0 ) dξ 0 , (40)
ξ

D̄b (t, ξ) = a(ξ)f (t, ξ). (41)

The application of eq. (2) to both sides of eq. (37) leads to the evolution equation for the moment Mk of
the NDF:
dMk (t)
= B̄ka (t) − D̄ka (t) + B̄kb (t) − D̄kb (t). (42)
dt
For consistency with prior work, in this example we use the moment vector M5 . The source terms for the
moments are defined as follows:
1 ∞
Z Z ∞
B̄ka (t) = f (t, ξ 0 ) β(ξ, ξ 0 )(ξ 3 + ξ 03 )k/3 f (t, ξ) dξ dξ 0 ,
2 0 0
Z ∞ Z ∞
a
D̄k (t) = k
ξ f (t, ξ) β(ξ, ξ 0 )f (t, ξ 0 ) dξ 0 dξ,
0 0
Z ∞ Z ∞ (43)
b k 0 0 0 0
B̄k (t) = ξ a(ξ )b(ξ|ξ )f (t, ξ ) dξ dξ,
Z0 ∞ 0

D̄kb (t) = ξ k a(ξ)f (t, ξ) dξ.

0

The integrals in eq. (43) are approximated as shown in eq. (4), when using GQMOM, by replacing n with N
to use all the additional nodes. The EQMOM integration strategy for eq. (42) is summarized in Appendix
A, and further details can be found in Madadi-Kandjani and Passalacqua (2015); Yuan et al. (2012). Table 1
summarizes the closure models used in the two test cases considered here, while the daughter distribution
functions b (ξ|ξ 0 ) are listed in table 2.
The time evolution of the volume-average diameter d43 = M4 /M3 is reported in fig. 7, where results are
shown for gamma–GQMOM, as well as for the lognormal–EQMOM (Madadi-Kandjani and Passalacqua,
2015), and the exact solution (Vanni, 2000). We consider the same cases to investigate the convergence of
13
 Table 2: Daughter distribution functions and their moment transforms.

No. Mechanism b((ξ|ξ 0 ) b̄k (ξ)

ξ0
2 ξ = 21/3
1 Symmetric fragmentation 2(3−k)/3 ξ k
0 otherwise
( 2
6ξ
ξ 03 ξ ∈ (0, ξ 0 ) 6ξ k
2 Uniform k+3
0 otherwise

5 5
Lognormal EQMOM (Madadi and Passalacqua, 2015)
4.5
4.5
Rigorous solution (Vanni, 2000)
4
4 GQMOM
3.5
3.5 3

3 2.5
d43

d43

2
2.5
1.5
2
Lognormal EQMOM (Madadi and Passalacqua, 2015) 1
1.5 Rigorous solution (Vanni, 2000) 0.5
GQMOM
1 0
0 50 100 150 200 0 2 4 6 8 10
t (s) t (s)

(a) Case 5 (b) Case 8

Figure 7: Time evolution of d43 obtained with gamma–GQMOM, lognormal–EQMOM Madadi-Kandjani and Passalacqua
(2015), and the exact solution Vanni (2000).

4.5

3.5

3
d43 (μm)

2.5

2 N=4
N=5
1.5
N = 10
1 N = 50
N = 100
0.5
Validation data - Rigorous solution of Vanni (2000)
0
0 50 100 150 200
t (s)

Figure 8: Case 5: Time evolution of d43 with increasing N .

14
 3.5

2.5

d43 (μm)
2
N=4
1.5 N=5
N = 10
1 N = 50
N = 100
N = 150
0.5
QMOM (3 nodes)
Validation data - Rigorous solution of Vanni (2000)
0
0 2 4 6 8 10
t (s)

Figure 9: Case 8: Time evolution of d43 with n = 3 and increasing N .

gamma–GQMOM with the number of nodes, while maintaining the number of QMOM quadrature nodes
constant and equal to n = 3 (6 moments), and increasing the number N of gamma–GQMOM nodes. We
consider N = 4, 5, 10, 50, 100 for Case 5 and N = 4, 5, 10, 50, 100, 150 for Case 8.
From the time evolution of d43 reported in fig. 8, no variation can be visually observed in the solutions
obtained with gamma–GQMOM using different N . A more interesting behavior is observed in Case 8, for
which the time evolution of d43 is shown in fig. 9. In this case, QMOM with n = 3 is insufficient to capture
the asymptotic behavior. Adding another gamma–GQMOM node for the same moments leads to correctly
predict the trend but not the asymptotic value. However, the gamma–GQMOM prediction appears to
converge for N > 50. Specifically, for the cases run with N = 100 and 150 the percentage difference between
the asymptotic values of d43 is 0.0462%.

4.3. Example with a localized source term

In this section, we consider an application of GQMOM to a PBE involving nucleation, growth and
filtration, as can occur in suspension crystallization processes (Grosch et al., 2007) or aerosol filtration.
Here, the nucleation is assumed to create particles of fixed size ξ0 , at a constant rate ϕ, the growth rate
G is assumed to be constant, and the filtration term describes the removal of particles of size larger than
ξ1 . This PBE is used, for example, to model the effect of an ideal fine trap (Grosch et al., 2007). We then
consider the following PBE for the NDF f (t, ξ):

∂f (t, ξ)
= ϕδξ0 (ξ) − ∂ξ (Gf (t, ξ)) − αf (t, ξ)I[ξ1 ,+∞) (ξ). (44)
∂t
 
where the identity function I[a,+∞) (x) is zero for x < a and 1 elsewhere. Let us remark that f t, ξξ−ξ
1
0
−ξ 0
is
the solution of a similar PBE with scaled parameters, in such a way that we can choose ξ0 = 0 and ξ1 = 1
without loss of generality. Finally, the values of the parameters used here are ϕ = 1, G = 2 and α = 10, in
this non-dimensional context. The main difficulty of this PBE lies in the filtration term, which is localized
in phase space so that it only affects particles with size ξ ≥ ξ1 . Indeed, quadrature-based methods usually
give very good results for the growth term, even with a non-constant rate G(ξ), and the nucleation term is
already closed in the moment equations.
The analytical solution of eq. (44), with the zero initial condition f (0, ξ) = 0, is

0 if ξ < ξ0
ϕ
f (t, ξ) = 1 if ξ0 ≤ ξ ≤ min(ξ1 , ξ0 + tG) , (45)
G α


exp − G (ξ − ξ1 ) if ξ1 ≤ ξ ≤ ξ0 + tG


15
 0.7 0.5
0.45
0.6
0.4
0.5 0.35

0.4 0.3
M0

M1
0.25
0.3 0.2
0.2 exact 0.15 exact
QMOM-Radau 0.1 QMOM-Radau
0.1 GQMOM-Radau GQMOM-Radau
GQMOM 0.05 GQMOM
EQMOM EQMOM
0 0
0 1 2 3 4 5 0 1 2 3 4 5
time time
0.5 0.6
0.45
0.5
0.4
0.35
0.4
0.3
M2

M3
0.25 0.3
0.2
0.2
0.15 exact exact
0.1 QMOM-Radau QMOM-Radau
GQMOM-Radau 0.1 GQMOM-Radau
0.05 GQMOM GQMOM
EQMOM EQMOM
0 0
0 1 2 3 4 5 0 1 2 3 4 5
time time
0.7

0.6

0.5

0.4
M4

0.3

0.2 exact
QMOM-Radau
0.1 GQMOM-Radau
GQMOM
EQMOM
0
0 1 2 3 4 5
time

Figure 10: Test case with nucleation, growth and filtration: moments of order 0, 1, 2, 3 and 4 for the analytical solution,
QMOM–Radau with five moments (i.e., M4 ), gamma–GQMOM and gamma–GQMOM–Radau with five moments and N = 20
and gamma–EQMOM with 10 secondary quadrature points.

so that the moments of the NDF are

!
max(0,ξ0 −ξ1 +tG)
ϕ min(ξ0 + tG, ξ1 )k+1 − ξ0k+1
Z  α 
k
mk (t) = + (ξ1 + s) exp − s ds . (46)
G k+1 0 G

Using GQMOM with M2n , the equations for the moments are written, for k = 0, 1, . . . , 2n, as
Z +∞
dMk (t) k
= ϕξ0 + kGMk−1 (t) − α ξ k dµ[M2n ] . (47)
dt ξ1

For their resolution, an operator splitting is introduced. During one time step ∆t, the growth term is
first solved, followed by the resolution of the others terms. Each operator is solved analytically, using
the underlying representation of the moments as a sum of Dirac delta functions. Then, for the growth
operator, each abscissa of the quadrature is increased by G∆t. For the nucleation and filtration operators,
the weights corresponding to abscissa greater than ξ1 are multiplied by exp(−a∆t) and ϕξ0k ∆t is added to
the corresponding moment of order k, for k = 0 . . . , 2n. Moreover, a small enough time step (5×10−6 ) is
chosen to ensure time convergence.
The simulations are done using five moments M4 and N = 20 abscissas with gamma–GQMOM and
gamma–GQMOM–Radau. Here, Radau quadrature refers to a Gaussian quadrature where one abscissa is

16
 0.7 0.5
0.45
0.6
0.4
0.5 0.35

0.4 0.3
M0

M1
0.25
0.3 0.2
exact exact
0.2 QMOM 0.15 QMOM
GQMOM-Radau 0.1 GQMOM-Radau
0.1 GQMOM GQMOM
EQ alpha=2 0.05 EQ alpha=2
EQ alpha=4 EQ alpha=4
0 0
0 1 2 3 4 5 0 1 2 3 4 5
time time
0.45 0.5
0.4 0.45

0.35 0.4
0.35
0.3
0.3
0.25
M2

M3
0.25
0.2
0.2
0.15 exact exact
QMOM 0.15 QMOM
0.1 GQMOM-Radau 0.1 GQMOM-Radau
GQMOM GQMOM
0.05 EQ alpha=2 0.05 EQ alpha=2
EQ alpha=4 EQ alpha=4
0 0
0 1 2 3 4 5 0 1 2 3 4 5
time time

Figure 11: Test case with nucleation, growth and filtration: moments of order 0, 1, 2 and 3 for the analytical solution, QMOM
with four moments (i.e., M3 ), gamma–GQMOM and gamma–GQMOM–Radau with four moments and entropic quadrature
(EQ) with multiplicative factor α = 2 and α = 4.

fixed (Gautschi and Li, 1991). These results are compared with the analytical solution in fig. 10. As seen
from these plots, gamma–GQMOM–Radau seems to be well adapted to this test case: compared to gamma–
GQMOM, an abscissa is fixed at ξ = 0, which is the nuclei size. And then, gamma–GQMOM–Radau gives
better results than gamma–GQMOM, even if gamma–GQMOM also has good results for a larger number of
abscissas (not shown here). These methods are also compared to QMOM–Radau, whose results for M0 are
quite good, but the error increases with the order of the moments. Gamma–EQMOM is also tested with
ten secondary quadrature points, so that the cost is quite similar to gamma–GQMOM. However, gamma–
EQMOM gives the worst results, even if these results can be improved by using a much larger number of
secondary quadrature points (not shown here).
Finally, to compare GQMOM with another method from the literature, entropic quadrature (EQ) de-
veloped by Böhmer and Torrilhon (2020), simulations using four moments M3 are done. In fig. 11 the
results are plotted with gamma–GQMOM and gamma–GQMOM–Radau using N = 20 quadrature points,
QMOM, and EQ with multiplicative factor α = 2 and α = 4 and the smallest abscissa fixed at ξ = 0. The
best results are obtained with gamma–GQMOM–Radau even if the moments are slightly overestimated,
whereas the other methods show an oscillatory behavior. In any case, it is very interesting to observe how
using Radau quadrature in the presence of nucleation significantly improves the predictions of GQMOM,
even with a relatively small number of moments. Also, comparing fig. 11 with fig. 10, we see that using
gamma–GQMOM–Radau with M4 is significantly more accurate than with M3 . This is consistent with
observations made with other systems, i.e., moment vectors terminating with an even-order moment usually
provide more accurate results.

4.4. Symmetric binary breakup

We consider next the apparently very simple test case of symmetric binary breakup. Depending on the
initial conditions, it has been shown that this test case can lead to some difficulties for QMOM, such as
convergence to an incorrect solution (Peterson et al., 2022). Thus, let us consider the following PBE for the
NDF f (t, v):

∂t f (t, v) = −g(v)f (t, v) + 4g(2v)f (t, 2v) (48)

17
 2 0.9
n=2 n=2
1.8 n=3 n=3
n=4 0.85 n=4
1.6 n=5 n=5

sqrt(M3M1/M22-1)
n=7 n=7
1.4 n=10 0.8 n=10
n=15 n=15
M2/M1

1.2
0.75
1
0.8 0.7
0.6
0.65
0.4
0.2 0.6
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
time time

Figure 12: Fragmentation, Case 1: mean (left) and variance (right) obtained for simulations with QMOM with a number of
abscissas n equal to 2, 3, 4, 5, 7, 10, 15.

with the initial condition f (0, v) = f 0 (v). Moreover, the simple case g(v) = v suffices to illustrate the
difficulties. Using GQMOM with M2n−1 , the equations for the moments are written, for k = 0 . . . , 2n − 1,
as
dMk (t)
= (−1 + 21−k )Mk+1 (t). (49)
dt
Let us remark that, due to the simple choice for g, only M2n has to be closed. This set of equations is solved
using an adaptive time-step algorithm (Nguyen et al., 2016) based on embedded SSP explicit Runge–Kutta
methods, with a time-step selection designed both to control the error and to ensure the realizability of the
moment set.
For the initial distribution, two cases are considered:

• Case 1: f 0 (v) = exp(−v).

• Case 2: f 0 (v) is a lognormal distribution: with v0 = 1, σ0 = 2:
  !
0 1 ln(v) − µ v0
f (v) = √
exp − , µ = ln , σ 2 = ln(1 + σ02 ).
2σ 2
p
vσ 2π 1 + σ02
p
As in Peterson et al. (2022), we look at the mean M2 /M1 and variance M3 M1 /M22 − 1 of the volume
distribution. For Case 1, we can see in fig. 12 that using QMOM, the results converge when the number of
moments increases to the value given in Peterson et al. (2022). With gamma–GQMOM, only one additional
quadrature point is used compared to QMOM. Due to the simple form of the equations, all simulations with
a higher number of quadrature points give the same results. It can be seen in fig. 13 that gamma–GQMOM
converges to the same mean value and variance as QMOM, when the number of moments increases, but also
faster than QMOM: about half the number of moments are needed for the same accuracy.
Case 2 was shown in Peterson et al. (2022) to be difficult for QMOM, i.e., the method converges to
the wrong solution. In fig. 14, we compared the mean value and variance obtained with QMOM, gamma–
GQMOM and lognormal–GQMOM, using a large enough number of moments (16) so that the solution is
converged. The reference solution is taken from Peterson et al. (2022). Gamma–GQMOM and QMOM
converge to the same solution, whereas lognormal–GQMOM converges to another one, closer to, but still
different, from the reference solution. For Cases 1 and 2, the number N of quadrature points for GQMOM
has no influence on the solution as soon as it is greater that n, which is the number of quadrature points for
QMOM. The convergence issue with Case 2 is due to the very long tails of the initial lognormal distribution.
In this sense, Case 2 is an extreme case that is unlikely to occur in physical systems. Nonetheless, using
lognormal–GQMOM improved the results, and it would also possible to define a GQMOM closure with even
longer tails to handle such cases. In real-world applications, GQMOM provides more flexibility than QMOM
to accommodate for the shape of the NDF.

18
 2 0.71
QMOM, n=15 QMOM, n=15
1.8 n=2, N=3 n=2, N=3
n=3, N=4 0.7 n=3, N=4
1.6 n=4, N=5 n=4, N=5

sqrt(M3M1/M22-1)
n=5, N=6 0.69 n=5, N=6
1.4
0.68
M2/M1

1.2
1 0.67
0.8
0.66
0.6
0.65
0.4
0.2 0.64
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
time time

Figure 13: Fragmentation, Case 1: reference solution using QMOM and mean (left) and variance (right) obtained for simulations
with gamma–GQMOM with a number of moments 2n equal to 4, 6, 8, 10; and a number of quadrature points equal to n + 1.

1 2
reference reference
0.9 QMOM 1.8 QMOM
Gamma GQMOM Gamma GQMOM
0.8 Lognormal GQMOM 1.6 Lognormal GQMOM

sqrt(M3M1/M22-1)
0.7 1.4
M2/M1

0.6 1.2
0.5 1
0.4 0.8
0.3 0.6
0.2 0.4
0.1 0.2
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
time time

Figure 14: Fragmentation, Case 2: reference solution from Peterson et al. (2022) and converged mean (left) and variance (right)
obtained for simulations with QMOM, gamma–GQMOM, and lognormal–GQMOM.

5. Discussion and conclusions

In this work, we have generalized the widely used moment closure QMOM for solving moment systems
derived from a 1-D PBE to allow for an arbitrarily large number N of Gauss quadrature nodes. Like QMOM
does for M2n−1 , GQMOM exactly reproduces the input moment vector M2n where 2n is the order of the
highest-order known moment. Unlike QMOM, GQMOM closes the higher-order moments such that the
moment vector M2N is (almost always) in the interior of moment space, i.e., the moments correspond to a
continuous NDF unless M2n is on the boundary of moment space. When the latter occurs, the GQMOM
quadrature is the same as the one found with QMOM. This is because, by construction, unless bn = 0,
GQMOM will never produce a set of recurrence coefficients on the boundary of moment space. As we
have shown through examples, the principal advantage of using GQMOM versus QMOM is the ability to
increase the number of quadrature points from n to N without solving for more moments. This allows for
a more accurate evaluation of the moment source terms at nearly the same computation cost. Nonetheless,
as shown in the examples, the overall accuracy of the moment method is controlled principally by n (i.e.,
by the number of solved moments), and, hence, a much larger N is only needed for “difficult” source terms.
In comparison to other moment closures based on a continuous NDF (e.g., EQMOM, EM, EQ), GQ-
MOM is much easier to implement because it only requires a straightforward modification of the Chebyshev
algorithm to compute the additional recurrence coefficients (i.e., ai , bi for i = n, n + 1, . . . , N ). Furthermore,
because this process results in a Gauss quadrature on the support B, we are guaranteed that the abscissae,
being the roots of a monic orthogonal polynomial, lie inside B and are distinct. This is not the case, for
example, with EQMOM where the abscissae from the secondary quadrature can overlap, which causes severe
problems, for example, when one attempts to find conditional moments (Cheng et al., 2010; Yuan and Fox,
2011). Moreover, the weights of the Gauss quadrature from GQMOM will be positive as long as M2n lies
in the interior of moment space; otherwise, some weights will be null.
In general, moment closures that reside in the interior of moment space are preferable because they
can tolerate small numerical errors without becoming non-realizable. For example, for a PBE that includes
19
spatial transport with a known velocity (e.g., the gas velocity for an aerosol), constructing realizable finite-
volume schemes is challenging (Wright, 2007), even if some realizable second-order schemes were developed
(Laurent and Nguyen, 2017; Marchisio and Fox, 2013; Passalacqua et al., 2020; Shiea et al., 2020; Vikas et al.,
2011). In this respect, GQMOM may allow for use of higher-order spatial reconstruction of the moment
vector for cases where QMOM is limited to low order. The same issue is faced when solving the moment
source terms numerically (Nguyen et al., 2016), so that, generally speaking, we expect that GQMOM will
generate more robust, and more accurate, numerical solvers for the multi-scale, multi-physics codes used in
real-world applications (Bryngelson et al., 2020; Heylmun et al., 2021, 2019; Ilgun et al., 2021; Passalacqua
et al., 2018; Wick et al., 2017).
Many real-world applications are described by a NDF with more than one internal variable (e.g., for
aerosols, the droplet volume, temperature, composition, etc.). It is well known that the numerical solution
of a multi-dimensional PBE, even for spatially homogeneous cases, is very challenging due to the high
dimension of the phase space. One possible simplification is to condition the internal variables on the variable
with the largest variance (Marchisio and Fox, 2013; Yuan and Fox, 2011). In applications involving droplets
(aerosols), solid particles or bubbles, the mass (or size) of the particle is very often the most important
internal variable. As done in the examples in section 4, GQMOM can be applied with the moments M2n
of the size variable ξ. Then, the conditional mean of another variable φ given ξ is denoted by hφ|ξi, and
is computed using the joint moments hφξ k i for k = 0, 1, . . . , n − 1 by applying the conditional quadrature
method of moments (CQMOM) (Cheng et al., 2010; Yuan and Fox, 2011). When QMOM is replaced by
GQMOM, it will be necessary to approximate hφ|ξi using the same set of joint moments as CQMOM.
However, the straightforward application of CQMOM with N > n abscissae yields an under-determined
linear system. Thus, it is necessary to modify CQMOM to handle N > n abscissae. This can be done
successfully using an interpolation function, and we will describe the resulting algorithm for generalized
CQMOM in a future publication.

Declaration of competing interest

The authors declare that they have no known competing interests or personal relationships that could
have appeared to influence the work reported in this paper.

Appendix A. Calculation of integrals with EQMOM

Differently from QMOM and GQMOM, in EQMOM the NDF is approximated as a weighted sum of
non-negative KDF (Yuan et al., 2012):
n
X
f (ξ) ≈ pn (ξ) = wα δσ (ξ, ξα ) (A.1)
α=1

where wα are the weights of each KDF δσ (ξ, ξα ) with standard deviation σ, ξα are the corresponding
abscissae, and n is the number of KDF being used. Based on this approximation, integrals involving
products of the NDF and a function of ξ are approximated as (Madadi-Kandjani and Passalacqua, 2015)
Z ∞ Z ∞ n
X X Nα
n X
g(ξ)pn (ξ) dξ = g(ξ) wα δσ (ξ, ξα ) dξ = wα wαβ g(ξαβ ). (A.2)
0 0 α=1 α=1 β=1

where, for fixed α, wαβ and ξαβ are the Nα Gaussian quadrature nodes corresponding to the KDF (Yuan
et al., 2012). In principle, Nα can be different for each α, but here we let Nα = 2n.

20
 Consequently, the evolution equation of the moment Mk in aggregation and breakup processes discussed
in section 4.2 is rewritten as
n Nα n Nα
dMk 1 X X X X
= wα1 wα1 β1 wα2 wα2 β2 (ξα3 1 β1 + ξα3 2 β2 )k/3 βα1 β1 α2 β2
dt 2 α =1 α =1
1 β1 =1 2 β2 =1
n
X Nα
X n
X Nα
X
− wα ξαk 1 β1 wα1 wα1 β1 wα2 wα2 β2 βα1 β1 α2 β2
α1 =1 β1 =1 α2 =1 β2 =1

X Nα
n X Nα
n X
X
+ wα aαβ b̄kαβ wαβ − k
wα ξαβ aαβ wαβ (A.3)
α=1 β=1 α=1 β=1

where b̄kαβ = b̄k (ξαβ ). In the numerical implementation of EQMOM, σ is found from M2n using an iterative
procedure (Pigou et al., 2018; Yuan et al., 2012). Thus, since N ≈ nNα , the principal advantage of GQMOM
versus EQMOM is to eliminate the need for iterations to find σ before evaluating the right-hand side of
eq. (A.3).

References
Böhmer, N., Torrilhon, M., 2020. Entropic quadrature for moment approximations of the Boltzmann–BGK equation. Journal
of Computational Physics 401, 108992. URL: http://www.sciencedirect.com/science/article/pii/S0021999119306977,
doi:https://doi.org/10.1016/j.jcp.2019.108992.
Bryngelson, S.H., Colonius, T., Fox, R.O., 2020. QBMMLib: A library of quadrature-based moment methods. SoftwareX 12,
100615.
Cheng, J.C., Vigil, R.D., Fox, R.O., 2010. A competitive aggregation model for flash nanoprecipitation. J. Colloid Interface
Sci. 351, 330–342.
Dette, H., Studden, W.J., 1997. The Theory of Canonical Moments with Applications in Statistics, Probability, and Analysis.
Wiley Series in Probability and Statistics: Applied Probability and Statistics, John Wiley & Sons Inc., New York.
Fox, R.O., Laurent, F., 2022. Hyperbolic quadrature method of moments for the one-dimensional kinetic equation. SIAM J.
Appl. Math. 82, 750–771.
Gautschi, W., 2004. Orthogonal Polynomials: Computation and Approximation. Oxford University Press, Oxford, UK.
Gautschi, W., Li, S., 1991. Gauss–Radau and Gauss–Labatto quadratures with double end points. Journal of Computational
and Applied Mathematics 34, 343–360.
Grosch, R., Briesen, H., Marquardt, W., Wulkow, M., 2007. Generalization and numerical investigation of QMOM. AIChE
Journal 53, 207–227.
Hamburger, H.L., 1944. Hermitian transformations of deficiency-index (1,1), Jacobian matrices, and undetermined moment
problems. Amer. J. Math. 66, 489–552.
Heylmun, J.C., Fox, R.O., Passalacqua, A., 2021. A quadrature-based moment method for the evolution of the joint size–velocity
number density function of a particle population. Computer Physics Communications 267, 108072.
Heylmun, J.C., Kong, B., Passalacqua, A., Fox, R.O., 2019. A quadrature-based moment method for polydisperse bubbly
flows. Computer Physics Communications 244, 187–204.
Ilgun, A.D., Fox, R.O., Passalacqua, A., 2021. Solution of the first-order conditional moment closure for multiphase reacting
flows using quadrature-based moment methods. Chemical Engineering Journal 405.
Jaynes, E.T., 1957. Information theory and statistical mechanics. The Physical Review 106, 620–630.
Lage, P.L., 2011. On the representation of QMOM as a weighted-residual method – the dual-quadrature method of generalized
moments. Computers and Chemical Engineering 35, 2186–2203.
Laurent, F., Nguyen, T.T., 2017. Realizable second-order finite-volume schemes for the advection of moment sets of the particle
size distribution. Journal of Computational Physics 337, 309–338.
Madadi-Kandjani, E., Passalacqua, A., 2015. An extended quadrature-based moment method with log-normal kernel density
functions. Chemical Engineering Science 131, 323–339.
Marchisio, D.L., Fox, R.O., 2013. Computational Models for Polydisperse Particulate and Multiphase Systems. Cambridge
University Press, Cambridge, UK.
Marchisio, D.L., Pikturna, J.T., Fox, R.O., Vigil, R.D., 2003. Quadrature method of moments for population-balance equations.
AIChE Journal 49, 1266–1276.
McGraw, R., 1997. Description of aerosol dynamics by the quadrature method of moments. Aerosol Science and Technology
27, 255–265.
Mead, L.R., Papanicolaou, N., 1984. Maximum entropy in the problem of moments. J. Math. Phys. 25, 2404–2417.
Nguyen, T.T., Laurent, F., Fox, R.O., Massot, M., 2016. Solution of population balance equations in applications with fine
particles: mathematical modeling and numerical schemes. Journal of Computational Physics 325, 129–156.

21
Passalacqua, A., Laurent, F., Fox, R.O., 2020. A second-order realizable scheme for moment advection on unstructured grids.
Computer Physics Communications 248, 106993.
Passalacqua, A., Laurent, F., Madadi-Kandjani, E., Heylmun, J.C., Fox, R.O., 2018. An open-source quadrature-based
population balance solver for OpenFOAM. Chemical Engineering Science 176, 306–318.
Peterson, J.D., Bagkeris, I., Michael, V., 2022. A new framework for numerical modeling of population balance equations:
Solving for the inverse cumulative distribution function. Chemical Engineering Science 259, 117781.
Pigou, M., Morchain, J., Fede, P., Penet, M.I., Laronze, G., 2018. New developments of the extended quadrature method of
moments to solve population balance equations. Journal of Computational Physics 365, 243–268.
Randolph, A.D., Larson, M.A., 1988. Theory of Particulate Processes: Analysis and Techniques of Continuous Crystallization.
Academic Press.
Schmüdgen, K., 2017. The Moment Problem. volume 277 of Graduate Texts in Mathematics. Springer, Cham.
Shiea, M., Buffo, A., Vanni, M., Marchisio, D.L., 2020. A novel finite-volume TVD scheme to overcome non-realizability
problem in quadrature-based moment methods. Journal of Computational Physics 409, 109337.
Vanni, M., 2000. Approximate population balance equations for aggregation–breakage processes. J. Colloid Interface Sci. 221,
143–160.
Vikas, V., Wang, Z.J., Passalacqua, A., Fox, R.O., 2011. Realizable high-order finite-volume schemes for quadrature-based
moment methods. J. Comput. Phys. 230, 5328–5352.
Wheeler, J.C., 1974. Modified moments and Gaussian quadratures. Rocky Mt. J. Math. 4, 287–296.
Wick, A., Nguyen, T.T., Laurent, F., Fox, R.O., Pitsch, H., 2017. Modeling soot oxidation with the extended quadrature
method of moments. Proceedings of the Combustion Institute 36, 789–797.
Wilck, M., 2001. A general approximation method for solving integrals containing a lognormal weighting function. J. Aerosol
Sci. 32, 1111–1116.
Wright, D.L., 2007. Numerical advection of moments of the particle size distribution in Eulerian models. J. Aerosol Sci. 38,
352–369.
Yuan, C., Fox, R.O., 2011. Conditional quadrature method of moments for kinetic equations. J. Comput. Phys. 230, 8216–8246.
Yuan, C., Laurent, F., Fox, R.O., 2012. An extended quadrature method of moments for population balance equations. J.
Aerosol Sci. 51, 1–23.

22