---
title: Modulated Relative Functional Overview
url: https://www.emergentmind.com/topics/modulated-relative-functional
type: topic
---

# Modulated Relative Functional Overview

“Modulated relative functional” is not a standardized term in the cited literature. The nearest formal usages are the **Riesz modulated energy** \(F_N(X_N,\mu)\), **modulated free energy** \(E_N(f_N,\mu)\), and **relative entropy** \(H_N(f_N\mid g_N)\) in mean-field analysis, together with a context-dependent **functional relative position encoding with progressive interpolation** in long-context Transformers and transmitter-state-aware modulation in functionalized vesicle-based molecular communication. Across these settings, the phrase denotes, either formally or by close conceptual analogy, a quantity defined relative to a reference state and modulated by transport, context scale, or an internal functional state [2511.13461] [2402.13977] [2310.04418] [2510.25676].

## 1. Terminology and conceptual range

In the singular-interaction literature, the formal object is a **modulated energy** or **modulated free energy**. In the long-context Transformer literature, the formal object is a **functional relative position encoding with progressive interpolation**, abbreviated FIRE. In molecular communication, the exact phrase does not appear; the formal names are **memory-aware modulation (MAM)** and **memory-erasing modulation (MEM)**, both designed around the transmitter’s internal biochemical state relative to a threshold [2511.13461] [2310.04418] [2510.25676].

| Domain | Formal object | Representative expression |
|---|---|---|
| Coulomb/Riesz mean-field theory | Riesz modulated energy | \(F_N(X_N,\mu)\) |
| Whole-space log/Riesz diffusions | Modulated free energy | \(E_N(f_N,\mu)\) |
| Long-context Transformers | FIRE | \(b_{ij}^{(h)}=f_h(\phi(i,j))\) |
| Vesicle-based molecular communication | MAM / MEM | \(c_{H,\mathrm{in}}(t)\) relative to \(c_H^{\mathrm{th}}\) |

A plausible unifying description is that a modulated relative functional compares a microscopic, discrete, or instantaneous state with a reference object, but does so in a geometry or control variable native to the system. In Coulomb/Riesz theory the geometry is induced by the interaction kernel \(g\); in FIRE it is induced by a learned function of normalized relative distance; in vesicle-based molecular communication it is induced by the internal proton concentration relative to an activation threshold. This synthesis is interpretive rather than a named common framework.

## 2. Canonical mathematical instance: Riesz modulated energy

The clearest formal realization of a modulated relative functional is the **Riesz modulated energy**. For pairwise distinct points \(X_N=(x_1,\dots,x_N)\in(\mathbb R^d)^N\), empirical measure
\[
\mu_N:=\frac1N\sum_{i=1}^N\delta_{x_i},
\]
and target probability measure \(\mu\), the functional is
\[
F_N(X_N,\mu):=\frac12\int_{(\mathbb R^d)^2\setminus\triangle} g(x-y)\,d\Big(\frac1N\sum_{i=1}^N\delta_{x_i}-\mu\Big)(x)\,d\Big(\frac1N\sum_{i=1}^N\delta_{x_i}-\mu\Big)(y),
\]
with
\[
g(x)=
\begin{cases}
\frac1s|x|^{-s}, & s\neq 0,\\
-\log|x|, & s=0,
\end{cases}
\qquad 0\le s<d,
\]
and \(\triangle=\{(x,y)\in(\mathbb R^d)^2:x=y\}\) excised to remove self-interactions [2511.13461].

This functional is “modulated” because it is the interaction energy of the **difference measure** \(\mu_N-\mu\), and “relative” because it compares a microscopic empirical state to a continuum reference. It is also a renormalized object rather than a metric in the ordinary sense. The singularity of \(g\) at the origin implies that \(F_N\) is not simply positive as written; the diagonal must be removed, and the paper emphasizes its **almost positivity** through lower bounds such as
\[
F_N(X_N,\mu)-\frac{\log\lambda}{2N}\mathbf 1_{s=0}+C\|\mu\|_{L^p}\lambda^{d-s}\ge 0,
\qquad
\lambda:=(N\|\mu\|_{L^\infty})^{-1/d}.
\]
For \(p=\infty\), the additive scale is \(N^{\frac{s}{d}-1}\), which is identified as optimal [2511.13461].

The same work derives a truncated-potential decomposition using
\[
g_\eta(x):=\mathsf{c}_{\phi,d,s}\lim_{T\rightarrow\infty}\Big(\int_\eta^T t^{d-s}\phi_t(x)\frac{dt}{t}-C_{\phi,T}\mathbf1_{s=0}\Big),
\qquad
f_\eta:=g-g_\eta,
\]
which isolates the diagonal singularity and yields principal nonnegative pieces. This is the structural reason the functional behaves like a relative discrepancy measure after renormalization. It also admits a Fourier-side coercive interpretation controlling negative Sobolev norms; in particular, the cited results state that
\[
\left\|\frac1N\sum_{i=1}^N\delta_{x_i}-\mu\right\|_{H^{-r/2}}^2
\]
is controlled by the modulated energy up to the optimal additive error, so \(F_N\to0\) implies \(\mu_N\rightharpoonup\mu\) [2511.13461].

A related development proves **sharp commutator estimates of all order** for the Coulomb and super-Coulomb range \(d-2\le s<d\), with arbitrary-order transport derivatives, sharp additive error, sharp density dependence, and localization to the support of the transport [2407.15650]. That work presents the same functional as a singular, renormalized, distance-like comparison between \(\mu_N\) and \(\mu\), and strengthens its differential control by establishing localized estimates in terms of a localized modulated energy \(F_N^\Omega(X_N,\mu)\).

## 3. Transport derivatives, commutators, and sharp mean-field control

The differential structure of the Riesz modulated energy is central to its use as a relative functional. For a vector field \(v:\mathbb R^d\to\mathbb R^d\), transporting both particles and background by \(I+t v\) gives
\[
\frac{d^n}{dt^n}\Big|_{t=0}F_N\big((I+t v)^{\oplus N}(X_N),(I+t v)\#\mu\big)
=
\frac12\int_{(\mathbb R^d)^2\setminus\triangle}\nabla^{\otimes n}g(x-y):(v(x)-v(y))^{\otimes n}\,d(\mu_N-\mu)^{\otimes2}(x,y).
\]
The first-order case is the basic transport derivative entering mean-field arguments [2511.13461].

The 2025 sharp commutator estimate shows that this derivative is controlled by the modulated energy itself, with the optimal additive \(N\)-dependent error for all \(0\le s<d\), including the sub-Coulomb case. The main bound takes the form
\[
\Bigg|\int_{(\mathbb R^d)^2\setminus\triangle}(v(x)-v(y))\cdot\nabla g(x-y)\,
d\Big(\frac1N\sum_{i=1}^N\delta_{x_i}-\mu\Big)^{\otimes2}(x,y)\Bigg|
\]
\[
\le
C\Big(\|\nabla v\|_{L^\infty}+\||\nabla|^{a/2}v\|_{L^{\frac{2d}{a-2}}}\mathbf1_{a>2}\Big)
\Big(F_N(X_N,\mu)-\frac{\log\lambda}{2N}\mathbf1_{s=0}+C\|\mu\|_{L^\infty}\lambda^{d-s}\Big),
\]
with \(\lambda=(N\|\mu\|_{L^\infty})^{-1/d}\). The additive term \(\lambda^{d-s}\sim N^{\frac{s}{d}-1}\) is identified as the optimal worst-case scale [2511.13461].

The proof reduces the derivative estimate to a commutator inequality through a new potential truncation scheme based on a **wavelet-type representation** of the Riesz potential, and then to averaged **Kato-Ponce type estimates**. The 2024 all-order paper in the Coulomb and super-Coulomb range develops a complementary local regularity theory for the associated commutators, using the Caffarelli–Silvestre extension structure and localized electric-field estimates [2407.15650].

These differential inequalities yield sharp mean-field rates. For first-order Hamiltonian and gradient flows, the cited results obtain the expected \(N^{\frac{s}{d}-1}\)-rate in modulated-energy distance, with logarithmic correction in the case \(s=0\) [2511.13461] [2407.15650].

## 4. Relative entropy and modulated free energy in self-similar variables

A second formal family of modulated relative functionals appears in the whole-space, no-confinement analysis of log/Riesz interacting diffusions. The basic objects are the normalized relative entropy
\[
H_N(f_N\mid g_N):=\frac1N\int_{(\mathbb R^d)^N}\log\!\left(\frac{f_N}{g_N}\right)\,df_N,
\]
the relative Fisher information
\[
I_N(f_N\mid g_N):=\frac1N\int_{(\mathbb R^d)^N}\left|\nabla_{X_N}\log\!\left(\frac{f_N}{g_N}\right)\right|^2\,df_N,
\]
the modulated energy \(F_N(X_N,\mu)\), and the modulated free energy
\[
E_N(f_N,\mu):=\frac1\beta H_N(f_N\mid \mu^{\otimes N})+\mathbb E_{f_N}\big[F_N(X_N,\mu)\big].
\]
Here the entropy is relative to the tensorized mean-field law, while the energetic term compares the empirical measure to the one-particle density \(\mu\) [2402.13977].

The novelty of the cited work is that these functionals are handled after a self-similar change of variables
\[
\xi=\frac{x}{\sqrt{t+1}},
\qquad
\tau=\log(t+1),
\]
which converts the whole-space problem into one with an effective quadratic confinement. Under this transformation, entropy is exactly invariant:
\[
H_N(f_N\mid g_N)=H_N(\bar f_N\mid \bar g_N),
\]
while the modulated energy becomes a time-dependent self-similar functional with
\[
F_N(X_N,\mu)=\bar F_N^\tau(\Xi_N,\bar\mu)+\frac{\tau}{4N}\mathbf1_{s=0},
\]
and
\[
\bar F_N^\tau(\Xi_N,\bar\mu)+\frac{\tau}{4N}\mathbf1_{s=0}
:=
\frac12\int_{(\mathbb R^d)^2\setminus\Delta}g_\tau(\xi-\eta)\,
d\Big(\frac1N\sum_{i=1}^N\delta_{\xi_i}-\bar\mu\Big)^{\otimes2}(\xi,\eta)
+\frac{\tau}{4N}\mathbf1_{s=0},
\]
where \(g_\tau(\xi)=e^{-\frac{s\tau}{2}}g(\xi)\) and \(k_\tau=\nabla g_\tau\) [2402.13977].

Because \(F_N\) is only almost positive, the paper introduces corrected coercive variants \(\mathcal E_N(f_N,\mu)\) and \(\mathcal E_N^\tau(\bar f_N,\bar\mu)\) by adding explicit \(o_N(1)\)-type terms. These corrected functionals satisfy Grönwall-type inequalities in self-similar time, conditional on bounds for
\[
\nabla^{\otimes2}\log\frac{\bar\mu^\tau}{\mu_\beta^\tau},
\]
where \(\mu_\beta^\tau\) is a time-dependent self-similar equilibrium. This yields finite-time propagation of chaos, uniform-in-time propagation under integrability assumptions, and polynomial-in-time generation of chaos when logarithmic Sobolev input is available [2402.13977].

A common misconception is that the whole-space problem is treated by the original functionals directly. The cited work instead shows that the effective confining structure emerges only after self-similar rescaling; in that frame, the relative and modulated functionals regain the coercive and commutator structure needed for the argument [2402.13977].

## 5. Functional relative position encodings as a modulated relative form

In long-context Transformers, the term appears most naturally as an interpretation of FIRE, **functional relative position encoding with progressive interpolation**, rather than as the paper’s official nomenclature. FIRE replaces a fixed positional table or hand-designed scalar bias by a learned function of relative distance. The recovered formulation is
\[
b_{ij}^{(h)}=f_h(x_{ij}),
\]
with \(f:\mathbb R\to\mathbb R^{N_H}\) implemented by a small MLP, for example
\[
f(x)=W_3\sigma(W_2\sigma(W_1x))
\quad\text{or}\quad
f(x)=W_3\sigma(W_1x),
\]
where \(\sigma=\mathrm{ReLU}\) and width \(d=32\) unless otherwise specified [2310.04418].

The position argument is not raw distance alone. The cited reconstruction gives normalized coordinates such as
\[
x_{ij}=\frac{i-j}{i+1},
\qquad
x_{ij}=\frac{j-i}{i+1}\in[-1,0),
\]
and a log-transformed version
\[
x_{ij}=\frac{\log(c|i-j|+1)}{\log(ci+1)+1}.
\]
This is why FIRE can be interpreted as a **modulated relative functional form**: it is relative because it depends on \(i-j\), functional because the bias is generated by a learnable function \(f\), and modulated because the argument of that function is normalized or interpolated according to current position or effective context length [2310.04418].

The abstract states two specific claims. First, the authors “theoretically prove that this can represent some of the popular relative position encodings, such as T5’s RPE, Alibi, and Kerple.” Second, they “empirically show that FIRE models have better generalization to longer contexts on both zero-shot language modeling and long text benchmarks” [2310.04418]. The supplied reconstruction further states that FIRE does not require a pre-defined max sequence length and can be applied directly to length generalization without tuning, whereas RoPE position interpolation methods require adaptation.

A common misunderstanding is to treat FIRE as merely another lookup-table relative position bias. The paper’s main conceptual move is instead a head-wise learned function on a bounded, normalized distance domain. This suggests that “modulated relative functional” is an apt descriptive label, but it remains an interpretive label rather than the paper’s title term [2310.04418].

## 6. Functionalized-transmitter-aware modulation and threshold-relative state control

In molecular communication, the exact phrase “modulated relative functional” is explicitly not used. The closest related idea is a modulation scheme designed around the **functionalized** state of a vesicle-based transmitter. The transmitter consists of an external binary LED modulator, a spherical vesicle-based transmitter, a free-space diffusion channel, and a fully absorbing receiver. The vesicle membrane contains a light-driven proton pump and an \(H^+\)/SM symporter, so signaling-molecule release occurs through the cascade
\[
\text{light}\rightarrow\text{proton pumping}\rightarrow\text{proton gradient}\rightarrow\text{symporter activity}\rightarrow\text{SM release}.
\]
The state variable driving memory is the intravesicular proton concentration \(c_{H,\mathrm{in}}(t)\), constrained by
\[
c_{H,\min}\le c_H(t)\le c_{H,\max}
\]
[2510.25676].

The paper proposes two novel schemes: **memory-aware modulation (MAM)** and **memory-erasing modulation (MEM)**. Both are built around a threshold proton concentration \(c_H^{\mathrm{th}}\), obtained from the Hill-kinetic activation law of the symporters. MAM chooses the LED pattern so that \(c_{H,\mathrm{in}}(t)\) reaches \(c_H^{\mathrm{th}}\) at the start of the next bit interval when a transition occurs, while MEM attempts to enforce
\[
c_{H,\mathrm{in}}(t=ib)=c_H^{\mathrm{th}},\qquad \forall i,
\]
so that every symbol boundary begins from approximately the same proton state [2510.25676].

The formal encoding rules are explicit. For MAM,
\[
l_{\mathrm{MAM}}(t)=\sum_{i=0}^{L-2}
\begin{cases}
b_i\,p(t-ib), & \text{if } b_i=b_{i+1},\\
\mathds{1}_{[0,b-d)}(t-ib), & \text{if } b_i=1,\ b_{i+1}=0,\\
\mathds{1}_{[b-l,b)}(t-ib), & \text{if } b_i=0,\ b_{i+1}=1,
\end{cases}
+b_{L-1}p(t-(L-1)b),
\]
where \(l\) is the loading time and \(d\) is the deloading time. For MEM,
\[
l_{\mathrm{MEM}}(t)=\sum_{i=0}^{L-1}
\begin{cases}
\mathds{1}_{[0,b-d)}(t-ib), & \text{if } b_i=1,\\
\mathds{1}_{[b-l,b)}(t-ib), & \text{if } b_i=0.
\end{cases}
\]
The central claim is that memory is mitigated directly at the transmitter, enabling low-complexity single-sample threshold detection at the receiver [2510.25676].

Here the most accurate terminology is not “modulated relative functional” but **functionalized-transmitter-aware modulation schemes**. Still, a plausible conceptual reading is that the modulation is relative to a target internal functional state \(c_H^{\mathrm{th}}\), rather than to absolute ON/OFF timing alone. The paper’s own terminology remains MAM and MEM [2510.25676].

## 7. Shared structure, limitations, and common misconceptions

Across the cited works, the phrase “modulated relative functional” does not designate a single universally accepted object. Instead, the literature exhibits several structurally similar constructions.

First, the functional is typically **relative to a reference state**: \(\mu_N-\mu\) in Riesz modulated energy, \(f_N\) relative to \(\mu^{\otimes N}\) in relative entropy, normalized distance relative to current context scale in FIRE, and transmitter proton concentration relative to \(c_H^{\mathrm{th}}\) in MAM and MEM. Second, the quantity is **modulated by a system-specific mechanism**: interaction kernel geometry, self-similar rescaling, context-length interpolation, or biochemical state shaping. Third, the resulting object is used because it matches the dynamics more faithfully than a naive discrepancy measure. This suggests a shared design principle, but not a uniform cross-disciplinary formalism.

Several misconceptions recur. One is that every modulated relative functional is nonnegative or metric-like. The Riesz modulated energy is not: the diagonal must be excised, and the lower bound is only recovered up to an optimal additive correction [2511.13461]. A second is that “modulated relative functional” is a formal label in the Transformer and molecular-communication papers. It is not; in those settings it is, at most, a close conceptual description of FIRE and of threshold-relative transmitter-state modulation [2310.04418] [2510.25676]. A third is that relative entropy and modulated energy are interchangeable. The whole-space self-similar analysis distinguishes them sharply: entropy is exactly invariant under the rescaling, whereas modulated energy acquires a time-dependent renormalization and must be corrected by explicit \(o_N(1)\)-terms to obtain a coercive functional [2402.13977].

In the strongest formal sense provided by the cited literature, a modulated relative functional is exemplified by the Coulomb/Riesz modulated energy and its free-energy extensions: a renormalized discrepancy functional, adapted to the governing interaction, possessing transport-derivative estimates, coercivity up to optimal finite-\(N\) errors, and direct applications to mean-field convergence, quasi-neutral limits, and fluctuation theory [2511.13461] [2407.15650]. In broader usage, the same phrase can describe any functional or modulation rule whose operative variable is not absolute state alone, but state measured relative to a reference and reshaped by a learned, dynamical, or physical modulation mechanism.

Source: https://www.emergentmind.com/topics/modulated-relative-functional