Sum Factorization: Unified Framework

• Sum factorization is a method that decomposes a global sum into simpler summands or sequential one-dimensional contractions, revealing underlying product structures.
• It is applied across diverse fields—from topology and algebra to numerical analysis and quantum physics—to improve computational efficiency and theoretical clarity.
• By converting high-dimensional or composite problems into controlled, lower-order components, sum factorization enables more tractable analysis and resource-efficient algorithms.

Sum factorization is a polysemous technical term used in several research areas to denote a decomposition of a global object into simpler summands or a reordering of sums that exposes hidden product structure. In topology it refers to connected-sum constructions for boundary-parabolic representations of knot groups, together with additive and multiplicative laws for invariants (Cho, 2014). In additive and arithmetic algebra it refers to decompositions into sums of atoms, coprime or gcd-restricted divisor-sum factorizations, and balanced factorizations whose factors have zero sum (Jiang et al., 2023, Mousavi et al., 2018, Klyachko et al., 2015). In scientific computing it denotes the reorganization of tensor-product quadrature and basis contractions into sequences of one-dimensional contractions (Bressan et al., 2018). In mathematical physics it appears as a sum over operator factorizations for the Coulomb Hamiltonian, as exact reduction of two-loop vacuum sum-integrals to products of one-loop objects, and as the interplay of factorization, resummation, and sum rules in heavy-to-light form factors (Lyu et al., 2022, Davydychev et al., 2023, Wang, 2016). The shared theme is not a single formal definition but a recurrent strategy: convert an apparently high-dimensional, non-unique, or composite structure into a controlled composition of lower-complexity pieces.

1. Terminological scope and structural motifs

Across the literature, sum factorization denotes either an additive decomposition of an object into constituent factors or a computational factorization of nested sums. In the additive-monoid setting, an atomic commutative monoid is studied through representations

$b=a_1+\cdots+a_k$

with each $a_i$ an atom, and structural finiteness is encoded by the length-finite factorization property, requiring $\#\mathsf{Z}_M^{(k)}(b)<\infty$ for every element $b$ and every prescribed length $k$ (Jiang et al., 2023). In isogeometric analysis, by contrast, sum factorization reorders multidimensional quadrature expressions so that a $d$-dimensional contraction is realized as a sequence of $d$ one-dimensional contractions, exploiting tensor-product structure in bases and quadrature rules (Bressan et al., 2018). In topology, the phrase is tied to connected sums of representations, where the central issues are existence of summand representations, non-uniqueness of assembly, and additivity or product formulas for invariants (Cho, 2014).

A common structural pattern is the replacement of a global object by a finite combination of local or lower-order building blocks. In the arithmetic setting of Lambert-series-type identities, the coefficient sequence of a generating function is represented through lower-triangular matrices built from partitions, coprimality kernels, or gcd-restricted sums (Mousavi et al., 2018). In finite-temperature field theory, massless bosonic two-loop vacuum sum-integrals are reduced exactly to finite linear combinations of products of one-loop sum-integrals (Davydychev et al., 2023). In operator-theoretic quantum mechanics, the hydrogen Hamiltonian is written as

$$\hat{\mathcal H}=\sum_{\alpha=x,y,z}\hat A_\alpha^\dagger(1)\hat A_\alpha(1)+E(1),$$

so that the Hamiltonian is represented as a sum over coupled Schrödinger factorizations rather than by conventional separation of variables (Lyu et al., 2022).

This suggests that “sum factorization” is best understood as a family of factorization paradigms rather than a single invariant definition. The relevant distinctions are whether the summands are algebraic atoms, tensor-product contractions, connected-sum components, matrix kernels, or operator factors.

2. Additive and arithmetic factorization frameworks

In additive commutative monoids, sum factorization is literal factorization by addition. For an atomic monoid $M$, one studies the factorization set $\mathsf{Z}_M(b)$, the set of lengths $a_i$0, and the fixed-length subsets $a_i$1 (Jiang et al., 2023). The paper on the length-finite factorization property identifies atomic co-well-ordered positive monoids as a large class satisfying LFF, including all decreasing positive monoids. It also establishes the equivalence

$a_i$2

and exhibits examples separating LFF, BFP, ACCP, and FFP. Grams’ monoid and certain geometric rational monoids are LFF but fail ACCP, while $a_i$3 is BFP but not LFF because the element $a_i$4 has infinitely many length-$a_i$5 factorizations (Jiang et al., 2023). The point is that finiteness can hold at each prescribed length even when total non-unique factorization remains unbounded.

Arithmetic generating-function factorization gives a different but related meaning. For Lambert series,

$a_i$6

the coefficients $a_i$7 are encoded by a lower-triangular partition matrix $a_i$8 (Mousavi et al., 2018). The same paper develops type I factorizations for coprime divisor sums

$a_i$9

and type II factorizations for Anderson–Apostol sums

$\#\mathsf{Z}_M^{(k)}(b)<\infty$0

The resulting kernels $\#\mathsf{Z}_M^{(k)}(b)<\infty$1 and $\#\mathsf{Z}_M^{(k)}(b)<\infty$2 are explicit, invertible, and tied to partition numbers, the Möbius function, Euler’s totient, Ramanujan sums, and discrete Fourier transforms over gcd-structured inputs (Mousavi et al., 2018). Here sum factorization means replacing arithmetic sums by matrix actions with partition-theoretic kernels.

Several papers study factorization under additive or zero-sum constraints. Balanced factorisations ask for

$\#\mathsf{Z}_M^{(k)}(b)<\infty$3

in a ring $\#\mathsf{Z}_M^{(k)}(b)<\infty$4 (Klyachko et al., 2015). Over $\#\mathsf{Z}_M^{(k)}(b)<\infty$5, every element admits a balanced factorisation into $\#\mathsf{Z}_M^{(k)}(b)<\infty$6 factors for every $\#\mathsf{Z}_M^{(k)}(b)<\infty$7, while the case $\#\mathsf{Z}_M^{(k)}(b)<\infty$8 fails uniformly and the paper notes an update that every rational number admits a balanced factorisation into four rational numbers (Klyachko et al., 2015). Over finite fields, the paper gives a complete classification of pairs $\#\mathsf{Z}_M^{(k)}(b)<\infty$9 for which every element of $b$0 has such a zero-sum product factorization.

Two further number-theoretic usages are more specialized. The function $b$1 is additive in the multiplicative argument,

$b$2

so it encodes a “sum factorization” of an integer by adding prime factors with multiplicity (Sharipov, 2011). The averaged function

$b$3

is studied numerically, leading to conjectures of the form $b$4 with suggested exponents $b$5 and $b$6 (Sharipov, 2011). In the theory of sum systems, the associated divisor functions $b$7 count ordered factorizations with prescribed non-triviality, and the number of sum systems with component cardinalities $b$8 is

$b$9

linking additive set-sum constructions to ordered factorisations and square-free filters (Lettington et al., 2019).

3. Connected sums, direct sums, and non-unique decomposition

In knot theory, sum factorization appears in the connected sum of boundary-parabolic representations of knot groups (Cho, 2014). Given oriented knots $k$0 with boundary-parabolic representations

$k$1

one uses arc-colorings in the quandle of parabolic elements to glue diagrams and obtain a representation of $k$2. The original claim that the connected sum representation is well-defined up to conjugation was corrected by an erratum: different choices of coinciding arcs and conjugations can yield non-conjugate connected-sum representations (Cho, 2014). What survives is a corrected factorization statement: for a boundary-parabolic representation $k$3 of a composite knot $k$4, there exist summand representations $k$5, unique up to conjugation, such that one connected-sum construction yields $k$6.

The significance of that corrected factorization is that invariants behave as expected even though assembly is non-unique. The complex volume satisfies

$k$7

and Wada’s twisted Alexander polynomial satisfies

$k$8

for boundary-parabolic lifts to $k$9 (Cho, 2014). The trefoil–figure-eight example $d$0 explicitly verifies both formulas.

A different decomposition theory appears for homogeneous polynomials. A form $d$1 is a direct sum if, after a linear change of variables, it can be written as

$d$2

for a nontrivial decomposition $d$3 (Fedorchuk, 2017). Under the assumptions $d$4 and good characteristic, direct-sum decomposability is equivalent to a product factorization of the associated form $d$5 of the Milnor algebra:

$d$6

after a linear change of dual variables (Fedorchuk, 2017). The additive splitting of $d$7 is thus encoded multiplicatively in the Macaulay inverse system.

Frieze patterns provide a third decomposition theory in which non-uniqueness is intrinsic. For a $d$8-quiddity cycle $d$9 defined by

$d$0

Cuntz’s sum operator

$d$1

glues two quiddity cycles into a longer one (Weber et al., 2018). The paper proves that $d$2 is neither commutative nor associative on raw sequences, though both properties are restored up to dihedral equivalence classes. More importantly, decomposition into irreducible factors is not unique, and “even under stronger assumptions, there is no canonical decomposition” (Weber et al., 2018). This is a precise instance of sum factorization without unique factorization.

4. Tensor-product sum factorization in numerical analysis

In isogeometric analysis, sum factorization is a computational technique for tensor-product discretizations rather than an algebraic decomposition (Bressan et al., 2018). For tensor-product basis functions and tensor-product quadrature,

$d$3

a naive quadrature-based bilinear form repeats the same one-dimensional multiplications across many basis-pair and quadrature combinations. Sum factorization reorders the multidimensional sum so that the $d$4-dimensional contraction is evaluated via a sequence of $d$5 one-dimensional contractions (Bressan et al., 2018).

The paper’s key contribution is to apply this idea globally or on macro-elements rather than element-wise. For Gauss quadrature in $d$6 dimensions with spline order $d$7, the resulting assembly complexity grows as $d$8 instead of $d$9 as previously achieved by per-element sum factorization (Bressan et al., 2018). With weighted quadrature, the bound improves further to $$\hat{\mathcal H}=\sum_{\alpha=x,y,z}\hat A_\alpha^\dagger(1)\hat A_\alpha(1)+E(1),$$0 for assembly and $$\hat{\mathcal H}=\sum_{\alpha=x,y,z}\hat A_\alpha^\dagger(1)\hat A_\alpha(1)+E(1),$$1 for matrix-free application. The tensor-product mass and stiffness forms on the parametric domain are written using the Jacobian $$\hat{\mathcal H}=\sum_{\alpha=x,y,z}\hat A_\alpha^\dagger(1)\hat A_\alpha(1)+E(1),$$2, the contravariant metric tensor $$\hat{\mathcal H}=\sum_{\alpha=x,y,z}\hat A_\alpha^\dagger(1)\hat A_\alpha(1)+E(1),$$3, and tensor-product B-spline or NURBS bases, and the same nested one-dimensional contraction principle applies to basis values, derivatives, and geometry terms (Bressan et al., 2018).

The computational significance is twofold. First, the method turns repeated multidimensional local work into reusable one-dimensional slices with better arithmetic intensity. Second, macro-element variants retain the asymptotic complexity of the global algorithm while improving memory locality, cache behavior, and parallelization. Numerical experiments on $$\hat{\mathcal H}=\sum_{\alpha=x,y,z}\hat A_\alpha^\dagger(1)\hat A_\alpha(1)+E(1),$$4 two-dimensional meshes and $$\hat{\mathcal H}=\sum_{\alpha=x,y,z}\hat A_\alpha^\dagger(1)\hat A_\alpha(1)+E(1),$$5 three-dimensional meshes show fitted $$\hat{\mathcal H}=\sum_{\alpha=x,y,z}\hat A_\alpha^\dagger(1)\hat A_\alpha(1)+E(1),$$6-exponents for macro/global methods substantially below standard and per-element approaches, with measured exponents about $$\hat{\mathcal H}=\sum_{\alpha=x,y,z}\hat A_\alpha^\dagger(1)\hat A_\alpha(1)+E(1),$$7–$$\hat{\mathcal H}=\sum_{\alpha=x,y,z}\hat A_\alpha^\dagger(1)\hat A_\alpha(1)+E(1),$$8 in 2D and about $$\hat{\mathcal H}=\sum_{\alpha=x,y,z}\hat A_\alpha^\dagger(1)\hat A_\alpha(1)+E(1),$$9–$M$0 in 3D for the macro/global variants (Bressan et al., 2018). The same paper also discusses HB and THB extensions, multi-patch assembly, and limitations such as trimmed geometries.

5. Operator, integral, and sum-rule factorizations in physics

One physics usage of sum factorization is operator-theoretic. For the hydrogenic Coulomb Hamiltonian,

$M$1

the Cartesian operator factorization method writes

$M$2

with

$M$3

(Lyu et al., 2022). The coupled Cartesian factors do not commute because of their $M$4-dependence. A second radial factorization using $M$5 and harmonic polynomials $M$6 generates the full hydrogenic bound spectrum

$M$7

together with coordinate-space Laguerre-polynomial wavefunctions and momentum-space Gegenbauer-polynomial wavefunctions (Lyu et al., 2022). The method thus replaces conventional separation into radial and angular variables by a sum over coupled first-order factorizations.

A second meaning arises in thermal quantum field theory. The paper on two-loop vacuum sum-integrals studies

$M$8

for bosonic Matsubara frequencies (Davydychev et al., 2023). The main theorem proves exact factorization into finite linear combinations of products of one-loop sum-integrals $M$9, with no remainder terms. The proof first maps the thermal object to double sums over massive continuum two-loop vacuum integrals with collinear masses, factorizes those exactly into one-loop tadpoles, and then performs the Matsubara sums in combinations where unknown double sums and multiple zeta values cancel (Davydychev et al., 2023). Example reductions include

$\mathsf{Z}_M(b)$0

In this setting, sum factorization is an exact analytic reduction of two-loop thermal structures to one-loop masters.

A third usage combines factorization, resummation, and sum rules in heavy-to-light QCD form factors. In the large-recoil limit, $\mathsf{Z}_M(b)$1 form factors satisfy

$\mathsf{Z}_M(b)$2

and the vacuum-to-$\mathsf{Z}_M(b)$3-meson correlator used in the light-cone sum-rule construction factorizes into hard coefficients, jet functions, and $\mathsf{Z}_M(b)$4-meson distribution amplitudes (Wang, 2016). The same paper derives NLL-resummed sum rules for $\mathsf{Z}_M(b)$5, exhibits one-loop hard and jet functions, and uses the resulting form factors to extract

$\mathsf{Z}_M(b)$6

(Wang, 2016). Here the phrase “sum factorization” refers not to a single theorem but to the concerted use of short-distance factorization, renormalization-group resummation, and dispersion-theoretic sum rules.

6. Integer factorization, smoothness engineering, and summation-based algorithms

Several papers use “sum factorization” for integer factorization methods based on exponential sums or engineered additive decompositions. In Gauss-sum factorization, one studies truncated sums

$\mathsf{Z}_M(b)$7

or their real-part variants and identifies factors $\mathsf{Z}_M(b)$8 by constructive interference (Wölk et al., 2012). Complete reciprocate sums satisfy exact factor-detection rules, while truncated versions suffer from ghost factors. A later transmon-qubit implementation analyzes the worst Type II ghost factors and introduces the discernability

$\mathsf{Z}_M(b)$9

together with preprocessing that removes dangerous small reduced denominators $a_i$00 by dividing out small prime powers of $a_i$01 and restricting allowed trial divisors (Zaw et al., 2021). For $a_i$02, preprocessing increases the reported discernability from $a_i$03 to $a_i$04 on resonance, and restores positive discernability under intentional detuning noise (Zaw et al., 2021).

A different classical heuristic is Smooth Subsum Search, which replaces sieving by congruence engineering (Hittmeir, 2023). With

$a_i$05

one fixes a small factor base subset $a_i$06, chooses residue conditions so that $a_i$07 divides $a_i$08 by CRT, and then searches for small shifts $a_i$09 that force additional primes from the larger factor base to divide $a_i$10 (Hittmeir, 2023). The remaining cofactor is therefore smaller, increasing the Dickman–de Bruijn smoothness probability. In Python benchmarks against SIQS implementations, the paper reports consistent speedups of about $a_i$11 to $a_i$12 times for $a_i$13–$a_i$14 digits and about $a_i$15 times for $a_i$16–$a_i$17 digits (Hittmeir, 2023). This is sum factorization in the literal sense of representing smoothness candidates as sums constructed to have prescribed divisibility.

The 2025 summation-based algorithm gives an even more direct additive-to-multiplicative conversion (Friedlander, 29 Apr 2025). Writing

$a_i$18

it seeks factors in the form

$a_i$19

with $a_i$20. Expanding

$a_i$21

reduces factoring to solving an integer divisibility condition for $a_i$22 after choosing $a_i$23 and scanning $a_i$24 (Friedlander, 29 Apr 2025). The stated complexity is on the order of $a_i$25, so the method is not asymptotically competitive with subexponential sieves, but it provides a deterministic binary-structure reinterpretation of factorization.

Taken together, these works show that “sum factorization” in computational number theory usually means one of two things: using oscillatory sums whose interference encodes divisibility, or constructing arithmetic candidates as sums chosen to impose factor-base divisibility constraints in advance. Both approaches exploit additive structure to expose multiplicative information.

7. Conceptual synthesis and recurrent limitations

Despite the diversity of contexts, the same structural tensions recur. One is the contrast between existence and uniqueness. Boundary-parabolic knot representations factor uniquely into prime-summand representations up to conjugation, but the connected-sum assembly is not unique (Cho, 2014). Quiddity cycles always decompose into irreducibles, but there is no canonical decomposition (Weber et al., 2018). Positive monoids may satisfy fixed-length finiteness while still having unbounded sets of lengths or infinite elasticity (Jiang et al., 2023). In these examples, sum factorization provides structure without unique factorization.

A second recurrent issue is normalization. Complex volume is only defined modulo $a_i$26 because of flattening and logarithm-branch choices in the extended Bloch group model (Cho, 2014). Twisted Alexander polynomials require the universal factor $a_i$27 for exact multiplicativity under connected sum (Cho, 2014). Weighted quadrature improves asymptotic complexity in IGA, but it breaks symmetry of the resulting matrix and requires care for accuracy and stability (Bressan et al., 2018). In Gauss-sum factorization, preprocessing and dynamic cutoffs are needed because finite coherence and Type II ghost factors otherwise destroy reliable discrimination (Zaw et al., 2021).

A third recurrent theme is that “sum factorization” often complements rather than replaces classical frameworks. In heavy-to-light QCD, it augments SCET/QCDF with sum-rule machinery (Wang, 2016). In thermal field theory, it eliminates a whole family of two-loop bosonic vacuum sum-integrals by expressing them through one-loop masters (Davydychev et al., 2023). In direct-sum decomposability of forms, additive splitting is detected through multiplicative factorization of an associated dual object (Fedorchuk, 2017). In IGA, tensor-product structure is not changed, only reorganized computationally (Bressan et al., 2018).

The term therefore names a methodology rather than a single theorem: expose separability, summand structure, or hidden product laws inside an apparently monolithic sum. Whether the aim is invariant additivity, fixed-length finiteness, reduced quadrature complexity, analytic reduction to master objects, or divisibility detection, the central move is the same—rewrite the problem so that summation no longer obscures factorization.