---
title: Random Multi-scale Sampling
url: https://www.emergentmind.com/topics/random-multi-scale-sampling
type: topic
---

# Random Multi-scale Sampling

Search arXiv for recent papers on multi-scale sampling and random sampling across domains.
Random multi-scale sampling denotes a family of stochastic sampling strategies in which randomization is used to probe, approximate, or couple structure across more than one spatial, temporal, spectral, geometric, or size scale. The phrase does not designate a single canonical algorithm. In the current literature it covers several distinct constructions: random local multi-sampling around a query point, random hierarchical subsampling across encoder stages, random scale or scale-offset selection during training, multilevel stochastic coupling across discretizations, and random sketching maps that compare objects of different sizes. Some methods are direct instances of random multi-scale sampling, whereas others are more accurately described as deterministic multiscale pipelines with stochastic components or as random sampling plus hierarchical feature aggregation [1901.07124], [2406.08924], [2607.07680].

## 1. Terminology and boundary cases

The term is used heterogeneously, and the main technical distinction is whether randomness acts on the **scale mechanism itself** or only inside a multiscale architecture. A concise taxonomy is useful.

| Pattern | Representative work | Characterization |
|---|---|---|
| Random local multi-sampling | [1901.07124] | Random auxiliary samples around each query point; effective support becomes scale-aware through the transformation |
| Random hierarchical subsampling | [2107.02389] | Random point removal across encoder levels, with multiscale feature aggregation compensating for discarded samples |
| Deterministic multiscale fusion | [2401.02614] | Scale pyramid plus fixed masking schedule; multiscale but not random |
| Exact multilevel stochastic coupling | [1703.08498] | Fine/coarse Gaussian-field samples coupled through projected white noise |
| Random scale curricula | [2406.08924] | Random patch scales and random scale offsets used explicitly during training |

Two recurrent misconceptions follow from this ambiguity. First, not every method with several scales is random. In image and video quality assessment, SAMA builds a pyramid \(p_x=\{x_0,x_1,\cdots,x_n\}\), samples fragments at each scale, and fuses them by a fixed mask
\[
\tilde{x}=M\cdot \hat{x}_0 + (1-M)\cdot \hat{x}_1,
\]
so its multiscale mechanism is structured and deterministic rather than stochastic [2401.02614]. Second, not every random sampler is explicitly multiscale. RandLA-Net uses random point sampling in every encoder layer, but its multiscale character arises from hierarchical receptive-field growth rather than from explicit multi-radius or pyramid sampling [2107.02389].

A further terminological boundary is negative rather than classificatory. The record titled "Accelerating Diffusion Decoders via Multi-Scale Sampling and One-Step Distillation" [2603.19570] does not actually contain a method section, a decoder, a sampling algorithm, or experiments on multi-scale sampling. It therefore does not provide evidence for a method called random multi-scale sampling.

## 2. Canonical stochastic mechanisms in imaging

A direct and technically explicit instance appears in differentiable image transformation. Linearized multi-sampling replaces bilinear interpolation’s reliance on direct neighbors by random auxiliary samples around each queried output pixel. Starting from
\[
\tilde{I}\left(\tilde{x}\right) = \sum_{x}I\left(T_{\theta}\left(x\right)\right) \: K\left(\tilde{x},T_{\theta}\left(x\right)\right),
\]
the method introduces a local first-order model
\[
\hat{I}\left(x\right) = I\left(T_{\theta}\left(x_i\right)\right) + A_i \left( T_{\theta}\left(x\right) - T_{\theta}\left(x_i\right) \right),
\]
with auxiliary samples
\[
x_i^{k} \sim N\left(x_i, \sigma\right),\;\; \forall k \in \left\{1,2,...,K\right\},
\]
and a regularized least-squares estimate
\[
A_i = \left(X_i^\top X_i + \epsilon E\right)^{-1}X_i^\top Y_i.
\]
The default experimental choice is \(K=8\), and \(\sigma\) is set to match the pixel width and height of the sample output. The method is random rather than deterministic; its scale robustness is indirect, because perturbations are drawn in output coordinates and then warped by \(T_\theta\), so downsampling automatically enlarges the effective support in the input image. The paper explicitly contrasts this with a deterministic multi-scale baseline using three scale levels with Gaussian kernels of standard deviations \(\{1,5,10\}\), and reports better GTSRB errors, including ICSTN + Ours \(4.85, 4.68, 4.86, 6.10\) at \(1\times,2\times,4\times,8\times\) downsampling [1901.07124].

A second imaging pattern is randomized candidate evaluation rather than randomized scale traversal. Monte Carlo Non-Local Means uses Bernoulli indicators
\[
\Pr[I_j=1]=p_j,\qquad \Pr[I_j=0]=1-p_j
\]
to evaluate only a random subset of patch similarities, then forms
\[
A(\mathbf p) = \frac{1}{n}\sum_{j=1}^n \frac{x_j w_j}{p_j} I_j, \qquad B(\mathbf p) = \frac{1}{n}\sum_{j=1}^n \frac{w_j}{p_j} I_j,
\]
and the self-normalized estimator
\[
Z(\mathbf p) = \frac{A(\mathbf p)}{B(\mathbf p)}.
\]
For uniform sampling, the paper proves
\[
\Pr\left[|Z(\mathbf p)-z|>\varepsilon\right] \le e^{-n\xi} + 2\exp\{-n\mu_B f(\varepsilon)\xi\},
\]
and reports that, with an external database containing ten billion patches, MCNLM returns a randomized solution within \(0.2\) dB of full NLM while reducing runtime by three orders of magnitude. Although the method is not itself multiscale, this suggests a general estimator design for scale-indexed candidate sets: randomized thinning with inverse-probability correction and concentration control [1312.7366].

## 3. Hierarchical subsampling, cropping, and scale fusion

In point-cloud learning, RandLA-Net is best characterized as **random hierarchical subsampling plus multi-scale feature aggregation**. Each of the four encoding layers applies local feature aggregation and then keeps only \(25\%\) of the points, so the hierarchy evolves as
\[
\widetilde{N} \rightarrow \frac{\widetilde N}{4} \rightarrow \frac{\widetilde N}{16} \rightarrow \frac{\widetilde N}{64} \rightarrow \frac{\widetilde N}{256}.
\]
Neighborhoods are built by \(K\)-nearest neighbors with \(K=16\). The local feature aggregation module combines local spatial encoding, attentive pooling, and a dilated residual block; after two neighborhood-aggregation units, a point can indirectly access approximately \(K^2\) points. The sampling step itself is uniform random selection with complexity \(\mathcal{O}(M)\), implemented with `numpy.random.choice()`. The paper stresses that this is not PointNet++-style explicit multi-scale sampling at each layer, but hierarchical subsampling whose effective support expands with depth. On SemanticKITTI validation, the same backbone reaches \(57.1\) mIoU with random sampling, compared with \(56.5\) for FPS and \(54.0\) for PDS, while remaining dramatically cheaper [2107.02389].

In image augmentation, CropMix is a direct random multi-scale sampling method. An image \(\mathbf X\) is cropped multiple times by operators \(\mathcal C=\{c_n\}_{n=1}^N\), each restricted to a distinct, non-overlapping scale interval, and the cropped views are then mixed. For an overall crop-scale range \((0.01,1.0)\) and \(N=3\), the assigned intervals are
\[
(0.01,0.34),\quad (0.34,0.67),\quad (0.67,1.0).
\]
On ImageNet classification, the number of cropping operations is randomly chosen from
\[
N\in\{2,3,4\}.
\]
This preserves label identity because all views come from the same source image. The paper reports that on 10% ImageNet with ResNet-50 and overall crop-scale range \((0.01,1)\), standard RRC gives \(52.57\) top-1 accuracy, whereas CropMix with random \(N\in\{2,3,4\}\) gives \(55.15\); it also shows that multi-scale CropMix is consistently better than single-scale CropMix at the same \(N\) [2205.15955].

By contrast, SAMA is explicitly presented as a deterministic alternative to random multi-scale sampling. It first constructs a scale pyramid, applies fragment sampling on every scale, and then interlaces the results with structured masks aligned to Swin or Video Swin patchization. On VQA trained on LSVQ, the reproduced baseline improves from SRCC/PLCC \(0.870/0.871\) to \(0.883/0.884\); on IQA, the baseline Swin-based model improves from \(0.923/0.934\) to \(0.930/0.942\) on KonIQ. The significance is conceptual as much as empirical: multiscale information can be preserved by deterministic scale interlacing rather than random scale choice [2401.02614].

## 4. Generative models and stochastic scale-space learning

A particularly clear random multi-scale training strategy appears in continuous scale-space generative modeling. The scale-space generator \(G(\mathbf z,\mathbf x_p,s_p)\) renders a fixed-resolution patch at arbitrary position and scale through
\[
g(\mathbf{x}_i; \mathbf{x}_p, s_p) = 2^{-s_p} \left( \mathbf{x}_i - \mathbf{x}_p \right).
\]
The paper states that simple uniform random sampling of patches across all scales does not converge satisfactorily. It therefore introduces two explicit scale-sampling mechanisms: a progressive adversarial sampling curriculum over scale bins—negative exponential early, uniform באמצע, and linearly increasing later—and a non-uniform random sampling rule for scale offsets in the consistency loss,
\[
p(\Delta s_p \mid s_p) = \frac {B\!\left(\frac{\Delta s_p}{s_p}; \alpha,\beta\right)} {s_p},
\]
with parameters chosen so that the mode is at \(\Delta s_p=1\). On the Milkyway pseudo-reconstruction task, Beta sampling gives FID \(9.00\), EMD \(2.08\), and PSNR\(_\text{inter}\) \(25.7\), compared with Uniform\(_\text{full}\) at FID \(9.5\), EMD \(2.77\), and PSNR\(_\text{inter}\) \(25.1\) [2406.08924].

Wavelet Multi-scale Generative Modeling illustrates a different category: deterministic coarse-to-fine factorization with stochastic components at each scale. The model writes
\[
p(\mathbf{x}^0) = \prod_{k=1}^S p(\mathbf{x}_H^k|\mathbf{x}_L^k) \, p(\mathbf{x}_L^S),
\]
uses reverse diffusion only on the coarsest low-frequency band, and samples high-frequency coefficients by
\[
x_H^k = G(x_L^k, z^k), \qquad z^k\sim \mathcal N(0,I),
\]
followed by
\[
x_L^{k-1} = A^T (x_H^k, x_L^k)^T.
\]
The scale order is fixed and deterministic; randomness enters through diffusion noise at the coarsest scale and through per-scale latent noise for high-frequency synthesis. At \(128\times128\), the paper reports 16 total sampling steps and sampling time \(2.31\) s for 100 images, versus \(15.093\) s for SGM and \(11.097\) s for WSGM [2411.09356].

A broader stochastic multiscale formulation is given by slow-fast score estimation for unnormalized targets. There the “scales” are dynamical rather than spatial: a slow variable follows a diffusion path while a fast ergodic process estimates the conditional expectations needed for the score. The generic system is
\[
\begin{cases}
d X_t = b(t, X_t, Y_t)\, d t + \sqrt{2}\, d B_t,\\
d Y_t = \frac{1}{\varepsilon} f(t, X_t, Y_t)\,d t + \sqrt{\frac{2}{\varepsilon}}\, d\tilde{B}_t,
\end{cases}
\]
with convergence to the averaged drift as \(\varepsilon\to0\). This is multiscale in the SDE sense rather than in the scale-space or pyramid sense, but it still exemplifies random multiscale sampling as stochastic averaging over coupled fast and slow processes [2508.15069].

## 5. PDEs, random fields, and inverse problems

In uncertainty propagation for PDEs, random multi-scale sampling acquires a literal coarse/fine meaning. A mixed finite-element SPDE sampler for Matérn-type Gaussian fields constructs a hierarchy of agglomerated meshes and solves, at each level,
\[
A_h u_h = -g\kappa^{-2} B_h^T f_h(\omega), \qquad 
\theta_h = \kappa^{-2} W_h^{-1}\big(B_h u_h + g f_h(\omega)\big),
\]
with white noise discretized exactly through the diagonal \(L^2\) mass matrix on piecewise constants:
\[
f_h = W_h^{1/2}\,\xi_h(\omega), \qquad \xi_h(\omega)\sim \mathcal N(0,I).
\]
The multilevel decomposition is
\[
\theta_\ell(\omega)=P_\theta \theta_{\ell+1}(\omega)+\delta\theta_\ell(\omega),
\]
and the projected coarse noise satisfies
\[
\xi_{\ell+1}(\omega) = W_{\ell+1}^{-1/2}P_\theta^T W_\ell^{1/2}\xi_\ell(\omega),
\qquad \operatorname{cov}(\xi_{\ell+1})=I.
\]
This yields exactly distribution-consistent coarse samples coupled to fine samples through shared underlying white noise. The paper reports near-linear weak scaling in 2D from 8 to 512 MPI processes, with fine-level stochastic dimension ranging from 3.7 million to 235 million, and similar behavior in 3D from 48 to 3072 MPI processes [1703.08498].

A related but distinct construction appears in randomized GMsFEM oversampling. Instead of solving one local problem for every fine-grid boundary degree of freedom on \(\partial\omega_i^+\), the method solves harmonic extensions with i.i.d. Gaussian random boundary traces:
\[
-\operatorname{div}(\kappa \nabla \psi_{l,\omega_i}^{+,\text{rsnap}})=0 \quad \text{in }\omega_i^+, \qquad
\psi_{l,\omega_i}^{+,\text{rsnap}}=r_l \quad \text{on }\partial\omega_i^+,
\]
then performs the usual local spectral compression. Oversampling is essential: with 20 local basis functions and \(p_{\text{bf}}=4\), the energy error drops from \(23.26\%\) at \(t=0\) to \(15.63\%\) at \(t=2\), and remains near \(15.24\%\) at \(t=7\). The randomized snapshot space is written as
\[
\Psi_{\omega_i}^{\text{rsnap}}=\mathcal{R}\Psi_{\omega_i}^{\text{snap}},
\]
making the relation to randomized range finding explicit [1409.7114].

The same random-local-solve principle is generalized for multiscale PDEs by sampling boundary conditions on subdomains and recovering low-rank local solution spaces automatically. For each patch \(K_m\), instead of enumerating all discrete boundary deltas, the method solves
\[
\begin{cases}
\mathcal{L}^\varepsilon r_{m,n}=0,\quad x\in K_m,\\
r_{m,n}=\omega_{m,n},\quad x\in \Gamma_m,
\end{cases}
\]
or on a buffered patch \(\widetilde K_m\), then couples the reduced local spaces globally. The central claim is that many homogenizable multiscale PDEs have numerically low-rank local Green’s maps, so random boundary excitations can recover the dominant local solution manifold without deriving the homogenized operator explicitly [1807.08848].

In Bayesian inverse modeling, “multiscale sampling” becomes a localized MCMC strategy. The permeability field is represented by a KLE,
\[
Y_k(x)=\sum_{i=1}^{N}\sqrt{\lambda_i}\,\theta_i\,\phi_i(x),
\]
then decomposed into subdomain-associated stochastic blocks \(\theta=(\theta^1,\ldots,\theta^{M_c})\). Local proposals use
\[
\theta_p^i = \sqrt{1-\beta^2}\,\theta^i + \beta\,\varepsilon^i,
\]
inside blocked Gibbs-style updates, embedded in delayed-acceptance MCMC with
\[
a_c(\eta,\eta_p)= \min\left( 1,\, \frac{I(\eta\mid \eta_p)P_c(\eta_p\mid R_p)}
     {I(\eta_p\mid \eta)P_c(\eta\mid R_p)} \right),
\]
followed by
\[
a_f(\eta,\eta_p)= \min\left( 1,\, \frac{P_f(\eta_p\mid R_p)P_c(\eta\mid R_p)}
     {P_f(\eta\mid R_p)P_c(\eta_p\mid R_p)} \right).
\]
The paper reports that sparse conditioning points further improve convergence: in one example, tail PSRF/MPSRF improves from approximately \(2.5/3.5\) without conditioning to approximately \(1.09/1.17\) with conditioning [2302.11149].

## 6. Abstract frameworks, matrices, and graphs

A general mathematical framework for random multi-scale sampling across variable-size objects is provided by any-dimensional learning by sampling. The central object is a family of random maps \(S_k\) that send any input, regardless of its original size, to a random \(k\)-dimensional summary. Objects are compared through the sampling metric
\[
d_{\mathrm{samp}}(x,y)=\sum_{k\ge 1} 2^{-k} W_1(S_k(x),S_k(y)),
\]
so agreement means that all finite sampled views are close in distribution. The paper distinguishes three admissible mechanisms—sampling with replacement, random binning, and species sampling—and ties each to a different cross-size symmetry: duplication for replacement sampling, and zero-padding for binning and species sampling. It then proves generalization and sketching bounds of the form
\[
\mathrm e_\infty(f,g)\le \mathrm e_n(f,g)+R_n,
\]
with sharper rates for classes that factor through low-dimensional sampled laws or moments [2607.07680].

A more algorithmic and explicitly multiscale construction arises in matrix heavy-hitter detection. There the cost of querying a row of an unknown right-stochastic matrix depends on precision \(\epsilon\), so querying all samples at the finest scale leads to \(\Theta((n/\Delta)^2)\) cost. The remedy is a precision schedule
\[
\epsilon_t = \frac1h \left\lceil \frac{t}{\tau}\right\rceil,
\]
combined with random row selection. In the vector precursor, the estimator
\[
\frac{n}{h}\sum_{t=1}^h \mathbf I[p_{s_t}\ge \epsilon_t]
\]
replaces fine-precision value estimation by threshold exceedance counts across many scales, exploiting the layer-cake identity behind \(\mathbb E[p_s]\). The full matrix algorithm aggregates sparse thresholded row approximations and achieves \(\tilde O(n/\Delta)\) cost, which the paper contrasts with traditional \(\Theta((n/\Delta)^2)\) sampling; this multiscale precision randomization is then used to detect significant PageRanks via the personalized PageRank matrix [1202.2771].

Graph sampling provides a useful boundary case. Frontier Sampling uses \(m\) dependent random walkers and is equivalent to a single random walk on the Cartesian power \(G^m\). Its stationary distribution over frontier states is
\[
P[L_\infty=(v_1,\ldots,v_m)] = \frac{\sum_{i=1}^m \deg(v_i)}
{m|V|^{m-1}\operatorname{vol}(V)},
\]
and the method improves finite-budget robustness on disconnected or loosely connected graphs. However, it is not a multiscale sampler in the usual pyramid, coarse-to-fine, or multi-resolution sense. It is more accurately a randomized multi-walker sampler whose relevance to random multi-scale sampling is partial rather than direct [1002.1751].

Across these literatures, the unifying principle is not a shared implementation but a shared decomposition of complexity. Random multi-scale sampling replaces exhaustive uniform evaluation by stochastic allocation across scales, levels, neighborhoods, or object sizes. The main design questions recur with remarkable regularity: which scales should be sampled randomly, how should samples be coupled across scales, what invariances define “the same object” at different resolutions, and whether randomness acts on scale selection itself or only within a fixed multiscale architecture. The literature surveyed here shows that those choices control not only computational cost, but also gradient quality, variance reduction, cross-level consistency, and out-of-scale generalization.

Source: https://www.emergentmind.com/topics/random-multi-scale-sampling