Code-Mixing Guided Preference-Learning Framework

• The paper introduces a novel framework that leverages code-mixing to guide preference learning, enabling robust handling of variance across multi-scale data.
• It employs techniques inspired by variance cascade analysis to integrate disparate information and reduce uncertainty in model predictions.
• The framework shows promise in fields from turbulence modeling to reinforcement learning, offering practical strategies for variance reduction and efficient inference.

Searching arXiv for the cited papers and closely related work on variance cascades. Variance cascade refers, in the cited literature, to the transfer of a variance-like quadratic quantity across an ordered hierarchy of scales or stages. In turbulence this quantity is typically temperature variance, thermal variance, scalar variance, or surface potential energy, and the cascade may be direct, toward smaller scales, or inverse, toward larger scales. The same vocabulary is extended in other fields to volatility across resolutions in financial time series and to variance propagation across sequential slots or iterations in learning algorithms. This suggests a unifying usage: redistribution of fluctuations through a cascade structure, with fluxes, spectra, or recursive variance terms providing the diagnostic language (Valadão et al., 2024, Vieweg et al., 2022, Beck et al., 2024, Maskawa et al., 2020).

1. Variance as a cascaded invariant in geophysical turbulence

In the $\alpha$-turbulence family, the active scalar $\theta(x,t)$ is advected by a two-dimensional incompressible velocity $v=\nabla^\perp\psi$, with constitutive law

$$\psi = |\Delta|^{-\alpha/2}\theta, \qquad \widehat v(k)= i\,k^\perp\,k^{-\alpha}\,\widehat\theta(k).$$

For $\alpha=1$ this is the Surface Quasi-Geostrophic model, while $\alpha=2$ gives the 2D Navier-Stokes vorticity equation. In the inviscid, unforced limit these models conserve the generalized energy $E=\tfrac12\langle \psi\theta\rangle$ and the generalized enstrophy, or in SQG language the surface potential energy (variance), $P=\tfrac12\langle \theta^2\rangle$. When forcing injects variance at a scale $\ell_f$, there is a direct cascade of $P$ toward $\theta(x,t)$0 and, for $\theta(x,t)$1, an inverse cascade of $\theta(x,t)$2 toward $\theta(x,t)$3 (Valadão et al., 2024).

A rigorous SQG-oriented treatment formulates the cascade through localized fluxes and multi-scale averaging rather than through a spectral closure. In that framework, the governing equation is $\theta(x,t)$4, with $\theta(x,t)$5 determined by $\theta(x,t)$6. The key object is the localized variance flux into balls of radius $\theta(x,t)$7, and a sufficient condition for a direct variance cascade is expressed by the inequality $\theta(x,t)$8, where $\theta(x,t)$9 is a cascade-cutoff length built from macro-scale averages of $v=\nabla^\perp\psi$0, its extension $v=\nabla^\perp\psi$1, and the dissipation density $v=\nabla^\perp\psi$2. Under this condition, for every $v=\nabla^\perp\psi$3 with $v=\nabla^\perp\psi$4, the ensemble-averaged localized flux obeys

$v=\nabla^\perp\psi$5

which identifies an inertial range carrying a direct temperature variance cascade (Bradshaw et al., 2013).

The geophysical significance of this formulation is that the cascaded quantity is not merely a passive scalar norm. In SQG, the velocity is slaved to $v=\nabla^\perp\psi$6 by a non-local law, and the dissipation is fractional. The cascade therefore depends on non-local inertial transport and on the Caffarelli-Silvestre extension that restores a local positive dissipation density in the extended variable $v=\nabla^\perp\psi$7 (Bradshaw et al., 2013).

2. Non-equilibrium corrections in the direct SQG cascade

A recent development studies not only the mean cascade but the temporal fluctuations of the flux of surface potential energy. The shell-integrated variance spectrum $v=\nabla^\perp\psi$8 satisfies the exact spectral balance

$v=\nabla^\perp\psi$9

In the inertial range $$\psi = |\Delta|^{-\alpha/2}\theta, \qquad \widehat v(k)= i\,k^\perp\,k^{-\alpha}\,\widehat\theta(k).$$0, forcing and dissipation are neglected, so $$\psi = |\Delta|^{-\alpha/2}\theta, \qquad \widehat v(k)= i\,k^\perp\,k^{-\alpha}\,\widehat\theta(k).$$1. The central claim is that the instantaneous imbalance in the energy budget originates a subleading correction to the spectrum of the turbulent cascade (Valadão et al., 2024).

The analysis assumes a multiple-scale expansion

$$\psi = |\Delta|^{-\alpha/2}\theta, \qquad \widehat v(k)= i\,k^\perp\,k^{-\alpha}\,\widehat\theta(k).$$2

with slow temporal drift $$\psi = |\Delta|^{-\alpha/2}\theta, \qquad \widehat v(k)= i\,k^\perp\,k^{-\alpha}\,\widehat\theta(k).$$3, and closes the flux through the Kovasznay-type dimensional Ansatz

$$\psi = |\Delta|^{-\alpha/2}\theta, \qquad \widehat v(k)= i\,k^\perp\,k^{-\alpha}\,\widehat\theta(k).$$4

At leading order one obtains $$\psi = |\Delta|^{-\alpha/2}\theta, \qquad \widehat v(k)= i\,k^\perp\,k^{-\alpha}\,\widehat\theta(k).$$5, the instantaneous dissipation-injection rate, and the classical inertial-range law

$$\psi = |\Delta|^{-\alpha/2}\theta, \qquad \widehat v(k)= i\,k^\perp\,k^{-\alpha}\,\widehat\theta(k).$$6

At first order, assuming $$\psi = |\Delta|^{-\alpha/2}\theta, \qquad \widehat v(k)= i\,k^\perp\,k^{-\alpha}\,\widehat\theta(k).$$7, exponent matching yields

$$\psi = |\Delta|^{-\alpha/2}\theta, \qquad \widehat v(k)= i\,k^\perp\,k^{-\alpha}\,\widehat\theta(k).$$8

Because $$\psi = |\Delta|^{-\alpha/2}\theta, \qquad \widehat v(k)= i\,k^\perp\,k^{-\alpha}\,\widehat\theta(k).$$9, the correction is steeper and subdominant at high $\alpha=1$0 (Valadão et al., 2024).

For $\alpha=1$1, corresponding to SQG, the equilibrium exponent is $\alpha=1$2 and the non-equilibrium correction exponent is $\alpha=1$3. The time-averaged inertial range displays the $\alpha=1$4 law, while the instantaneous deviation exhibits a $\alpha=1$5 tail. As $\alpha=1$6, the prefactor $\alpha=1$7 diverges, so the perturbation theory breaks down and one recovers the well-known logarithmic corrections of the enstrophy cascade (Valadão et al., 2024).

The same work gives an operational decomposition of instantaneous spectra into equilibrium and non-equilibrium parts. Defining

$\alpha=1$8

one isolates the pure $\alpha=1$9 law in $\alpha=2$0 and the steeper correction in $\alpha=2$1. This disentangling procedure is explicitly proposed as generalizable to other turbulent systems (Valadão et al., 2024).

3. Linear stochastic realization and fractional-Gaussian scaling

A different realization of variance cascade is provided by a linear stochastic PDE designed to model the spatial structure of turbulent velocity fields. In Fourier space, the field $\alpha=2$2 evolves according to

$\alpha=2$3

with $\alpha=2$4 a real-valued, mean-zero Gaussian forcing that is $\alpha=2$5-correlated in time and space and supported in an annulus $\alpha=2$6. The cascade mechanism is the radial transport operator $\alpha=2$7, which in polar coordinates becomes $\alpha=2$8. In the inviscid, unforced limit, and ignoring the weaker $\alpha=2$9 term, the dynamics reduces to $E=\tfrac12\langle \psi\theta\rangle$0, so any packet of variance originally at scale $E=\tfrac12\langle \psi\theta\rangle$1 is carried to larger wavenumbers $E=\tfrac12\langle \psi\theta\rangle$2 (Beck et al., 2024).

Because the white-in-time forcing injects variance only into $E=\tfrac12\langle \psi\theta\rangle$3, transport carries that injected variance outward to all $E=\tfrac12\langle \psi\theta\rangle$4. In the stationary inviscid limit, and for $E=\tfrac12\langle \psi\theta\rangle$5, the mode-by-mode covariance obeys

$E=\tfrac12\langle \psi\theta\rangle$6

which is the classic power-law variance spectrum of a fractional Gaussian field of Hölder index $E=\tfrac12\langle \psi\theta\rangle$7. In physical space, the exact relation

$E=\tfrac12\langle \psi\theta\rangle$8

leads to $E=\tfrac12\langle \psi\theta\rangle$9 for $P=\tfrac12\langle \theta^2\rangle$0. In particular, for $P=\tfrac12\langle \theta^2\rangle$1 and $P=\tfrac12\langle \theta^2\rangle$2, one recovers $P=\tfrac12\langle \theta^2\rangle$3, the shell-integrated energy density $P=\tfrac12\langle \theta^2\rangle$4, and $P=\tfrac12\langle \theta^2\rangle$5 (Beck et al., 2024).

The numerical evidence is based on one-, two-, and three-dimensional finite-volume discretizations in $P=\tfrac12\langle \theta^2\rangle$6-space with a splitting integrator in time. Angle-averaged periodograms display the exact power-law slope $P=\tfrac12\langle \theta^2\rangle$7 over an increasingly wide inertial range at smaller $P=\tfrac12\langle \theta^2\rangle$8, and the structure functions transition from $P=\tfrac12\langle \theta^2\rangle$9 to $\ell_f$0 in the viscous subrange. In this model, the cascade of variance is analytically explicit: large-scale stochastic injection, deterministic outward transport in Fourier space, and viscous arrest at high $\ell_f$1 (Beck et al., 2024).

4. Inverse and coupled variance cascades in convection and electrokinetic turbulence

Variance cascades are not confined to direct small-scale transfer. In three-dimensional, horizontally extended turbulent Rayleigh-Bénard convection with imposed heat-flux boundary conditions, the total thermal variance is $\ell_f$2, with spectral density $\ell_f$3. The mode-to-mode transfer is

$\ell_f$4

and the cumulative planar flux among vertically homogeneous modes is denoted $\ell_f$5. A negative sign of $\ell_f$6 on some range indicates an inverse cascade, transferring variance to larger horizontal scales (Vieweg et al., 2022).

In the non-rotating case, at all Rayleigh numbers up to $\ell_f$7, one finds $\ell_f$8 for $\ell_f$9, and positive flux above that scale, so the dynamics contains both an inverse and a direct branch. The peak of the thermal spectrum $P$0 drifts toward the domain fundamental $P$1, and once the aggregation fills the domain the inverse cascade ceases. The negative transfer is dominated by planar triads with $P$2, while interactions involving nonzero $P$3 do not produce inverse transfer. Weak rotation arrests the inverse cascade at a finite scale $P$4, with the empirical law

$P$5

A parallel inverse cascade of kinetic energy persists even after the thermal variance saturates, sustaining the large $P$6 rolls that feed the planar thermal-variance cascade (Vieweg et al., 2022).

Electrokinetic turbulence provides a different coupled setting in which scalar variance and kinetic energy are linked by an externally forced conservation law. Under statistical stationarity, and in the subrange where electric body force dominates and dissipation is negligible, the forcing components satisfy $P$7, giving the universal conservation law

$P$8

This produces a quad-cascade picture with four asymptotic regimes. For the scalar-variance spectrum $P$9, the predicted exponents are $\theta(x,t)$00 when both $\theta(x,t)$01 and $\theta(x,t)$02 are constant, $\theta(x,t)$03 when $\theta(x,t)$04 is constant and $\theta(x,t)$05 is variable, $\theta(x,t)$06 when $\theta(x,t)$07 is constant and $\theta(x,t)$08 is variable, and $\theta(x,t)$09 when both are variable (Shi et al., 2023).

Microfluidic experiments support three of these regimes. At $\theta(x,t)$10, a clear $\theta(x,t)$11 range appears; between $\theta(x,t)$12–$\theta(x,t)$13, the spectrum shifts toward a $\theta(x,t)$14 range; and above $\theta(x,t)$15, a $\theta(x,t)$16 inertial subrange emerges. The cross-over wavenumber scales as $\theta(x,t)$17, in close agreement with the model prediction $\theta(x,t)$18 (Shi et al., 2023).

5. Cascade constructions beyond fluid turbulence

In financial modeling, a volatility cascade is formulated as a stochastic differential equation in the log-scale variable $\theta(x,t)$19, with $\theta(x,t)$20 as the volatility proxy: $\theta(x,t)$21 The multiplicative source governs the cascade, while the additive source is treated perturbatively through $\theta(x,t)$22, $\theta(x,t)$23, with $\theta(x,t)$24. The stationary Fokker-Planck density

$\theta(x,t)$25

has a Student-$\theta(x,t)$26 tail when $\theta(x,t)$27, and the $\theta(x,t)$28th moments scale as $\theta(x,t)$29 with

$\theta(x,t)$30

The non-quadratic dependence of $\theta(x,t)$31 is identified as the hallmark of multifractality, and the paper explicitly frames the mechanism as a variance-cascade analogy to turbulence (Maskawa et al., 2020).

In ranking-policy evaluation under the cascade behavior model, the cascade is sequential rather than spectral. A user examines slot $\theta(x,t)$32 only if no click occurred at earlier slots, with $\theta(x,t)$33 and $\theta(x,t)$34. The Cascade Doubly Robust estimator introduces slot-wise baselines $\theta(x,t)$35 into a recursive estimator so that the dominant variance term in the Recursive Importance-weighted Per-Decision Sampling recursion,

$\theta(x,t)$36

is replaced by

$\theta(x,t)$37

The cited analysis states that whenever $\theta(x,t)$38, the largest variance component at each slot is reduced. Under the cascade assumption alone, the estimator is unbiased for $\theta(x,t)$39 (Kiyohara et al., 2022).

A further terminological extension appears in model-free reinforcement learning. Variance-Reduced Cascade Q-learning combines direct variance reduction with a proposed Cascade Q-learning scheme, using the intertwined recursions

$\theta(x,t)$40

$\theta(x,t)$41

The stated role of the cascade step is to filter the noise in Bellman updates, and the combined algorithm is reported to be minimax optimal in the $\theta(x,t)$42-norm; when the action set is a singleton, it achieves non-asymptotic instance optimality while requiring the minimum number of samples theoretically possible (Boveiri et al., 2024). This indicates a distinct usage of “cascade,” now attached to recursive variance reduction rather than to transport of a conserved scalar.

6. Interpretation, limitations, and recurrent misconceptions

One recurrent misconception is that a variance cascade necessarily means a single direct $\theta(x,t)$43 law. The cited literature does not support that restriction. In SQG and more general $\theta(x,t)$44-turbulence, the classical direct-cascade spectrum is accompanied by a steeper non-equilibrium correction when the instantaneous flux fluctuates (Valadão et al., 2024). In horizontally extended convection, the salient diagnostic is not a universal power law but the sign of the planar flux $\theta(x,t)$45, and no simple power-law exponents were reported for $\theta(x,t)$46 (Vieweg et al., 2022). In electrokinetic turbulence, multiple exponents arise from coupled kinetic-energy and scalar-variance fluxes (Shi et al., 2023).

A second misconception is that stationarity implies an exactly constant instantaneous flux. The non-equilibrium SQG analysis states the opposite: real turbulent cascades display intermittent bursts of dissipation and forcing, so that $\theta(x,t)$47 fluctuates and $\theta(x,t)$48. The resulting correction is subleading but observable, and it can be extracted directly from instantaneous spectra (Valadão et al., 2024).

There are also explicit breakdown regimes. As $\theta(x,t)$49 in the $\theta(x,t)$50-turbulence calculation, the prefactor of the non-equilibrium correction diverges and the perturbation theory breaks down (Valadão et al., 2024). In the earlier SQG theorem, the triggering condition for the cascade is described as more exotic than standard phenomenology, reflecting the non-locality introduced by fractional dissipation (Bradshaw et al., 2013). In electrokinetic experiments, the highest-$\theta(x,t)$51 variable-flux subrange $\theta(x,t)$52 was not clearly observed, with the paper attributing this to the imaging noise floor and finite spatial resolution (Shi et al., 2023).

Taken together, these results show that “variance cascade” is best understood as a family of scale-transfer mechanisms rather than as a single exponent or single model class. In fluids, the central objects are fluxes, invariant-like quadratic forms, and inertial-range transfers; in finance, the same logic is translated into multiplicative scale dynamics and multifractal moments; in sequential decision problems, it becomes a language for recursive variance propagation and its control. The common mathematical theme is the organization of fluctuations by a cascade structure, while the physical or algorithmic meaning of the transported quantity depends on the domain (Beck et al., 2024, Maskawa et al., 2020, Kiyohara et al., 2022, Boveiri et al., 2024).