Multi-Scale Random Fourier Features

• Multi-Scale Random Fourier Features are techniques that combine multiple frequency scales to better capture heterogeneous data patterns and enhance kernel approximation.
• They employ scale mixtures, variable-specific bandwidths, and data-adaptive frequency learning to overcome the limitations of single-scale random features.
• Applications span forecasting fast-slow systems, multi-kernel learning, and quantum or deep learning settings, improving computational efficiency and representational richness.

Multi-Scale Random Fourier Features are Random Fourier Feature (RFF) constructions in which the spectral sampling mechanism is no longer tied to a single kernel bandwidth or a single fixed frequency law, but instead incorporates multiple scales, learned frequency bands, or scale mixtures in order to represent heterogeneous structure in data. In the standard RFF setting, a shift-invariant kernel is approximated by mapping inputs through random projections and periodic nonlinearities so that inner products in feature space approximate the kernel (Langrené et al., 2024). The multi-scale perspective generalizes this idea in several directions: by representing isotropic kernels as scale mixtures of $\alpha$-stable random vectors (Langrené et al., 2024), by assigning variable-specific bandwidths in reservoir computing for fast-slow dynamical systems (Laha, 4 Nov 2025), by learning useful frequency bands directly in the Fourier domain (Băzăvan et al., 2012), and by extending random-feature methodology to tensorized, operator-valued, sequential, and quantum settings (Wesel et al., 2021, Brault et al., 2016, Toth et al., 2023, Sakurai et al., 29 Jan 2026).

1. Spectral principle and the meaning of “multi-scale”

Random Fourier Features arise from the Fourier representation of translation-invariant kernels. For a continuous, shift-invariant, isotropic positive definite kernel $K(\mathbf{u})$ on $\mathbb{R}^d$, Bochner’s theorem gives

$$K(\mathbf{u}) = \int_{\mathbb{R}^d} \exp(i \mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x},$$

and, when $f$ is symmetric,

$$K(\mathbf{u}) = \int_{\mathbb{R}^d} \cos(\mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x} = \mathbb{E}[\cos(\bm{\eta}^\top \mathbf{u})].$$

RFF approximates this expectation by Monte Carlo sampling from the spectral distribution (Langrené et al., 2024).

In the standard Gaussian-kernel case, the feature map is written as

$$\phi(\boldsymbol{x}) = \sqrt{\frac{2}{D}} \left[\cos(\omega_1 \cdot \boldsymbol{x} + b_1), \dots, \cos(\omega_D \cdot \boldsymbol{x} + b_D)\right]^\top,$$

with $\omega_j \sim p(\omega)$ and $b_j \sim \mathrm{Unif}[0,2\pi]$, so that

$$k(\boldsymbol{x},\boldsymbol{x}') \approx \phi(\boldsymbol{x})^\top \phi(\boldsymbol{x}').$$

This construction is efficient, but in its simplest form it reflects a single spectral scale or a fixed parametric frequency law (Sakurai et al., 29 Jan 2026).

Within this framework, “multi-scale” refers to feature maps whose frequencies encode more than one characteristic scale. The supplied literature realizes this in three principal ways. First, the spectral law itself can be a scale mixture, so that a single feature generator samples across multiple effective bandwidths (Langrené et al., 2024). Second, different variables or feature groups can be assigned different bandwidth parameters, as in multi-scale RFF reservoirs for fast-slow systems (Laha, 4 Nov 2025). Third, the scale or frequency-band parameters can be optimized from data or selected among groups by group Lasso, which amounts to learning which bands are useful for prediction (Băzăvan et al., 2012).

A recurrent theme across these works is that single-scale constructions can be insufficient when the target signal contains heterogeneous frequency content. One paper states this explicitly for forecasting: a single kernel bandwidth cannot balance the need for high-frequency features for fast variables with smoothness for slow variables, leading to information loss or overfitting (Laha, 4 Nov 2025). Another frames the issue spectrally: changing the kernel bandwidth in the Fourier domain effectively selects different frequency bands or scales, with small $K(\mathbf{u})$0 sampling higher frequencies and large $K(\mathbf{u})$1 sampling lower frequencies (Băzăvan et al., 2012).

2. Scale mixtures for isotropic kernels

A general theoretical route to multi-scale RFF is given by the spectral mixture representation of isotropic kernels. The paper “A spectral mixture representation of isotropic kernels to generalize random Fourier features” shows that the spectral distribution of positive definite isotropic kernels in $K(\mathbf{u})$2 for all $K(\mathbf{u})$3 can be decomposed as a scale mixture of $K(\mathbf{u})$4-stable random vectors, with the mixing distribution identified as a function of the kernel (Langrené et al., 2024).

The key construction introduces a symmetric $K(\mathbf{u})$5-stable random vector $K(\mathbf{u})$6 satisfying

$K(\mathbf{u})$7

a nonnegative scalar random variable $K(\mathbf{u})$8, and a scale parameter $K(\mathbf{u})$9. The random projection is then

$\mathbb{R}^d$0

and the resulting kernel is

$\mathbb{R}^d$1

where $\mathbb{R}^d$2 is the characteristic function of $\mathbb{R}^d$3 (Langrené et al., 2024).

This result turns the spectral law into a random-scale mechanism. The paper states that the distribution for $\mathbb{R}^d$4 encodes the mixing over scales; different choices of $\mathbb{R}^d$5 yield different degrees of heavy-tailedness or smoothness in the constructed kernel, in a manner that is transparent and computationally easy to simulate (Langrené et al., 2024). In that sense, the multi-scale property is not an external heuristic but an intrinsic spectral parameterization.

The same paper provides explicit kernel families covered by this construction, including exponential power kernels, generalized Cauchy kernels, Matérn and generalized Matérn kernels, rational quadratic or Student kernels, power kernels, and newly introduced Beta, Kummer, and Tricomi kernels (Langrené et al., 2024). It also states the special cases

• Gaussian kernel: $\mathbb{R}^d$6, $\mathbb{R}^d$7, $\mathbb{R}^d$8;
• Laplace kernel: $\mathbb{R}^d$9, $$K(\mathbf{u}) = \int_{\mathbb{R}^d} \exp(i \mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x},$$0, $$K(\mathbf{u}) = \int_{\mathbb{R}^d} \exp(i \mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x},$$1;
• Matérn kernel: $$K(\mathbf{u}) = \int_{\mathbb{R}^d} \exp(i \mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x},$$2 with inverse-Gamma mixing (Langrené et al., 2024).

The simulation recipe is also explicit. One samples $$K(\mathbf{u}) = \int_{\mathbb{R}^d} \exp(i \mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x},$$3, samples $$K(\mathbf{u}) = \int_{\mathbb{R}^d} \exp(i \mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x},$$4, and forms

$$K(\mathbf{u}) = \int_{\mathbb{R}^d} \exp(i \mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x},$$5

where $$K(\mathbf{u}) = \int_{\mathbb{R}^d} \exp(i \mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x},$$6 and $$K(\mathbf{u}) = \int_{\mathbb{R}^d} \exp(i \mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x},$$7 is a positive random variable constructed from independent uniforms via the formula given in the paper (Langrené et al., 2024). The same source notes that the approach is immediately adaptable to existing RFF codebases based on Gaussian RFF: one simply replaces standard normal draws with draws from these multi-scale mixtures (Langrené et al., 2024).

A plausible implication is that multi-scale RFF, in this formulation, converts kernel design into a problem of choosing or fitting a mixing law on scales. That implication is stated directly in the paper as an applicability claim: the approach reduces kernel design or learning to choosing or fitting a suitable mixing distribution $$K(\mathbf{u}) = \int_{\mathbb{R}^d} \exp(i \mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x},$$8 on scales (Langrené et al., 2024).

3. Learning scales and selecting frequency bands

A second meaning of multi-scale RFF is data-adaptive scale selection. “Learning Random Kernel Approximations for Object Recognition” develops a Fourier-domain optimization view in which kernel hyperparameters controlling the spectral measure are learned from data (Băzăvan et al., 2012). The paper’s premise is that different kernel-induced Fourier sampling distributions correspond to different kernel parameters, and that optimization in the Fourier domain can identify the different frequency bands that are useful for prediction on training data (Băzăvan et al., 2012).

For kernel ridge regression, the linear model in random-feature space is

$$K(\mathbf{u}) = \int_{\mathbb{R}^d} \exp(i \mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x},$$9

with solution

$f$0

The kernel parameter $f$1 is then learned by minimizing validation loss plus regularization (Băzăvan et al., 2012).

The key technical device is a reparameterized sampling rule

$f$2

where $f$3 is a quantile function and $f$4 are fixed uniform random samples. Because the Monte Carlo randomness is fixed while $f$5 scales the frequencies, the validation objective becomes differentiable with respect to the kernel parameters, enabling efficient gradient-based optimization of frequency scales (Băzăvan et al., 2012).

The same source gives the operational interpretation of scale. Small $f$6 samples higher frequencies and emphasizes fine detail, whereas large $f$7 samples lower frequencies and emphasizes coarse structure (Băzăvan et al., 2012). Multi-scale RFF in this setting therefore means either continuously optimizing a scale parameter or assembling multiple random-feature groups corresponding to multiple kernels and multiple scales.

That second strategy is realized through group Lasso. If $f$8 concatenates the RFF embeddings of several kernels, then the optimization

$f$9

selects among groups of kernel features (Băzăvan et al., 2012). The paper states that applying group Lasso to concatenated RFFs is equivalent to classical multiple kernel learning, while avoiding full Gram-matrix computation and scaling linearly with the number of examples rather than quadratically for standard MKL (Băzăvan et al., 2012).

This line of work addresses a common misconception: that RFF necessarily means fixed, non-adaptive Monte Carlo features. In the supplied literature, standard RFF uses a fixed sampling distribution, but Fourier-domain optimization and grouped feature selection allow the sampling law or the active feature groups to reflect the data (Băzăvan et al., 2012). That distinction is central to multi-scale random-feature methods.

4. Variable-specific bandwidths in reservoir computing

A concrete application of multi-scale RFF appears in forecasting fast-slow dynamical systems. “Reservoir Computing via Multi-Scale Random Fourier Features for Forecasting Fast-Slow Dynamical Systems” introduces a reservoir computing framework that combines delay embedding with RFF mappings, and compares a single-scale reservoir using a fixed kernel bandwidth to a multi-scale reservoir integrating multiple bandwidths (Laha, 4 Nov 2025).

Given a $$K(\mathbf{u}) = \int_{\mathbb{R}^d} \cos(\mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x} = \mathbb{E}[\cos(\bm{\eta}^\top \mathbf{u})].$$0-dimensional time series $$K(\mathbf{u}) = \int_{\mathbb{R}^d} \cos(\mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x} = \mathbb{E}[\cos(\bm{\eta}^\top \mathbf{u})].$$1, the delayed predictor is built from

$$K(\mathbf{u}) = \int_{\mathbb{R}^d} \cos(\mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x} = \mathbb{E}[\cos(\bm{\eta}^\top \mathbf{u})].$$2

and

$$K(\mathbf{u}) = \int_{\mathbb{R}^d} \cos(\mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x} = \mathbb{E}[\cos(\bm{\eta}^\top \mathbf{u})].$$3

In the single-scale case, the feature map is

$$K(\mathbf{u}) = \int_{\mathbb{R}^d} \cos(\mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x} = \mathbb{E}[\cos(\bm{\eta}^\top \mathbf{u})].$$4

with $$K(\mathbf{u}) = \int_{\mathbb{R}^d} \cos(\mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x} = \mathbb{E}[\cos(\bm{\eta}^\top \mathbf{u})].$$5 and a single global $$K(\mathbf{u}) = \int_{\mathbb{R}^d} \cos(\mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x} = \mathbb{E}[\cos(\bm{\eta}^\top \mathbf{u})].$$6 (Laha, 4 Nov 2025).

In the multi-scale case, each variable $$K(\mathbf{u}) = \int_{\mathbb{R}^d} \cos(\mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x} = \mathbb{E}[\cos(\bm{\eta}^\top \mathbf{u})].$$7 has its own bandwidth $$K(\mathbf{u}) = \int_{\mathbb{R}^d} \cos(\mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x} = \mathbb{E}[\cos(\bm{\eta}^\top \mathbf{u})].$$8 and feature count $$K(\mathbf{u}) = \int_{\mathbb{R}^d} \cos(\mathbf{x}^\top \mathbf{u}) f(\mathbf{x})\,d\mathbf{x} = \mathbb{E}[\cos(\bm{\eta}^\top \mathbf{u})].$$9. The per-variable map is

$$\phi(\boldsymbol{x}) = \sqrt{\frac{2}{D}} \left[\cos(\omega_1 \cdot \boldsymbol{x} + b_1), \dots, \cos(\omega_D \cdot \boldsymbol{x} + b_D)\right]^\top,$$0

with $$\phi(\boldsymbol{x}) = \sqrt{\frac{2}{D}} \left[\cos(\omega_1 \cdot \boldsymbol{x} + b_1), \dots, \cos(\omega_D \cdot \boldsymbol{x} + b_D)\right]^\top,$$1, and the full reservoir feature is the concatenation

$$\phi(\boldsymbol{x}) = \sqrt{\frac{2}{D}} \left[\cos(\omega_1 \cdot \boldsymbol{x} + b_1), \dots, \cos(\omega_D \cdot \boldsymbol{x} + b_D)\right]^\top,$$2

The key design principle is explicit: narrow bandwidths, meaning small $$\phi(\boldsymbol{x}) = \sqrt{\frac{2}{D}} \left[\cos(\omega_1 \cdot \boldsymbol{x} + b_1), \dots, \cos(\omega_D \cdot \boldsymbol{x} + b_D)\right]^\top,$$3, are assigned to fast variables to resolve sharp transients, while wider bandwidths, meaning large $$\phi(\boldsymbol{x}) = \sqrt{\frac{2}{D}} \left[\cos(\omega_1 \cdot \boldsymbol{x} + b_1), \dots, \cos(\omega_D \cdot \boldsymbol{x} + b_D)\right]^\top,$$4, are assigned to slow variables to capture low-frequency modulations (Laha, 4 Nov 2025).

The readout is multivariate ridge regression,

$$\phi(\boldsymbol{x}) = \sqrt{\frac{2}{D}} \left[\cos(\omega_1 \cdot \boldsymbol{x} + b_1), \dots, \cos(\omega_D \cdot \boldsymbol{x} + b_D)\right]^\top,$$5

and evaluation uses the Normalized Root Mean Squared Error,

$$\phi(\boldsymbol{x}) = \sqrt{\frac{2}{D}} \left[\cos(\omega_1 \cdot \boldsymbol{x} + b_1), \dots, \cos(\omega_D \cdot \boldsymbol{x} + b_D)\right]^\top,$$6

(Laha, 4 Nov 2025).

The paper evaluates six canonical fast-slow systems: the Rulkov map, Morris-Lecar model, Hindmarsh-Rose model, Izhikevich model, Ricker map with seasonal forcing, and a predator-prey model (Laha, 4 Nov 2025). Across all cases, the multi-scale RFF reservoir consistently outperforms its single-scale counterpart, achieving lower NRMSE and more robust long-horizon predictions (Laha, 4 Nov 2025). The detailed results include, for example, one-step NRMSE for Rulkov $$\phi(\boldsymbol{x}) = \sqrt{\frac{2}{D}} \left[\cos(\omega_1 \cdot \boldsymbol{x} + b_1), \dots, \cos(\omega_D \cdot \boldsymbol{x} + b_D)\right]^\top,$$7 improving from $$\phi(\boldsymbol{x}) = \sqrt{\frac{2}{D}} \left[\cos(\omega_1 \cdot \boldsymbol{x} + b_1), \dots, \cos(\omega_D \cdot \boldsymbol{x} + b_D)\right]^\top,$$8 to $$\phi(\boldsymbol{x}) = \sqrt{\frac{2}{D}} \left[\cos(\omega_1 \cdot \boldsymbol{x} + b_1), \dots, \cos(\omega_D \cdot \boldsymbol{x} + b_D)\right]^\top,$$9, and for Izhikevich $\omega_j \sim p(\omega)$0 from $\omega_j \sim p(\omega)$1 to $\omega_j \sim p(\omega)$2; in closed-loop prediction, the single-scale model is reported as diverging quickly for Izhikevich, whereas the multi-scale model is stable for the full test interval (Laha, 4 Nov 2025).

The significance claimed by the paper is not merely empirical. It argues that, rather than relying on implicit temporal memory as in standard ESNs or leaky reservoirs, multi-scale RFF-mapped reservoirs project delay vectors into a feature space where different time frequencies are naturally resolved (Laha, 4 Nov 2025). This makes multi-scale RFF an explicit mechanism for aligning spectral resolution with known fast-slow structure.

5. High-dimensional, structured, and sequential generalizations

Multi-scale RFF interacts with several extensions of random-feature methodology beyond the scalar Euclidean setting. These works do not all use the phrase “multi-scale” in the same way, but they expand the representational scope of Fourier-feature approximations and clarify where scale heterogeneity enters.

“Large-Scale Learning with Fourier Features and Tensor Decompositions” focuses on deterministic rather than random Fourier features, but it is directly relevant to multi-scale constructions because deterministic tensor-grid features can converge exponentially in the number of basis functions while suffering from the curse of dimensionality (Wesel et al., 2021). The feature map is tensorized,

$\omega_j \sim p(\omega)$3

and the weight tensor is represented by a low-rank CP decomposition,

$\omega_j \sim p(\omega)$4

The resulting block coordinate descent algorithm has total complexity

$\omega_j \sim p(\omega)$5

which the paper describes as linear in both sample size $\omega_j \sim p(\omega)$6 and input dimension $\omega_j \sim p(\omega)$7 under small $\omega_j \sim p(\omega)$8 and $\omega_j \sim p(\omega)$9 (Wesel et al., 2021). This work is not a multi-scale RFF method in the strict Monte Carlo sense, but it addresses an adjacent issue: how to retain rich spectral approximations when the dimension is large.

“Random Fourier Features for Operator-Valued Kernels” generalizes Bochner’s theorem to translation-invariant operator-valued Mercer kernels and constructs matrix-valued random features (Brault et al., 2016). The approximated kernel is

$b_j \sim \mathrm{Unif}[0,2\pi]$0

and the feature map stacks blocks involving $b_j \sim \mathrm{Unif}[0,2\pi]$1 where $b_j \sim \mathrm{Unif}[0,2\pi]$2 (Brault et al., 2016). The paper proves uniform convergence for both bounded and unbounded operator random Fourier features using Bernstein matrix concentration inequalities (Brault et al., 2016). A plausible implication is that multi-scale spectral parameterizations could be combined with operator-valued kernels by placing mixtures or learned parameterizations in $b_j \sim \mathrm{Unif}[0,2\pi]$3 or $b_j \sim \mathrm{Unif}[0,2\pi]$4, but the supplied text presents that as a route rather than as a developed result (Brault et al., 2016).

“Random Fourier Signature Features” transfers RFF methodology to signature kernels for sequential data, where the base computational bottleneck is quadratic in sequence length and number of sequences (Toth et al., 2023). The truncated signature kernel is

$b_j \sim \mathrm{Unif}[0,2\pi]$5

and the proposed Random Fourier Signature Feature construction yields an unbiased estimator with linear dependence on dataset size and sequence length (Toth et al., 2023). The diagonal-projection and tensor-random-projection variants further reduce the complexity to

$b_j \sim \mathrm{Unif}[0,2\pi]$6

respectively, while retaining favourable concentration properties (Toth et al., 2023). Here the notion of scale is coupled to signature level and tensor degree rather than only to bandwidth, but the method still uses Fourier randomization to capture multi-level structure in sequences.

These extensions show that the multi-scale agenda in random features is not restricted to classical scalar kernels. It appears whenever the target representation must cover heterogeneous spectral, tensorial, or output-structured behavior while remaining computationally tractable.

6. Deep and quantum analogues

Two additional lines of work place multi-scale Fourier features in architectures that are not standard kernel approximators.

“Deep Learning without Global Optimization by Random Fourier Neural Networks” introduces residual networks with random complex exponential activation functions

$b_j \sim \mathrm{Unif}[0,2\pi]$7

trained block-wise by an adaptive MCMC procedure rather than global gradient-based optimization (Davis et al., 2024). The frequency sampling in each block is adapted to the residual through distributions proportional to the magnitude of the Fourier transform of the target or residual: $b_j \sim \mathrm{Unif}[0,2\pi]$8 The paper states that this procedure sidesteps spectral bias because, at each stage, the MCMC samples the frequencies where most of the residual energy is, enabling efficient learning of multiscale and high-frequency features (Davis et al., 2024). It also reports that Gibbs phenomena are not observed in approximating discontinuous target functions (Davis et al., 2024). Although this setting uses complex exponential activations in a deep residual network rather than a kernel machine, it is an explicit example of adaptive multi-scale Fourier feature allocation.

A quantum analogue is provided by “Quantum Random Features: A Spectral Framework for Quantum Machine Learning” (Sakurai et al., 29 Jan 2026). The paper introduces Quantum Random Features (QRF) and Quantum Dynamical Random Features (QDRF), lightweight quantum reservoir models inspired by classical RFF that generate high-dimensional spectral representations without variational optimization (Sakurai et al., 29 Jan 2026). Inputs are encoded by $b_j \sim \mathrm{Unif}[0,2\pi]$9-rotations,

$$k(\boldsymbol{x},\boldsymbol{x}') \approx \phi(\boldsymbol{x})^\top \phi(\boldsymbol{x}').$$0

interleaved either with random permutation matrices or with Ising-type Hamiltonian evolution

$$k(\boldsymbol{x},\boldsymbol{x}') \approx \phi(\boldsymbol{x})^\top \phi(\boldsymbol{x}').$$1

and followed by Hadamard gates and measurement (Sakurai et al., 29 Jan 2026). The resulting probability distribution over bitstrings defines a $$k(\boldsymbol{x},\boldsymbol{x}') \approx \phi(\boldsymbol{x})^\top \phi(\boldsymbol{x}').$$2-dimensional quantum feature map.

The paper states that, as the number of encoding layers $$k(\boldsymbol{x},\boldsymbol{x}') \approx \phi(\boldsymbol{x})^\top \phi(\boldsymbol{x}').$$3 increases, the spectrum of random frequencies induced by QRF and QDRF approaches that of RFF, with the effective transformation generating a weight matrix $$k(\boldsymbol{x},\boldsymbol{x}') \approx \phi(\boldsymbol{x})^\top \phi(\boldsymbol{x}').$$4 whose row vectors approximate i.i.d. normal variables as $$k(\boldsymbol{x},\boldsymbol{x}') \approx \phi(\boldsymbol{x})^\top \phi(\boldsymbol{x}').$$5 grow (Sakurai et al., 29 Jan 2026). It also states that randomization of the bias is necessary to prevent probability condensation into one computational basis state, with scaling

$$k(\boldsymbol{x},\boldsymbol{x}') \approx \phi(\boldsymbol{x})^\top \phi(\boldsymbol{x}').$$6

(Sakurai et al., 29 Jan 2026). On Fashion-MNIST, the models reach up to $$k(\boldsymbol{x},\boldsymbol{x}') \approx \phi(\boldsymbol{x})^\top \phi(\boldsymbol{x}').$$7 accuracy, matching or surpassing classical baselines, while using scalable qubit requirements (Sakurai et al., 29 Jan 2026).

The quantum setting makes explicit a broader point also visible in the classical literature: multi-scale random features are fundamentally spectral constructions. Whether realized by Gaussian scale mixtures, variable-specific bandwidths, adaptive frequency learning, or entangling quantum dynamics, the core objective is to produce diverse frequency coverage without full kernel-matrix computation or deep variational optimization.

7. Limitations, trade-offs, and recurring themes

Across the supplied literature, the main advantage of multi-scale RFF is more faithful representation of heterogeneous structure. In forecasting, variable-specific bandwidths improve both one-step accuracy and long-horizon robustness on systems with intrinsic fast-slow interactions (Laha, 4 Nov 2025). In isotropic-kernel design, scale mixtures broaden the class of kernels for which explicit RFF sampling is available, including generalized Cauchy, generalized Matérn, Beta, Kummer, and Tricomi kernels (Langrené et al., 2024). In Fourier-domain learning, optimization or group selection identifies the frequency bands most useful for prediction (Băzăvan et al., 2012).

The main trade-offs are also clear. Standard random features retain Monte Carlo error, and one paper contrasts this with deterministic Fourier features whose approximation error can decrease exponentially in the number of basis functions but whose tensor-product extension suffers heavily from the curse of dimensionality (Wesel et al., 2021). In sequential and operator-valued settings, extra structure requires tensor projections, matrix-valued features, or stronger concentration analysis (Toth et al., 2023, Brault et al., 2016). In the quantum case, there is measurement or sampling overhead, and deep entangling circuits or long dynamical evolutions can pose noise and decoherence challenges (Sakurai et al., 29 Jan 2026).

A compact comparison of the main multi-scale mechanisms in the supplied papers is as follows.

Setting	Multi-scale mechanism	Paper
Isotropic kernels	Scale mixture of $$k(\boldsymbol{x},\boldsymbol{x}') \approx \phi(\boldsymbol{x})^\top \phi(\boldsymbol{x}').$$8-stable random vectors via $$k(\boldsymbol{x},\boldsymbol{x}') \approx \phi(\boldsymbol{x})^\top \phi(\boldsymbol{x}').$$9	(Langrené et al., 2024)
Kernel learning / MKL	Optimize $K(\mathbf{u})$00 in Fourier domain or select RFF groups by group Lasso	(Băzăvan et al., 2012)
Reservoir computing	Variable-specific bandwidths $K(\mathbf{u})$01 for fast and slow variables	(Laha, 4 Nov 2025)
Deep residual networks	Residual-adaptive frequency sampling by MCMC	(Davis et al., 2024)
Quantum random features	Layered scrambling induces RFF-like spectral behavior	(Sakurai et al., 29 Jan 2026)

One recurring misconception is that “more features” alone solve the approximation problem. The supplied papers repeatedly tie performance not just to feature count but to spectral allocation: the law from which frequencies are drawn, the scales attached to variables, the grouping of kernels, or the residual-adapted distribution from which frequencies are sampled (Băzăvan et al., 2012, Laha, 4 Nov 2025, Davis et al., 2024). A plausible implication is that multi-scale random features are best understood not as a single algorithm, but as a design principle for spectral approximation under heterogeneity.

In that broader sense, Multi-Scale Random Fourier Features denote a family of methods that preserve the computational advantages of random-feature approximations while replacing fixed, single-scale spectral sampling by richer constructions that encode multiple bandwidths, multiple temporal scales, learned bands, or structured multi-level frequency representations.