---
title: Interval Partition Functions Overview
url: https://www.emergentmind.com/topics/interval-partition-functions
type: topic
---

# Interval Partition Functions Overview

“Interval partition functions” does not denote a single standardized object. In current arXiv usage, the phrase covers several distinct constructions in which a partition of an interval—or of an interval-partition-valued state space—organizes an analytic, probabilistic, spectral, or physical quantity. In multifractal analysis, it refers to regularity-filtered Dirichlet series attached to refining interval partitions [1007.1467]. In probability, the nearest analogues are semigroups, transition kernels, and transform functionals of interval-partition-valued diffusions [2402.08548]. In spectral optimization, it refers to partition-dependent cost functionals on the unit interval [1807.05973]. In 3d supersymmetric field theory, it is a literal partition function on \(T^2\times[0,1]\), interpreted as a transition amplitude between boundary conditions [2509.23902]. In numerical global optimization, interval partitioning itself becomes the central mechanism controlling how sampled function values are attached to subregions of a domain [1103.2657].

## 1. Terminological scope and conceptual organization

The term is best understood as a family of context-dependent constructions rather than a universal invariant. The common datum is an interval partition together with an additional rule: a coarse regularity assignment, a branching law, a spectral cost, a gluing amplitude, or a reuse scheme for function evaluations. What varies is the mathematical role played by the resulting object.

| Context | Basic object | Representative source |
|---|---|---|
| Multifractal analysis | Regularity-indexed Dirichlet series \(\zeta^\mu_{\mathfrak P}(\alpha,s)\) | [1007.1467] |
| Interval-partition probability | Semigroups, excursion laws, Laplace-type functionals | [2402.08548] |
| Spectral optimization | Costs built from \(\lambda(I_j)\) over interval partitions | [1807.05973] |
| 3d \(\mathcal N=2\) QFT | Transition amplitude on \(T^2\times[0,1]\) | [2509.23902] |
| Global optimization | Partition strategy governing one-point summaries of cells | [1103.2657] |

A recurring misconception is that “partition function” here should always mean either a statistical-mechanical partition function or an integer-partition counting function. Neither is generally correct. In several of the relevant literatures, the phrase is used only by analogy, and in some papers the interval partition is primary while the associated “function” is a semigroup, a cost functional, or a distributional transform rather than a scalar series.

## 2. Multifractal partition zeta functions

A canonical meaning of interval partition function appears in the theory of partition zeta functions. For a Borel measure \(\mu\) on \([0,1]\) and a sequence of finite interval partitions \(\mathfrak P=\{\mathcal P_n\}_{n=1}^\infty\) with mesh tending to zero, the coarse Hölder regularity of an interval \(U\subset[0,1]\) of positive Lebesgue measure is defined by
\[
A(U):=\frac{\log \mu(U)}{\log |U|}.
\]
The values \(\alpha\in[-\infty,\infty]\) select the intervals whose mass-to-scale relation realizes that regularity. The associated \(\alpha\)-lengths form
\[
\mathcal L_{\mathfrak P}^\mu(\alpha):=\{\ell\mid \ell=|P_n^i|\text{ and }A(P_n^i)=\alpha\},
\]
and the partition zeta function is
\[
\zeta^\mu_{\mathfrak P}(\alpha,s)
=\sum_{\ell\in\mathcal L^\mu_{\mathfrak P}(\alpha)}\ell^s
=\sum_{n=1}^\infty\sum_{A(P_n^i)=\alpha}|P_n^i|^s,
\]
initially for \(\operatorname{Re}(s)\) larger than its abscissa of convergence [1007.1467].

The resulting abscissa-of-convergence function
\[
f_{\mathfrak P}^\mu(\alpha)
:=\inf\{\gamma\in\mathbb R\mid \zeta^\mu_{\mathfrak P}(\alpha,\gamma)<\infty\}
\]
is itself a multifractal spectrum. In self-similar settings generated by weighted iterated function systems satisfying the OSC, the theory becomes explicit. If a stage-\(K\) interval is encoded by a frequency vector \(k=(k_1,\dots,k_N)\), then
\[
|P|=r_1^{k_1}\cdots r_N^{k_N},\qquad
\mu(P)=p_1^{k_1}\cdots p_N^{k_N},
\]
and therefore
\[
\alpha(k)=\frac{\sum_{i=1}^N k_i\log p_i}{\sum_{i=1}^N k_i\log r_i}.
\]
Under the distinctness hypothesis denoted (H), the partition zeta function admits the closed form
\[
\zeta^\mu_{\mathfrak P}(\alpha(k),s)
=\sum_{n=1}^\infty \binom{nK}{nk_1,\dots,nk_N}(r_1^{k_1}\cdots r_N^{k_N})^{ns},
\qquad K=\sum_{i=1}^N k_i,
\]
and its abscissa of convergence equals the Hausdorff dimension of the corresponding Besicovitch subset [1007.1467].

This framework is designed as a multifractal analogue of geometric zeta functions of fractal strings, but with a decisive modification: intervals are grouped by coarse regularity rather than by geometric complement structure. The paper explicitly distinguishes partition zeta functions from earlier multifractal zeta functions by requiring an actual sequence of interval partitions as part of the data. In the binomial examples, the concave envelope of \(f_{\mathfrak P}^\mu\) recovers the classical geometric and symbolic multifractal spectrum, while for certain atomic measures the theory extends to a “tapestry of complex dimensions”
\[
\mathcal T_{\mathfrak P}^\mu
=\{(\alpha,\omega)\mid \omega\in\mathcal D_{\mathfrak P}^\mu(\alpha,W_\alpha)\},
\]
showing that interval-partition-based Dirichlet series can encode oscillatory complex-analytic structure as well as real multifractal exponents [1007.1467].

## 3. Interval-partition-valued diffusions and transform functionals

In probability, interval partitions usually appear as state spaces rather than as scalar partition functions. The basic state is an ordered collection of disjoint open intervals whose lengths sum to a total mass. In the most general abstract among the supplied sources, branching interval partition diffusions are built by letting interval widths evolve independently according to a general real-valued diffusion, under assumptions ensuring finite lifetimes and continuous generation of new intervals. Births are governed by the Pitman–Yor excursion measure of the underlying diffusion, while an associated spectrally positive Lévy process orders both the excursions and their start times [2402.08548].

This construction generalizes the self-similar interval-partition framework of Forman, Pal, Rizzolo, and Winkel. A natural implication is that the analytic surrogates for “interval partition functions” in this setting are the Markov semigroup, transition kernels, and Laplace- or multiplicative-type observables of block widths. The supplied extract explicitly identifies these as the nearest analogues of partition functions in the probabilistic sense and stresses that the ordered interval partition contains more information than a ranked mass partition because left-to-right order is part of the state [2402.08548].

A more explicit self-similar theory is developed for \({\rm SSIP}^{(\alpha)}(\theta_1,\theta_2)\)-evolutions. Here the total mass process is squared Bessel with dimension \(2\theta\), where
\[
\theta=\theta_1+\theta_2-\alpha\ge -\alpha.
\]
The parameters \(\theta_1\) and \(\theta_2\) act as left and right immigration parameters, while the subtraction of \(\alpha\) reflects emigration through the middle \({\tt BESQ}(-2\alpha)\)-type component. The paper proves pseudo-stationarity for a new interval-partition law whose ranked sequence of lengths has Poisson–Dirichlet distribution with parameters \(\alpha\) and \(\theta\), and it establishes a reversal symmetry exchanging left and right immigration [2011.13378]. In this regime, the main interval-partition functionals are total mass \(\|\beta\|\), reversal \(\mathrm{rev}(\beta)\), scaling \(c\odotip\beta\), concatenation, and diversity.

A further development is the type-2 interval partition evolution related to the Aldous diffusion. Its state is \((m_1^y,m_2^y,\alpha^y)\), consisting of two distinguished leftmost masses and a residual interval partition. The total mass process
\[
m_1^y+m_2^y+\|\alpha^y\|
\]
is \({\rm BESQ}[-1]\), and after de-Poissonization the three-mass projection becomes a Wright–Fisher diffusion with parameters \(\left(-\frac12,-\frac12,\frac12\right)\) up to degeneration. Resampling at degeneration produces a recurrent extension with Dirichlet stationary distribution and yields the first stationary Markovian projection of the conjectured Aldous diffusion described in the paper [1804.01205].

A different interval-partition dynamic appears in fragmentations with erasure. Starting from the trivial partition \(P_0=(0,1]\), each interval is split in two and then old break points are erased, producing a partition
\[
P_n=\{(a_{n,0},a_{n,1}],\ldots,(a_{n,n},a_{n,n+1}]\}
\]
with \(n+1\) subintervals and \(n\) break points. The recursive update
\[
a_{n,k}=p_{n,k}a_{n-1,k-1}+(1-p_{n,k})a_{n-1,k}
\]
admits a random-walk representation, and the empirical break-point measure
\[
g_n=\frac1n\sum_{k=1}^n\delta_{a_{n,k}}
\]
has a multiscale asymptotic theory: weak convergence to \(p\delta_0+(1-p)\delta_1\), interior vague limits after normalization by \(n/\sigma_n\), and endpoint limits after the nonlinear rescaling \(x\mapsto x^{1/n}\) governed by large deviations [2502.09114]. This literature again treats interval partitions as primary stochastic objects, with empirical measures and transforms playing the role of observables.

## 4. Spectral and algorithmic functionals built from interval partitions

A distinct usage arises in optimization over partitions of the unit interval. For
\[
I=(0,1),
\]
the admissible class
\[
\mathcal C_n=\{\{I_j\}_{j=1}^n:I_j=(x_{j-1},x_j),\ 0=x_0\le x_1\le\cdots\le x_n=1\}
\]
supports a family of spectral costs built from first Dirichlet Sturm–Liouville eigenvalues. For a bounded open interval \(A\),
\[
\lambda(A):=\min_{u\in H_0^1(A),\,u\neq 0}
\frac{\int_A p(x)u'(x)^2\,dx+\int_A q(x)u(x)^2\,dx}
{\int_A w(x)u(x)^2\,dx},
\]
and the finite-\(n\) optimization problem is
\[
\min_{\{I_j\}\in\mathcal C_n}\sum_{j=1}^n
\varphi\!\left(\frac{n}{\sqrt{\lambda(I_j)}}\right).
\]
The partition is encoded by a probability measure \(\mu_{\{I_j\}}\), and the averaged functionals \(F_n\) \(\Gamma\)-converge in \(\mathcal P(I)\) to
\[
F_\infty(\mu)
=
\frac1\pi\int_I f(x)\,\varphi\!\left(\frac{\pi s(x)}{f(x)}\right)\,dx
+\varphi^\infty\mathcal L(\{f=0\})
+\varphi(0)\mu^s(I),
\]
where \(\mu=f\mathcal L+\mu^s\) and \(s(x)=\sqrt{w(x)/p(x)}\) [1807.05973]. The unique minimizing density is
\[
f_\infty(x)=\frac{s(x)}{\int_I s(t)\,dt}.
\]
This asymptotic optimizer is independent of the particular convex cost \(\varphi\) and of the potential \(q\). The paper treats sums of positive and negative powers of eigenvalues and a heat-trace-type approximation within this framework, so here an interval partition function is literally a partition-dependent spectral functional.

In numerical global optimization, the emphasis shifts from a scalar cost to the partition mechanism itself. For a vector function
\[
f(x)=(f_1(x),\ldots,f_r(x))
\]
defined on an \(N\)-dimensional hyperinterval \(D\subset\mathbb R^N\), one-point-based algorithms seek a minimal description of \(f\) by recursively subdividing \(D\) and associating only one trial point with each subinterval. The paper isolates four design choices—partition strategy, interval selection rule, number of trial points, and trial-point location—and studies the first and fourth [1103.2657].

Two traditional strategies are contrasted. The first evaluates \(f\) at an internal point of each new sub-box, but each function value is used by only one interval. The second cuts through an internal point by \(N\) coordinate hyperplanes, creating \(2^N\) sub-boxes and sharing one evaluation across all of them; however, it generates combinatorial growth and structural redundancy, formalized by Proposition 1.1: every partition made by the second strategy creates one or two sub-cells having \(f(x)\) evaluated at two vertices [1103.2657].

The proposed strategy adapts a diagonal-algorithm partition rule. For a chosen cell \(D(t)=[a(t),b(t)]\), one selects a longest edge, defines split points \(u\) and \(v\) at the \(2/3\) and \(1/3\) positions along that coordinate, and replaces the parent by exactly three subintervals. In the one-point-based adaptation, \(f\) is initially evaluated at the original vertex \(a\), and thereafter only at \(u\); the point \(v\) is used solely for geometry. The resulting mesh is regular, each stored vertex value is evaluated at most once, and a database plus direct-address numeration allows later reuse by as many as \(2^N\) cells. The example reported in the paper produces \(21\) intervals after \(11\) iterations using only \(7\) function evaluations [1103.2657]. This suggests an algorithmic meaning of interval partition function in which the partition rule and the evaluation-reuse map together constitute the effective functional description of the domain.

## 5. Interval partition functions in 3d \(\mathcal N=2\) quantum field theory

In 3d supersymmetric gauge theory, the phrase takes its most literal form. The interval partition function is the supersymmetric partition function on
\[
T^2\times[0,1],
\]
with \((0,2)\) boundary conditions \(B_1,B_2\) imposed on the two boundary tori. It is interpreted as a transition amplitude between boundary conditions and is shown to factorize into sums of products of hemisphere partition functions on \(HS^2\times S^1\), together with normalization factors that encode the non-orthonormality of intermediate states [2509.23902].

For SQED with \(N\) flavors, the interval partition function localizes to a contour integral in a gauge holonomy \(s\), with poles at
\[
s=x_\alpha,\qquad \alpha=1,\dots,N.
\]
The corresponding hemisphere partition functions are
\[
Z_\alpha(q,x)
=
(q;q)_\infty\oint_{s=x_\alpha}\frac{ds}{2\pi i s}
\frac{1}{\prod_{i=1}^N (s^{-1}x_i;q)_\infty}
=
\prod_{i\neq\alpha}\frac{1}{(x_i/x_\alpha;q)_\infty}.
\]
The factorization formula is
\[
Z_{\mathrm{int}}(q,x)
=
\sum_{\alpha=1}^N
Z_\alpha(q,x)\,Z_\alpha(q,x^{-1})
\prod_{i\neq\alpha}
\Bigl(\sqrt{x_\alpha/x_i}-\sqrt{x_i/x_\alpha}\Bigr).
\]
Geometrically, the Higgs branch is \(\mathbb P^{N-1}\), the labels \(\alpha\) are its torus fixed points, and the normalization factor is the inverse \(K\)-theoretic norm square of the fixed-point class. The paper interprets the factorization as the factorization of an equivariant \(K\)-theory class \(E\) into two factors \(F\) and \(\widetilde F\), each represented by a hemisphere amplitude [2509.23902].

For \(U(k)\) Chern–Simons–Yang–Mills theory with \(N\) boundary Fermi multiplets, the interval partition function again localizes to a holonomy integral, and the hemisphere partition functions are proved to be affine \(\widehat{SU(N)}_k\) characters. The interval amplitude admits a diagonal factorization
\[
Z_{\mathrm{int}}^{N,k}(q,x,y)
=
\sum_\mu
f_{\mu\mu}^{N,k}(q)\,
\chi_\mu^{\widehat{SU(N)}_k}(q,x)\,
\chi_\mu^{\widehat{SU(N)}_k}(q,y),
\]
where \(\mu\) ranges over integrable highest weights of \(SU(N)\) at level \(k\) [2509.23902]. The normalization factors \(f_{\mu\mu}^{N,k}(q)\) arise from a deformed inner product on affine characters. In this setting, an interval partition function is neither a combinatorial counting function nor a multifractal Dirichlet series, but a gluing amplitude in a supersymmetric boundary-value problem.

## 6. Adjacent literatures and sources of confusion

Several nearby topics use similar words but study different objects. In the theory of antimonotonic Boolean functions, “intervals” are order intervals
\[
[\alpha,\beta]=\{\kappa:\alpha\le \kappa\le \beta\}
\]
in the poset of antichains, and the paper develops an interval algebra based on joins, projections, and an external product. These intervals partition the whole space into disjoint families and lead to recursive enumeration formulae for Dedekind numbers, but they are not interval partition functions in the analytic or probabilistic senses described above [1103.2877].

In number theory, the restricted partition function \(p_N(n)\) counts partitions of \(n\) into at most \(N\) parts, equivalently with largest part at most \(N\). Its analysis via \(q\)-partial fractions, Sylvester waves, Ramanujan sums, degenerate Bernoulli numbers, and Gaussian-Ramanujan sums concerns integer partitions rather than partitions of geometric intervals [2202.03603]. Likewise, interval parking functions are parking functions in which each car accepts a fixed interval of parking spaces; they define a reachability relation on permutations whose transitive closure is the bubble-sort order, but the subject is unrelated to interval-partition-valued functionals [2006.09321].

The practical consequence is that “interval partition function” is polysemous across fields. In multifractal geometry it is a regularity-stratified Dirichlet series; in stochastic analysis it is usually a semigroup or law on ordered interval partitions; in spectral optimization it is a partition-dependent cost; in supersymmetric QFT it is a genuine partition function on a manifold with interval factor; and in algorithmic optimization the decisive structure may be the interval partition rule itself. The shared theme is not a single formula, but the use of interval decompositions as the primary carriers of analytic information.

Source: https://www.emergentmind.com/topics/interval-partition-functions