---
title: Divergence Amplification Factor
url: https://www.emergentmind.com/topics/divergence-amplification-factor
type: topic
---

# Divergence Amplification Factor

Searching arXiv for the cited papers to ground the article in current records.
Divergence amplification factor denotes a quantity that tracks how a measurable response grows, saturates, or is effectively transformed when an underlying control parameter approaches a critical, singular, or information-theoretic boundary. In the cited literature, the term appears in three distinct technical settings: the field-amplitude amplification factor \(F_{\rm amp}(\theta)\) for light modes in a homogeneous interface layer near critical conditions [2606.16841], the geometric amplification factor \(A_{\rm geo}[{\cal C}]\) in non-Hermitian adiabatic evolution [2409.13595], and the worst-case amplification curve \(Post_{d,k,\alpha}(\epsilon)\) governing Rényi-divergence behavior under Bernoulli post-sampling in differential privacy [2105.10594]. The shared vocabulary reflects the presence of a divergence or near-divergence structure, but the underlying objects, observables, and operational meanings are different.

## 1. Domain-specific definitions

The principal definitions are structurally similar in that each introduces a normalized or worst-case factor, but they operate on different mathematical objects.

| Setting | Quantity | Definition |
|---|---|---|
| Interface optics | \(F_{\rm amp}(\theta)\) | \(F_{\rm amp}(\theta)\equiv |A(\theta)|/|A_0|\) |
| Non-Hermitian adiabatic evolution | \(A_{\rm geo}[{\cal C}]\) | \(\exp\!\Bigl[-2\int_{\cal C}\Im[{\cal A}_n(R)]\!\cdot dR\Bigr]\) |
| Bernoulli post-sampling and RDP | \(Post_{d,k,\alpha}(\epsilon)\) | \(\sup_{A:\epsilon_A(\alpha)\le\epsilon}\epsilon_{B_k\circ A}(\alpha)\) |

In the optical setting, the quantity is a field-amplitude ratio normalized to a reference angle away from criticality. In the non-Hermitian setting, it is the purely geometric contribution to norm change under adiabatic transport. In the privacy setting, it is a worst-case curve defined through a supremum over mechanisms or, equivalently, over distributions \(P,Q\) on \([0,1]^d\) subject to two Rényi-divergence constraints [2606.16841] [2409.13595] [2105.10594].

A common source of ambiguity is that “amplification” does not always mean literal increase of an experimentally measured amplitude. In optics it refers to the divergence of plane-wave basis amplitudes near \(k_1\to 0\); in non-Hermitian dynamics it refers to norm change induced by the imaginary part of a Berry-type connection; in privacy it measures how much smaller the \(\alpha\)-RDP parameter becomes after Bernoulli post-processing. This suggests that the phrase is best understood as a family resemblance rather than a single invariant notion.

## 2. Critical-incidence amplification in a homogeneous interface layer

For a homogeneous layer 1 in total reflection geometry, the field-amplitude amplification factor is defined by
\[
F_{\rm amp}(\theta)\equiv \frac{|A(\theta)|}{|A_0|},
\]
where \(A_0\equiv A(\theta_0)\) is taken at a reference angle well away from criticality. The normal wave-vector component in layer 1 is
\[
k_1(\theta)=\sqrt{(2\pi n_0/\lambda)^2\sin^2\theta-(2\pi n_1/\lambda)^2},
\]
and the amplification behaves as
\[
F_{\rm amp}(k_1)\sim |k_1|^{-1},
\]
up to a non-singular prefactor [2606.16841].

The transfer-matrix solution gives the two field coefficients in layer 1 in the schematic form
\[
A_\pm(\theta)=N_0(\theta)\exp(\mp ik_1d_1/2)/C(\theta),
\]
with common denominator
\[
C(\theta)\equiv k_1[k_0+k_2]\cos(k_1d_1)-i[k_1^2+k_0k_2]\sin(k_1d_1).
\]
Expanding numerator and denominator near criticality \(k_1\to 0\) yields
\[
A_\pm(\theta)=\pm i\,k_1^{-1}+B+O(k_1),
\]
where \(B\) is finite and depends on \(n_0,n_1,n_2,d_1,\lambda\) but not on \(k_1\). The singular contribution is therefore \(A(\theta)\sim k_1(\theta)^{-1}\) [2606.16841].

The relevant critical angle in the absence of absorption is
\[
\theta_1=\arcsin(n_1/n_0).
\]
Writing \(\Delta\theta\equiv \theta-\theta_1\), the small-\(\Delta\theta\) expansion gives
\[
k_1^2\simeq -(2\pi/\lambda)^2\,2n_0n_1\cos\theta_1\,\Delta\theta,
\]
hence
\[
k_1\simeq e^{\pm i\pi/4}(2\pi/\lambda)\sqrt{2n_0n_1\cos\theta_1|\Delta\theta|}.
\]
It follows that
\[
|A(\theta)|\propto |k_1|^{-1}\propto |\Delta\theta|^{-1/2}.
\]
The critical exponent is therefore \(-0.5\) [2606.16841].

The physical picture given for this divergence is that, exactly at critical incidence, the refracted wave in layer 1 travels tangent to the interfaces and “piles up” over an infinite distance, producing a \(1/k_1\to\infty\) amplification in the plane-wave basis amplitudes. Absorption and finite spatial coherence then act as cutoffs rather than as changes to the singular mechanism itself.

## 3. Absorption regularization and observable peak structure

In a weakly absorbing layer, one writes \(n_1^2=\epsilon_1'+i\epsilon_1''\). The exact \(k_1\) then satisfies
\[
k_1'^2+k_1''^2=(2\pi n_0/\lambda)^2\sqrt{(\epsilon_1'/n_0^2-\sin^2\theta)^2+(\epsilon_1''/n_0^2)^2}
\pm (2\pi n_0/\lambda)^2(\epsilon_1'/n_0^2-\sin^2\theta).
\]
The minimum possible \(|k_1|\) occurs at the “absorbing” critical angle
\[
\sin\theta_1=\sqrt{\epsilon_1'}/n_0,
\]
and has value
\[
|k_1|_{\min}=(2\pi/\lambda)\sqrt{\epsilon_1''}.
\]
The divergence is therefore regularized into a finite maximum whose saturated field peak-height is
\[
H_{\rm abs}\simeq \frac{1}{|k_1|_{\min}}=(\lambda/2\pi)\,\epsilon_1''^{-1/2}.
\]
For the half-width at half-maximum in angle, solving \(|k_1(\theta)|^{-2}\simeq \frac12 |k_1|_{\min}^{-2}\) gives
\[
\Delta\theta_{1/2}\simeq \frac{15\,\epsilon_1''}{16\,n_0^2\sin\theta_1\cos\theta_1},
\]
identified as the half-width at half maximum of the intensity [2606.16841].

The no-absorption treatment also introduces a peak-base width,
\[
\Delta\theta_{\rm base}=\lambda^2\,[8\pi^2 n_0n_1\cos\theta_1(1+d_1/2\,\ell_1)^2]^{-1},
\]
where \(\ell_1\) is the evanescent penetration depth in layer 2. This separates two related but distinct notions of width: a base width in the idealized singular setting, and an HWHM once absorption regularizes the singularity [2606.16841].

The experimental significance emphasized for visible-light Evanescent-Wave Dynamic Light Scattering is that the scattered intensity from an interface layer is proportional to \(|A(\theta)|^2\). Without absorption it would diverge as \(|\Delta\theta|^{-1}\); with finite absorption and finite beam-profile width, the result is a sharply peaked but finite scattering intensity at \(\theta_1\). The paper summary identifies this as the origin of the dramatic \(10^2\)–\(10^7\times\) amplification observed experimentally, and it links the effect to depth-selective (“tomographic”) measurements of interfacial fluctuations. The same summary notes a relation to surface plasmon resonance and states that in SPR second-harmonic generation the SHG field is proportional to \(A^2\), so the SHG intensity scales as \(|A|^4\), producing \(|\Delta\theta|^{-2}\) wings in the no-absorption regime [2606.16841].

## 4. Geometric amplification in non-Hermitian adiabatic evolution

In non-Hermitian adiabatic dynamics, the relevant object is not a spatially diverging field coefficient but a geometric contribution to intensity change along a path in parameter space. Let \(H(R)\) be a non-Hermitian Hamiltonian with nondegenerate right eigenstate \(|R(R)\rangle\) and left eigenstate \(\langle L(R)|\),
\[
H(R)|R(R)\rangle=E(R)|R(R)\rangle,\qquad
\langle L(R)|H(R)=\langle L(R)|E(R).
\]
Two Berry connections are introduced,
\[
{\cal A}^{LR}(R)\equiv i\frac{\langle L|\nabla_R R\rangle}{\langle L|R\rangle},
\qquad
{\cal A}^{RR}(R)\equiv i\frac{\langle R|\nabla_R R\rangle}{\langle R|R\rangle},
\]
and the non-Hermitian Berry connection is defined by
\[
{\cal A}_n(R)\equiv {\cal A}^{LR}(R)-{\cal A}^{RR}(R).
\]
For adiabatic variation \(R(t)\), the total intensity amplification factor
\[
\frac{I(T)}{I(0)}=\frac{\langle\psi(T)|\psi(T)\rangle}{\langle\psi(0)|\psi(0)\rangle}
\]
splits into a dynamical part, coming from \(\Im E\), and a purely geometric part. The geometric amplification factor is
\[
A_{\rm geo}[{\cal C}]\equiv
\exp\Bigl[-2\int_{\cal C}\Im[{\cal A}_n(R)]\cdot dR\Bigr]
\]
for a path \({\cal C}:R(0)\to R(T)\) [2409.13595].

Path-independence is controlled by the Berry curvature two-form
\[
\Omega_n(R)\equiv \nabla_R\times {\cal A}_n(R).
\]
If the parameter space is simply-connected, then
\[
\Im[\Omega_n(R)]=0
\]
is equivalent to the line integral being path-independent, so that \(A_{\rm geo}[{\cal C}]\) depends only on the endpoints \(R(0),R(T)\) and not on the detailed shape of \({\cal C}\) [2409.13595].

This formulation makes the amplification factor a geometric quantity in a precise sense: it is generated by the imaginary part of a Berry-type connection rather than by the instantaneous imaginary part of the eigenvalue. A plausible implication is that “divergence” in this context refers to the accumulation of geometric norm change under adiabatic steering, not to a local singularity in real space.

## 5. Symmetry classes and Petermann-factor formulations

The non-Hermitian analysis identifies four general situations in which \(\Im\Omega_n(R)=0\), often described as the existence of a metric or particle–hole symmetry relating left and right eigenvectors. With \(M\) a fixed parameter-independent matrix, these are:

\[
|L(R)\rangle=M|R(R)\rangle,\quad M^\dagger=M,
\]
\[
|L(R)\rangle=M|R(R)^*\rangle,\quad M^T=M,
\]
and the inverse relations
\[
|R(R)\rangle=M|L(R)\rangle,\quad M^\dagger=M,
\]
\[
|R(R)\rangle=M|L(R)^*\rangle,\quad M^T=M.
\]

In each case,
\[
\Xi(R)\equiv \Im[{\cal A}^{LR}-{\cal A}^{RL}]
\]
vanishes identically, and hence \(\Im\Omega_n=0\). The geometric amplification factor is then endpoint-only [2409.13595].

The Petermann factor is
\[
K(R)=\frac{\langle L|L\rangle\,\langle R|R\rangle}{|\langle L|R\rangle|^2}
=\frac{1}{\operatorname{Tr}[P_LP_R]},
\]
with
\[
P_L=|L\rangle\langle L|/\langle L|L\rangle,\qquad
P_R=|R\rangle\langle R|/\langle R|R\rangle.
\]
When \(M\) is a rank-1 projector, one obtains the linear ratio
\[
A_{\rm geo}=\frac{K(R(T))}{K(R(0))}.
\]
When \(M\) is unitary, and Hermitian or symmetric, one obtains the square-root ratio
\[
A_{\rm geo}=\sqrt{\frac{K(R(T))}{K(R(0))}}.
\]
The projector case is highlighted because it permits direct measurement of the change in Petermann factor from initial to final parameter [2409.13595].

Two examples are given. For the two-level Hamiltonian
\[
H(\Delta,J,\delta)=
\begin{pmatrix}
-\Delta & J+\delta\\[6pt]
J-\delta & \Delta
\end{pmatrix},
\]
in the real-eigenvalue regime \(\Delta^2+J^2\ge \delta^2\), the Petermann factor is
\[
K=\frac{\Delta^2+J^2}{\Delta^2+J^2-\delta^2}.
\]
When \(\Delta=0\), the Hamiltonian is unitarily equivalent to a symmetric one, so \(M\) is unitary and \(A_{\rm geo}=\sqrt{K_T/K_0}\); the same \(\sqrt{K_T/K_0}\) is reached for different driving paths in \((J,\delta)\). In the nonreciprocal robotic metamaterial example, an effective non-Hermitian SSH chain has a single zero-mode with right eigenvector \(\langle n|R\rangle\propto (a/b)^n\) and left eigenvector chosen fixed by setting \((a',b')\) constant. Then one is in the rank-1 projector case and
\[
A_{\rm geo}=\frac{K[a(t=T)]}{K[a(t=0)]}.
\]
The same study states that a slow adiabatic experiment can directly extract the final Petermann factor through
\[
K_f=K_i\,\frac{I(T)}{I(0)}.
\]
In a reciprocal realization \(a'=a,\ b'=b\), one instead has \(|L\rangle=|R^*\rangle\) and \(A_{\rm geo}=\sqrt{K_T/K_0}\) [2409.13595].

## 6. Divergence amplification in Bernoulli post-sampling and Rényi differential privacy

In the privacy setting, the divergence-amplification factor is formulated as a worst-case curve rather than as a local singular law. Fix \(\alpha>1\), let \(A\) be any randomized mechanism satisfying \((\alpha,\epsilon)\)-RDP, and assume \(A\) outputs \(\theta\in\Theta\subseteq[0,1]^d\). After running \(A\), one draws \(k\) independent Bernoulli samples with biases given by the coordinates of \(\theta\). The Bernoulli-sampling process \(B_k(\theta)\) returns \(k\) i.i.d. draws \(b^{(1)},\dots,b^{(k)}\in\{0,1\}^d\) with
\[
\Pr[b^{(j)}_\ell=1]=\theta_\ell
\]
independently for all \(j,\ell\). Writing
\[
\epsilon_A(\alpha)=\sup_{D\sim D'}R_\alpha(A(D)\|A(D')),
\]
the post-processed mechanism \(B_k\circ A\) has
\[
\epsilon_{B_k\circ A}(\alpha)=\sup_{D\sim D'}R_\alpha(B_k(A(D))\|B_k(A(D'))).
\]
The worst-case amplification curve is then
\[
Post_{d,k,\alpha}(\epsilon)
=
\sup_{A:\epsilon_A(\alpha)\le \epsilon}\epsilon_{B_k\circ A}(\alpha)
=
\sup_{\substack{P,Q\text{ on }[0,1]^d:\\ R_\alpha(P\|Q),R_\alpha(Q\|P)\le \epsilon}}
R_\alpha(B_k(P)\|B_k(Q)).
\]
This is the quantity called the divergence-amplification factor in the Bernoulli post-sampling analysis [2105.10594].

The exact computation begins from an infinite-dimensional optimization over \(P,Q\in\Delta([0,1]^d)\). Two reductions are stated. First, by a convexity and “corner-point” argument, worst-case \(P,Q\) can be taken on
\[
C_d=\{c,1-c\}^d,\qquad c\in(0,\tfrac12).
\]
Second, writing the Rényi-divergence constraints and the post-sampling divergence in closed form yields a convex program in the \(2\cdot 2^d\) variables \(\{x_i\}\cup\{y_i\}\), where
\[
P(z_i)=x_i,\qquad Q(z_i)=y_i,
\]
and
\[
z_i=(c^{i_1}(1-c)^{1-i_1},\dots,c^{i_d}(1-c)^{1-i_d}).
\]
For \(k=1\), the constraints are
\[
\sum_i (x_i/y_i)^\alpha y_i\le e^{(\alpha-1)\epsilon},
\qquad
\sum_i (y_i/x_i)^\alpha x_i\le e^{(\alpha-1)\epsilon},
\]
and for each \(b\in\{0,1\}^d\),
\[
x_b'=\sum_i x_i\,c^{\Delta(i,b)}(1-c)^{d-\Delta(i,b)},
\qquad
y_b'=\sum_i y_i\,c^{\Delta(i,b)}(1-c)^{d-\Delta(i,b)},
\]
with objective
\[
\max_{x,y}\sum_b (x_b'/y_b')^\alpha y_b'.
\]
The output is \((1/(\alpha-1))\log \mathrm{MaxVal}\). For general \(k\), \(b\) is replaced by \(k\)-tuples and the exponents become products over \(j=1,\dots,k\). The resulting program has \(O(2^d)\) variables and is solvable for small \(d\) by standard convex solvers [2105.10594].

The upper and lower bounds given for the amplification curve are
\[
Post_{d,k,\alpha}(\epsilon)\le \min\{\epsilon,\ dkr_\alpha(c)\},
\]
where
\[
r_\alpha(p)=\frac{1}{\alpha-1}\log\!\bigl[p^\alpha(1-p)^{1-\alpha}+(1-p)^\alpha p^{1-\alpha}\bigr].
\]
The \(\epsilon\) term follows from post-processing, while the \(dkr_\alpha(c)\) term comes from quasi-convexity of \(R_\alpha\) and independence across coordinates. The lower bound uses a two-point construction placing mass \(p\) and \(1-p\) on the all-\(c\) and all-\((1-c)\) corners, which gives \(R_\alpha(P\|Q)=r_\alpha(p)\) and
\[
R_\alpha(B_k(P)\|B_k(Q))
=
\frac{1}{\alpha-1}\log
\sum_{j=0}^{dk}\binom{dk}{j}P_j^\alpha Q_j^{1-\alpha}
\le
Post_{d,k,\alpha}(r_\alpha(p)),
\]
with
\[
P_j=p\,c^j(1-c)^{dk-j}+(1-p)c^{dk-j}(1-c)^j,
\]
\[
Q_j=(1-p)c^j(1-c)^{dk-j}+p\,c^{dk-j}(1-c)^j.
\]
The summary states that in many regimes, including large \(\epsilon\) or moderate-to-large \(d\), the upper and lower bounds coincide or become arbitrarily close, yielding either \(Post(\epsilon)\approx \epsilon\) or \(Post(\epsilon)\approx dkr_\alpha(c)\) depending on parameters [2105.10594].

A worked example is provided for \(d=2\), \(k=1\), \(\alpha=2\), \(c=0.3\), \(p=0.2\). In that case,
\[
r_2(p)=\epsilon=\log\!\bigl[p^2/(1-p)+(1-p)^2/p\bigr]\simeq 1.178,
\]
the upper bound gives
\[
Post(\epsilon)\le \min\{1.178,1.135\}=1.135,
\]
the lower bound yields
\[
R_2(B(P)\|B(Q))\simeq 0.391,
\]
and the convex solver finds \(Post(1.178)\approx 0.39\), so the two-point lower bound is tight in that case [2105.10594].

## 7. Comparative interpretation and limits of unification

Across these settings, the divergence amplification factor always quantifies a transformed response under a limiting process, but the limiting processes are not the same. In interface optics, the relevant limit is \(k_1\to 0\) at critical incidence, giving \(|A|\propto |\Delta\theta|^{-1/2}\) and \(|A|^2\propto |\Delta\theta|^{-1}\) in the absence of absorption [2606.16841]. In non-Hermitian adiabatic transport, the factor is a path integral of \(\Im[{\cal A}_n]\), with path-independence controlled by \(\Im[\Omega_n]=0\) and, in specific symmetry classes, reducible to ratios of Petermann factors [2409.13595]. In Bernoulli post-sampling, the factor is a worst-case Rényi-divergence curve obtained from a supremum over admissible mechanisms or distributions, with exact convex-program computation and upper/lower bounds [2105.10594].

A frequent misconception would be to treat these as interchangeable manifestations of a single formalism. The available material does not support that. The optical quantity is a normalized field amplitude in a transfer-matrix treatment; the non-Hermitian quantity is an adiabatic geometric norm factor; the privacy quantity is an extremal divergence functional over randomized mechanisms. What unifies them is not a common algebraic object but the repeated appearance of singular, endpoint, or worst-case amplification behavior under precisely defined constraints.

A plausible implication is that the phrase “divergence amplification factor” is best reserved for context-specific use, with the relevant state space, observable, and regularization mechanism made explicit. In the optical case, absorption and finite spatial coherence cut off the divergence to a finite peak. In the non-Hermitian case, the central distinction is between dynamical amplification from \(\Im E\) and geometric amplification from \(\Im[{\cal A}_n]\). In the privacy case, the central distinction is between exact worst-case optimization and computable bounds such as \(\min\{\epsilon,dkr_\alpha(c)\}\). These distinctions determine both the interpretation and the practical use of the factor in each domain.

Source: https://www.emergentmind.com/topics/divergence-amplification-factor