---
title: 'Positive Magnitude: Concepts & Applications'
url: https://www.emergentmind.com/topics/positive-magnitude
type: topic
---

# Positive Magnitude: Concepts & Applications

“Positive magnitude” is not a single standardized term in the recent arXiv literature. It appears in several technically distinct senses: as a forecasted positive-valued climatic amplitude, as a magnitude-based selection rule in representation learning, as a robust statistic built from positive earthquake magnitude differences, and as part of the theory of Leinster magnitude on positive definite metric spaces. A broader adjacent literature studies positivity of quantities governed by size, scale, or amplitude, including operator forms, EFT couplings, parton distributions, and scattering amplitudes [1909.03789], [2102.07981], [2101.10030], [2404.15706], [1012.5863].

## 1. Positive El Niño magnitude as a complexity-prediction relation

In ENSO forecasting, the phrase acquires a directly predictive meaning. The paper “Complexity based approach for El Nino magnitude forecasting before the ‘spring predictability barrier’” introduces System Sample Entropy, or SysSampEn, as a complexity measure for the system of temperature anomaly time series in the Niño 3.4 region, using the 22 grid points in that region as a coupled system [1909.03789]. The quantity is defined by
\[
SysSampEn(m,p,l_{eff},\gamma)=-\log\!\left(\frac{A}{B}\right),
\]
where \(A\) counts similar subsequences of length \(m+p\), \(B\) counts similar subsequences of length \(m\), \(l_{eff}\) is the effective number of data points used per yearly calculation, and \(\gamma\) sets the similarity tolerance. In the formulation given in the paper, SysSampEn measures how likely two sequences that match for \(m\) days remain matching for the next \(p\) days.

The central empirical result is a strong positive correlation between the previous calendar year’s SysSampEn and the following El Niño’s magnitude. Across ERA-Interim, ERA5 air temperature, ERA5 SST, and JRA55-do SST, the average reported correlation is approximately \(r\approx 0.90\) for the best parameter choices, while the hindcast/forecast comparison of realized events gives about \(r=0.84\). For ERA-Interim, the strongest reported case uses \(m=60\) days, \(p=15\) days, \(l_{eff}=345\) days, and \(\gamma=9\), yielding \(r=0.99\). Older approaches based on averaged SampEn or Cross-SampEn perform much worse, with average correlation only about \(r=0.42\) [1909.03789].

Forecasting is implemented through a linear map,
\[
Y=aX+b,
\]
with \(Y\) the El Niño magnitude and \(X\) the previous year’s SysSampEn. The paper reports an average hindcast RMSE of \(0.23^\circ C\) for El Niño events from 1984 to 2017, and for the ongoing 2018 event it forecasts a weak El Niño with magnitude \(1.11\pm 0.23^\circ C\), where the uncertainty is taken as \(1\times RMSE\) [1909.03789]. In this setting, “positive magnitude” denotes a positive forecasted ENSO amplitude inferred from a prior complexity state. The authors also suggest a physical interpretation: stronger El Niños may be preceded by weaker horizontal synchronization and higher temporal disorder in temperature anomalies.

## 2. Magnitude-based binarization in neural networks

In binary neural networks, “positive magnitude” denotes a rule in which magnitude, rather than sign, determines the discrete code. The paper “SiMaN: Sign-to-Magnitude Network Binarization” replaces standard \(\{-1,+1\}\) sign binarization with a \(\{0,+1\}\) representation that keeps high-magnitude weights as \(+1\) and maps the rest to \(0\) [2102.07981]. The core optimization is
\[
\mathop{\arg\max}_{\bar{\mathbf{b}_w}} \frac{(\bar{\mathbf{b}_w})^T |\mathbf{w}|}{\|\bar{\mathbf{b}_w}\|_2 \, \||\mathbf{w}|\|_2},
\qquad \text{s.t.}\quad \bar{\mathbf{b}_w}\in\{0,+1\}^n,
\]
so the alignment target is \(|\mathbf w|\), not \(\mathbf w\).

The paper states that the discrete problem has an analytical global solution in \(\mathcal O(n\log n)\) time: sort the magnitudes, assign \(+1\) to the largest \(k\) entries, and assign \(0\) to the rest. The final deployment rule is the half-half version,
\[
\tilde{\mathbf{b}_w}= \begin{cases} 
+1, & \text{top half of sorted } |\mathbf{w}|,\\
0, & \text{otherwise}.
\end{cases}
\]
This is the paper’s explicit “positive magnitude” interpretation: binarization retains the strongest weights by absolute value, irrespective of their original sign [2102.07981].

A second contribution concerns the distribution of learned weights. The paper argues that BNN weights roughly follow a Laplacian distribution under standard training with \(\ell_2\) regularization, leading to a fraction of selected \(+1\) weights
\[
p_{+1}\approx 0.37.
\]
The reported empirical range is about \(36\%\sim 38\%\), contrasted with a Gaussian-based expectation of about \(0.54\). Removing \(\ell_2\) regularization shifts the \(+1\) fraction to around \(0.50\sim 0.52\), reduces the angle gap from about \(30.66^\circ\)–\(36.35^\circ\) to about \(7.97^\circ\)–\(8.05^\circ\), and lowers the complexity from \(\mathcal O(n\log n)\) to \(\mathcal O(n)\) because only the median of \(|\mathbf w|\) is needed [2102.07981].

The reported experiments place this rule in a performance-oriented context. On CIFAR-10, SiMaN reaches \(92.5\%\) top-1 on ResNet-18 and \(92.5\%\) on VGG-small. On ImageNet, it reports \(60.1\%\) top-1 and \(82.3\%\) top-5 on ResNet-18, and \(63.9\%\) top-1 and \(84.8\%\) top-5 on ResNet-34 [2102.07981]. A plausible implication is that, in this line of work, positivity is not a sign constraint but the occupancy of the \(+1\) state by large-magnitude coefficients.

## 3. Positive-instance recognition by feature magnitude

In weakly supervised video anomaly detection, “positive magnitude” refers to the use of large feature norms as evidence for positive instances, namely anomalous snippets inside abnormal videos. The paper “Weakly-supervised Video Anomaly Detection with Robust Temporal Feature Magnitude Learning” formulates the task as MIL, with each video represented as a bag of snippets and only the video-level label observed [2101.10030]. Because abnormal videos contain many normal snippets, the authors replace pure score-based instance selection with feature-magnitude selection.

The top-\(k\) score of a video is defined as
\[
g_{\theta,k}(\mathbf{X}) = \max_{\Omega_k(\mathbf{X}) \subseteq \{ \mathbf{x}_t\}_{t=1}^T} \frac{1}{k}\sum_{\mathbf{x}_t \in \Omega_k(\mathbf{X})} \|\mathbf{x}_t\|_2,
\]
where \(\Omega_k(\mathbf X)\) is the subset of \(k\) snippets with the largest \(\ell_2\)-norms. Separability between abnormal and normal videos is
\[
d_{\theta,k}(\mathbf{X}^+,\mathbf{X}^-) = g_{\theta,k}(\mathbf{X}^{+}) - g_{\theta,k}(\mathbf{X}^{-}),
\]
and the key statistical assumption is
\[
\mathbb{E}[\|\mathbf{x}^+\|_2] \ge \mathbb{E}[\|\mathbf{x}^-\|_2].
\]
Under this assumption, the paper proves that for \(0<k<\mu\), expected separability increases with \(k\), while for finite \(\mu\),
\[
\lim_{k\to\infty}\mathbb{E}[D_{\theta,k}(\mathbf{X}^+,\mathbf{X}^-)]=0,
\]
so overly large \(k\) values dilute the anomaly signal [2101.10030].

The temporal feature extractor combines Pyramid Dilated Convolutions with dilation rates \(1,2,4\) and Temporal Self-Attention, with the stated purpose of capturing both local and long-range dependencies. The classifier is then trained only on the top-\(k\) magnitude snippets. In this framework, “positive magnitude” means that larger \(\ell_2\)-norm features are treated as more likely to be positive instances.

The reported experiments are consistent with that interpretation. The paper gives best results of \(97.21\%\) AUC on ShanghaiTech, \(84.30\%\) AUC on UCF-Crime, \(77.81\%\) AP on XD-Violence, and \(98.6\%\) AUC on UCSD-Peds using I3D-RGB features. It also states that adding the feature-magnitude module raises AUC by over \(7\%\) on ShanghaiTech and over \(4\%\) on UCF-Crime, and reports mean top-\(k\) magnitude \(53.4\) for abnormal snippets versus \(7.7\) for normal snippets on UCF-Crime, with \(k=3\) in experiments [2101.10030]. This suggests a precise operational meaning: magnitude is used as a surrogate for latent positive-instance membership.

## 4. Positive earthquake magnitude differences

In seismology, the expression has a narrower and more literal meaning. The paper “A positive answer on the existence of correlations between positive earthquake magnitude differences” studies
\[
\delta m_i=m_{i+1}-m_i,
\]
and then restricts attention to the positive part,
\[
q_i=m_{i+1}-m_i>0,
\]
that is, cases in which the later earthquake is larger than the previous one [2404.15706]. The motivation comes from the observation by Van der Elst that, under appropriate conditions, positive magnitude differences are much less affected by missed small earthquakes than raw magnitudes or signed differences.

The paper analyzes relocated Southern California and Northern California catalogs. The Southern California catalog covers January 1981 to March 2022 and contains \(N=800499\) earthquakes with \(m\ge -1\); the Northern California catalog covers January 1984 to December 2021 and contains \(N=879547\) earthquakes with \(m\ge -0.8\). The completeness magnitudes are estimated as \(M_c=2.6\) for Southern California and \(M_c=2.7\) for Northern California [2404.15706].

To test correlation structure, the paper compares real catalogs with reshuffled catalogs and defines
\[
\Delta P_+(Q,Y)= P(\delta q_i>Q \vert \Delta y_i<Y)-P(\delta q^*_i> Q \vert \Delta y_i<Y).
\]
Under the null hypothesis of i.i.d. magnitudes following a Gutenberg-Richter law, the authors argue that \(\Delta P_+\) should be small and should vanish when positive differences are well above the detection threshold. Empirically, however, they find the opposite trend: as the threshold on positive differences increases, \(E[\Delta P_+(Q,Y)]\) increases rather than decreases, and for \(\delta M_{\text{th}}\ge 0.5\), above the estimated completeness threshold for differences \(\delta M_c\simeq 0.4\), the correlations remain strong [2404.15706].

The paper further reports similar results for both California catalogs, stronger correlations when events are closer in time, and only weak dependence on spatial separation once proximity is imposed. It interprets the findings as consistent with a time-dependent \(b\)-value in the Gutenberg-Richter law and suggests that ETAS-type models may need explicit magnitude-memory effects [2404.15706]. Here, “positive magnitude” is not a forecast label or a representation rule, but a catalog statistic chosen for robustness to incompleteness.

## 5. Magnitude as an invariant of positive definite metric spaces

A different tradition uses “magnitude” in Leinster’s sense: a numerical invariant of metric spaces. The paper “Positive definite metric spaces” develops the theory for spaces whose similarity kernel
\[
\zeta_A(x,y)=e^{-d(x,y)}
\]
is positive definite [1012.5863]. For a finite metric space \(A\), a weighting is a vector \(w\) satisfying \(\zeta_A w=\mathbf 1\), and the magnitude is
\[
|A|=\sum_{x\in A} w(x).
\]
If \(A\) is positive definite, then \(\zeta_A\) is invertible, the weighting is unique,
\[
w=\zeta_A^{-1}\mathbf 1,
\]
and the magnitude is positive [1012.5863].

The paper emphasizes the structural advantages of positive definiteness. For finite spaces, magnitude is always defined and monotone under inclusion: if \(\emptyset\neq B\subseteq A\), then \(0<|B|\le |A|\). For compact positive definite spaces, the proposed definitions of magnitude coincide, including the supremum over finite subspaces, the approximation-by-finite-subsets definition, the weight-measure definition, and the variational Rayleigh-quotient formula
\[
|A|=\sup\left\{\frac{\mu(A)^2}{Z_A(\mu,\mu)} \,\middle|\, \mu\in M(A),\ Z_A(\mu,\mu)\neq 0\right\},
\]
where
\[
Z_A(\mu,\nu)=\int_A\int_A e^{-d(x,y)}\,d\mu(x)\,d\nu(y).
\]
Magnitude is lower semicontinuous with respect to Gromov–Hausdorff distance on compact positive definite metric spaces, while maximum diversity is continuous [1012.5863].

This framework connects magnitude to negative type. The same paper proves that stable positive definiteness is equivalent to negative type, and lists many classes of examples, including \(L_p\) for \(0<p\le 2\), ultrametric spaces, round spheres, real and complex hyperbolic spaces, weighted trees, and spaces with at most four points [1012.5863]. In this domain, positivity is a property of the similarity kernel, and positive magnitude means that the invariant itself is forced to be well defined and positive by that kernel structure.

## 6. Geometric and continuity refinements of magnitude

Recent work gives geometric formulas for magnitude and sharp continuity conditions. The paper “Geometric interpretation of magnitude” studies a positive definite symmetric matrix \(Z\) with \(Z_{ii}=1\) and realizes it as a Gram matrix \(Z=V^tV\) of unit vectors \(v_1,\dots,v_n\in\mathbb R^n\). If \(R\) is the radius of the circumsphere of \(V_Z=\{v_1,\dots,v_n\}\), then
\[
\operatorname{Mag}(Z)=\frac{1}{1-R^2}.
\]
The same paper states that a positive weighting exists if and only if the center of the circumsphere lies in the interior of the convex hull of the \(v_i\). For an \(n\)-point metric space \(X\) of negative type, it proves
\[
\lim_{t\to 0}\operatorname{Mag}(Z_{tX})=1
\]
and
\[
\operatorname{Mag}(Z_X)<n \qquad (n>1),
\]
thereby giving a negative answer to the Gomi–Meckes problem [2510.26118].

The paper “Convergence of Magnitude of Finite Positive Definite Metric Spaces” reformulates the same invariant through a similarity embedding \(Y=\{y_1,\dots,y_k\}\subset\mathbb R^k\) with Gram matrix \(\zeta_X\), yielding
\[
|X|=\frac{1}{1-\rho_Y^2},
\]
where \(\rho_Y\) is the circumradius of \(Y\). It then classifies when magnitude converges under Gromov–Hausdorff convergence by a clustering type \(r=\langle r_1,\dots,r_m\rangle\). The main theorem states that \(\lim_{n\to\infty}|X_n|=|X|\) holds if and only if
\[
\|r\|_1\le 2
\]
for any \(k\), or
\[
\|r\|_1=3 \text{ and } k=1.
\]
Outside these cases, the paper constructs counterexamples in which Hausdorff convergence of similarity embeddings does not imply convergence of circumradii [2511.10331].

A complementary continuity program is developed in “Tractable Metric Spaces and the Continuity of Magnitude.” A tractable metric space is defined by three conditions: positive definiteness, the Heine–Borel property, and finite magnitude for every closed ball. In this setting, continuity of magnitude on compact subspaces is equivalent to uniform continuity on bounded subspaces, and on intervals \([a,b]\subset\mathbb R\) the magnitude map is Lipschitz with constant
\[
1+\frac{b-a}{2}.
\]
The paper concludes that magnitude is continuous on \(+(\mathbb R)\), with bounded restrictions Lipschitz [2506.21128].

For odd-dimensional Euclidean balls, potential-function methods give explicit determinant formulas. “On the magnitude of odd balls via potential functions” uses reverse Bessel polynomials and a Hankel linear system to derive a determinant expression for \(|B_R^{2p+1}|\), and reports conjectured Hankel determinant formulas for both the magnitude and its derivative [1804.02174]. This suggests that, within positive definite settings, magnitude often admits rigid geometric or algebraic normal forms.

## 7. Positivity programs adjacent to “positive magnitude”

Several adjacent literatures study positivity of magnitude-bearing quantities even when the phrase itself is not used in the same way. In operator theory, “Positivity of \(|p|^a|q|^b+|q|^b|p|^a\)” proves that
\[
J_{a,b,n}:=\frac12\bigl(|p|^a|q|^b+|q|^b|p|^a\bigr)\ge 0
\]
as a quadratic form on \(C_0^\infty(\mathbb R^n)\) under the conditions \(n>a+b\) and \(\min\{a,b\}\in[0,2]\). The same paper derives the sharper bound
\[
J_{a,b,n}\ge L_{a,b,n}|q|^{b-a},
\]
connecting the result to generalized Hardy inequalities for fractional Laplacians [1301.1524].

In EFT, “Higher-Point Positivity” studies shift-symmetric \(P(X)\) theories with
\[
\mathcal L=-\frac12 X+\sum_{i=2}^\infty \lambda_i X^i,
\qquad X=(\partial\phi)^2,
\]
and proves a parity-dependent sign rule for the first relevant higher operator:
\[
\lambda_n>0 \text{ for even } n,\qquad \lambda_n<0 \text{ for odd } n
\]
in mostly-plus signature. The result is derived from causality, analyticity of scattering amplitudes, and unitarity of the spectral representation [1804.03153].

In perturbative QCD, “On the positivity of MSbar distributions” shows that \(\overline{\mathrm{MS}}\) PDFs inherit positivity from physical PDFs above the perturbative bound
\[
Q^2\gtrsim 5~\mathrm{GeV}^2.
\]
The argument uses the perturbative scheme transformation
\[
f(x,Q^2)=\left[C(\alpha_s(Q^2))\right]^{-1}\otimes f^{\rm PHYS}(Q^2),
\]
and the same lower bound is stated to apply to longitudinally polarized PDFs [2405.08643].

In amplitude theory, “Positivity properties of scattering amplitudes” promotes ordinary non-negativity to complete monotonicity,
\[
(-\partial_x)^n f(x)\ge 0,
\]
and its multivariable generalizations. The paper argues that many planar and non-planar Feynman integrals with a Euclidean region, certain Euler integrals, and several observables in planar maximally supersymmetric Yang–Mills theory satisfy this stronger positivity hierarchy. It also states that the QCD and QED cusp anomalous dimensions exhibit the same property to three and four loops, respectively [2407.05755].

Taken together, these works indicate that the wider research ecosystem around “positive magnitude” is organized less by a single definition than by recurring positivity mechanisms: kernel positivity, magnitude ordering, positive differences, positive weightings, and complete monotonicity. A plausible implication is that “magnitude” functions as a bridge term between statistical amplitude, discrete salience, metric size, and formal positivity constraints, with each field fixing its own operational meaning.

Source: https://www.emergentmind.com/topics/positive-magnitude