CAPMix: Multifaceted Approaches Across Domains

• CAPMix is a polysemous term designating distinct methods: tail-risk allocation in finance, synthetic anomaly generation in time series detection, and capacitive mixing for blue-energy harvesting.
• In finance, CAPMix extends traditional CTE methods by incorporating higher-order tail central moments to reveal nuanced risk allocation and diversification effects.
• For time series anomaly detection, CAPMix integrates CutAddPaste, label revision, and dual-space mixup to generate diverse anomalies and boost detection performance.

CAPMix is a polysemous research term rather than a single established method. In recent arXiv literature, it denotes a tail-central-moment-based capital allocation framework for multivariate normal mean-variance mixture distributions, a controllable time series anomaly detection framework built around synthetic anomaly generation and dual-space mixup, and it also stands adjacent to the older electrochemical acronym CAPMIX, which denotes capacitive mixing for blue-energy harvesting and capacitive deionization in porous electrodes (Calderín-Ojeda et al., 2 Jan 2026, Mou et al., 8 Sep 2025, Härtel et al., 2014). Orthographic similarity has also led to potential confusion with CAP in communications, where it denotes carrier-less amplitude and phase modulation, and with CAP in modulation recognition, where it denotes capsule networks.

1. Terminological scope and disambiguation

The term is best understood as domain-specific shorthand whose meaning depends entirely on context. In finance and insurance, CAPMix refers to a capital allocation method based on tail central moments. In time series anomaly detection, CAPMix refers to a framework combining CutAddPaste anomaly synthesis, label revision, and dual-space mixup. In electrochemistry, the uppercase acronym CAPMIX refers to capacitive mixing in salinity-gradient energy harvesting and is historically distinct from the later CAPMix usages. In visible light communications and modulation recognition, the visually similar abbreviation CAP refers instead to carrier-less amplitude and phase modulation or capsule networks, respectively (Calderín-Ojeda et al., 2 Jan 2026, Mou et al., 8 Sep 2025, Härtel et al., 2014, Haigh et al., 2019, Snoap et al., 2023).

Term	Domain	Meaning
CAPMix	Finance and insurance	Tail-central-moment capital allocation
CAPMix	Time series anomaly detection	Anomaly augmentation with dual-space mixup
CAPMIX	Electrochemistry	Capacitive mixing / blue-energy harvesting

A common misconception is that CAPMix denotes a single transferable methodology across these literatures. The published record instead shows independent developments that share nomenclature but not mathematical machinery, datasets, or application objectives.

2. CAPMIX in electrochemical energy harvesting and desalination

In electrochemical literature, CAPMIX is a porous-electrode process for harvesting blue energy from salinity gradients, closely linked to capacitive deionization (CDI). The relevant model treats the electrode as a slit-pore capacitor of width $L$, with a symmetric 1:1 restricted primitive model electrolyte of hard-sphere ions of radius $R=0.34\ \text{nm}$, and formulates equilibrium through a grand potential functional whose excess part is decomposed as $F_{\rm exc}=F_{\rm C}+F_{\rm HS}+F_{\rm corr}$. The hard-sphere contribution uses the White-Bear mark II fundamental measure theory functional, and the residual electrostatic correction is MSA-inspired; the full theory is denoted FMT-PB-MSA (Härtel et al., 2014).

The physical point of the framework is that counter-ionic packing, Stern-layer exclusion, and confinement dominate the electric double layer in narrow pores, especially for $L=2$ nm micropores. Relative to PB, PB+S, and mPB models, FMT-PB-MSA predicts inhomogeneous density profiles even at $\Psi=0$, oscillatory layering near charged walls, coion exclusion at sufficiently high $\Psi$, a packing-limited surface charge of about $e\sigma_{\max}\approx 3.9\ e/\text{nm}^2$, and a more realistic decay of the differential capacitance $$C_{\rm diff}=\frac{\partial e\sigma}{\partial \Psi}$$ at large potentials. These steric and correlation effects are central to the concentration-dependent potential rise

$$\Delta\Psi := \Psi_{\rho_{\rm s}=0.02\rm M} - \Psi_{\rho_{\rm s}=0.5\rm M},$$

which governs CAPMIX work extraction (Härtel et al., 2014).

The work delivered by a cycle is

$W=-\oint \Psi\, dQ.$

Because charging occurs in high-salinity water and discharging in low-salinity water, the CAPMIX cycle is counterclockwise in the $R=0.34\ \text{nm}$0-$R=0.34\ \text{nm}$1 plane and yields positive work. The DFT treatment predicts a limiting potential difference of about 83 mV in PB and about 107 mV in FMT-type theories, implying a larger work output for blue-energy cycles than less elaborate mean-field models. The same packing physics increases the predicted energy demand for CDI, and when a concentration-dependent dielectric constant is included, the maximum work output is reduced and the optimal operating potential shifts to around $R=0.34\ \text{nm}$2 (Härtel et al., 2014).

3. CAPMix as tail-central-moment capital allocation

In finance and insurance, CAPMix is introduced as a capital allocation framework based on tail central moments (TCM) for the class of normal mean-variance mixture (NMVM) distributions. Its starting point is the observation that CTE-based allocation is coherent and widely used, but it captures only the conditional mean of the tail. CAPMix generalizes this by incorporating tail dispersion and higher-order tail shape. The paper defines the tail moment

$R=0.34\ \text{nm}$3

with $R=0.34\ \text{nm}$4, and the $R=0.34\ \text{nm}$5-th order tail central moment

$R=0.34\ \text{nm}$6

For $R=0.34\ \text{nm}$7, this reduces to the tail variance $R=0.34\ \text{nm}$8 (Calderín-Ojeda et al., 2 Jan 2026).

The modeling class is

$R=0.34\ \text{nm}$9

specialized to $F_{\rm exc}=F_{\rm C}+F_{\rm HS}+F_{\rm corr}$0, with $F_{\rm exc}=F_{\rm C}+F_{\rm HS}+F_{\rm corr}$1 and $F_{\rm exc}=F_{\rm C}+F_{\rm HS}+F_{\rm corr}$2. This structure is attractive because it generates heavy tails through the mixing variable $F_{\rm exc}=F_{\rm C}+F_{\rm HS}+F_{\rm corr}$3, skewness through $F_{\rm exc}=F_{\rm C}+F_{\rm HS}+F_{\rm corr}$4, dependence through $F_{\rm exc}=F_{\rm C}+F_{\rm HS}+F_{\rm corr}$5, and closure under linear combinations. A key special case is the generalized hyperbolic family obtained when $F_{\rm exc}=F_{\rm C}+F_{\rm HS}+F_{\rm corr}$6; the paper notes that GH includes the normal, skewed Student-$F_{\rm exc}=F_{\rm C}+F_{\rm HS}+F_{\rm corr}$7, variance gamma, and normal inverse Gaussian as special cases (Calderín-Ojeda et al., 2 Jan 2026).

The main theoretical contribution is a set of analytical and recursive expressions for tail moments, tail central moments, and the associated CAPMix allocations under NMVM assumptions. Conceptually, the allocation extends CTE-based decomposition by measuring each component’s contribution to centered tail fluctuations rather than to tail mean alone. The paper’s numerical analysis on daily log losses of Boeing, American Express, ExxonMobil, and Chevron over 2020–2024 shows that CTE, TV, and $F_{\rm exc}=F_{\rm C}+F_{\rm HS}+F_{\rm corr}$8 can produce materially different allocation patterns: BA and CVX remain relatively stable, AXP receives a larger share when moving from CTE to TV and again from TV to $F_{\rm exc}=F_{\rm C}+F_{\rm HS}+F_{\rm corr}$9, while XOM’s allocated proportion shrinks and can become negative under $L=2$0, indicating diversification effects that CTE does not reveal (Calderín-Ojeda et al., 2 Jan 2026).

4. CAPMix in time series anomaly detection

In time series anomaly detection, CAPMix is a framework for the anomaly-assumption setting, in which synthetic anomalies are injected into training data to compensate for scarce labels. The method is motivated by two stated failure modes of earlier approaches: patchy generation, where injected anomalies are overly simplistic or incoherent, and anomaly shift, where synthetic anomalies are either too close to normal data or unrealistically far from real anomalies. CAPMix addresses these issues with three components: CutAddPaste, label revision, and dual-space mixup within a temporal convolutional network (Mou et al., 8 Sep 2025).

The structural model used to motivate anomaly synthesis is

$L=2$1

where $L=2$2 captures shape structure, $L=2$3 seasonality, and $L=2$4 trend. CutAddPaste chooses a patch from a source sequence, adds a random linear trend term, and pastes the modified patch into a destination sequence. According to the paper, this single mechanism can generate shape anomalies, correlation anomalies, seasonality anomalies, trend anomalies, and point-wise anomalies. To control anomaly shift, CAPMix estimates the normal center

$L=2$5

and revises labels using the Dynamic Time Warping distance between each synthetic sample and $L=2$6, assigning a hard anomaly label when the sample is sufficiently far from the normal center and a softer anomaly label in a borderline region (Mou et al., 8 Sep 2025).

The classifier is a three-block TCN followed by an MLP, with mixup applied in both input and latent space. Input-space mixup is defined by

$L=2$7

and feature-space mixup is applied analogously at intermediate layers. Training uses binary cross-entropy,

$L=2$8

and inference thresholds the anomaly probability $L=2$9 to produce the anomaly score $\Psi=0$0 (Mou et al., 8 Sep 2025).

Evaluation is reported on AIOps, UCR, SWaT, WADI, and ESA, using Revised Point Adjusted (RPA) F1. CAPMix outperforms the CutAddPaste baseline on all five datasets, with reported scores of 80.46\% vs 77.44\% on AIOps, 71.89\% vs 69.98\% on UCR, 47.04\% vs 43.43\% on SWaT, 34.08\% vs 26.55\% on WADI, and 84.46\% vs 18.56\% on ESA. The ablation study further distinguishes the effects of label revision and mixup: label revision contributes more on univariate datasets, whereas mixup matters more on multivariate datasets (Mou et al., 8 Sep 2025).

5. Orthographically related but distinct CAP usages

A major source of confusion is the similarity between CAPMix and the acronym CAP in communications. In visible light communication, CAP means carrier-less amplitude and phase modulation, a passband-like modulation formed digitally from two orthogonal pulse-shaped branches rather than by explicit RF upconversion. Its multi-band extension, m-CAP, splits the transmission band into multiple subbands, each carrying its own QAM stream. A real-time FPGA-based VLC demonstrator using off-the-shelf LEDs reports transmission speeds up to approximately 30 Mb/s, with specific throughputs of 11.3 Mb/s for 4-QAM with 7% FEC, 22.61 Mb/s for 16-QAM with 7% FEC, and 33.91 Mb/s for 16-QAM with 20% FEC, sufficient for HDTV streaming (Haigh et al., 2019).

A related VLC paper proposes non-orthogonal multi-band CAP (NM-CAP), in which subcarrier spacing is compressed below the orthogonality limit. The carrier frequencies are parameterized by a compression factor $\Psi=0$1, and the spectral efficiency becomes

$\Psi=0$2

The reported result is up to 30% bandwidth saving and 44% improvement in the measured spectral efficiency with no further bit error rate performance degradation compared to traditional m-CAP, using the same transmitter and receiver functional blocks (Haigh et al., 2018).

In modulation recognition, CAP can instead mean capsule network. One line of work feeds capsule networks with cyclic cumulant features extracted by cyclostationary signal processing; another replaces generic front ends with custom nonlinear layers inspired by CSP. The custom layers include a squaring layer, a Pow3 layer, and an FFT layer, and the proposed CAP reports overall correct-classification probabilities of 89.8% on CSPB.ML.2018 and 86.7% on CSPB.ML.2022 in one study, while a broader dissertation reports a CC-trained CAP reaching 92.3%, 93.1%, 92.5%, and 91.6% across within-dataset and cross-dataset evaluations on CSPB.ML.2018 and CSPB.ML.2022 (Snoap et al., 2023, Snoap, 25 Mar 2025).

6. Comparative perspective

Across the cited literatures, CAPMix is not a family of mathematically homologous methods. In electrochemistry, CAPMIX is driven by electric-double-layer thermodynamics in nanoporous carbon electrodes and analyzed through DFT with FMT-PB-MSA. In finance, CAPMix is a higher-order tail-risk allocation framework for NMVM and GH models, built around $\Psi=0$3. In time series anomaly detection, CAPMix is a training-time augmentation and regularization pipeline centered on CutAddPaste, distance-based label revision, and dual-space mixup (Härtel et al., 2014, Calderín-Ojeda et al., 2 Jan 2026, Mou et al., 8 Sep 2025).

The overlap is therefore lexical rather than methodological. A plausible implication is that literature searches for “CAPMix” require explicit domain qualifiers, especially because the neighboring acronyms CAPMIX, CAP, m-CAP, and CAP as capsule network all have active and technically sophisticated literatures of their own. Without that disambiguation, electrochemical blue-energy harvesting, tail-risk allocation, TSAD, visible light communications, and modulation classification can be conflated despite having no shared problem formulation or implementation pipeline (Haigh et al., 2019, Snoap et al., 2023).

The most stable encyclopedic usage is thus contextual: CAPMix in finance refers to TCM-based capital allocation; CAPMix in TSAD refers to controllable anomaly augmentation with mixup; CAPMIX in electrochemistry refers to capacitive mixing; and CAP/m-CAP in communications refers to carrier-less amplitude and phase modulation.