---
title: Scale-Adaptive Interpolation Schedules
url: https://www.emergentmind.com/topics/scale-adaptive-interpolation-schedules
type: topic
---

# Scale-Adaptive Interpolation Schedules

Searching arXiv for the cited papers to ground the article and verify identifiers.
Scale-adaptive interpolation schedules are adaptive procedures for selecting interpolation coordinates, collocation nodes, basis functions, or time/noise parameters across heterogeneous problem scales such as polynomial degrees, spatial regimes, spectral bands, or circuit depths. Across the literature, the unifying idea is to replace fixed interpolation schedules with data- or model-dependent refinement rules that allocate resolution where local error, score mismatch, spectral roughness, optimization benefit, or physical stiffness is largest. In the context of adaptive sparse polynomial chaos expansions, the notion appears as a dimension-adaptive, anisotropic schedule over multi-indices and Leja nodes [1911.08312]. Closely related constructions arise in QAOA parameter interpolation [2504.01694], molecular force prediction [2606.09480], adaptive tempering in filtering [2405.14408], adiabatic state preparation [2512.10329], diffusion and stochastic-interpolant generative models [2412.07877], [2509.02971], [2509.01629], [2602.03789], and adaptive interpolation for molecular potential energy surfaces [1605.09751]. This body of work suggests that “scale-adaptive interpolation schedules” are best understood not as a single algorithm, but as a design principle: schedules should be matched to the local geometry, sensitivity, and computational difficulty of the interpolation path.

## 1. Formal concept and recurring mathematical structure

A common formal pattern is the interpolation of a target object between two or more representations, together with an adaptive rule for deciding where to place the next interpolation point or how fast to traverse the path. In adaptive sparse polynomial chaos expansions, a scalar quantity of interest \(u\) depending on \(d\) independent random variables \(\xi=(\xi_1,\dots,\xi_d)\) is approximated by a finite polynomial chaos expansion
\[
u(\xi) \approx \sum_{\alpha\in\mathcal{I}} c_\alpha \psi_\alpha(\xi),
\]
with multivariate orthonormal polynomials \(\psi_\alpha\), and the schedule is the adaptive growth of the downward-closed multi-index set \(\mathcal{I}\) together with its associated Leja collocation nodes [1911.08312]. In that setting, each refinement step adds exactly one node and one polynomial term, and the resulting expansion is interpolatory at the selected nodes.

In QAOA, the interpolated object is a smooth parameter schedule over normalized layer index \(t=k/p\in[0,1]\), with
\[
\gamma(t)=\sum_{i=1}^{m} c_i^{(\gamma)} \phi_i(t), \qquad
\beta(t)=\sum_{i=1}^{m} c_i^{(\beta)} \phi_i(t),
\]
where the schedule is adaptive in both circuit depth \(p\) and basis dimension \(m\) [2504.01694]. In molecular force prediction, the interpolation coordinate is a scale variable \(\alpha_i\in[0,1]\) mediating between short-scale and long-range experts,
\[
\hat{F}_i = \alpha_i f_{\text{short}}(x_i) + (1-\alpha_i) f_{\text{long}}(x_i),
\]
with schedule refinement implemented through routing, differentiable updates, and scale-pool expansion [2606.09480].

In diffusion and stochastic interpolants, the interpolated object is typically a path of distributions. For denoising diffusion models, the schedule is a discretization of the diffusion path between \(p_1\) and \(p_0\), with adaptive time allocation derived from a local cost \(\mathcal{L}(t,t')\) or from spectral properties of the instance [2412.07877], [2603.19222]. In stochastic interpolants, schedules are the scalar functions \(\alpha_t,\beta_t\) in
\[
I_t = \alpha_t z + \beta_t x_1,
\]
and scale adaptivity is imposed by controlling the Lipschitzness of the drift or the spectral conditioning of the interpolation [2509.02971], [2509.01629], [2602.03789].

These formulations differ in application domain, but they share three structural components. First, there is an interpolation family parameterized by time, degree, scale, or index. Second, there is a criterion measuring local difficulty, importance, or error. Third, there is an update rule that redistributes computational effort toward the scales where that criterion is largest. This suggests that the essential content of a scale-adaptive interpolation schedule is the coupling of interpolation geometry with problem-dependent refinement indicators.

## 2. Adaptive sparse interpolation and polynomial chaos

The 2019 Leja-interpolation construction provides a canonical example of a scale-adaptive interpolation schedule in uncertainty quantification [1911.08312]. The method starts from stochastic collocation on Leja nodes and exploits a hierarchical Newton-like basis whose multivariate degrees are unique. The interpolating polynomial chaos expansion is obtained either through an explicit hierarchical basis transform or directly in the orthonormal basis, because each Leja node is in one-to-one correspondence with a basis polynomial degree.

The univariate Leja sequence is defined recursively by
\[
x_{n+1} = \arg\max_{x\in D} w(x)\prod_{k=0}^{n}|x-x_k|,
\]
with \(w(x)\) a positive weight linked to the orthogonality measure. For polynomial chaos with respect to \(\rho_i\), the paper states that a common and effective choice is \(w(x)\approx\sqrt{\rho_i(x)}\), and this empirically improves stability and adaptivity across distributions including uniform, Gaussian, truncated, and Gumbel cases [1911.08312]. Because Leja sequences are nested, new points append to the previous set rather than replacing it.

The adaptive schedule is organized over a downward-closed multi-index set \(\mathcal{I}\subset\mathbb{N}_0^d\), satisfying
\[
\text{if } \alpha\in\mathcal{I} \text{ and } \alpha_i>0 \text{ for some } i,\text{ then } \alpha-e_i\in\mathcal{I}.
\]
Downward-closedness ensures hierarchical consistency and sparse-grid structure. Refinement occurs on the admissible frontier using indicators that measure contribution at different scales. The data block lists two examples:
\[
\Delta_\alpha = |a_\alpha| \cdot \|N_\alpha\|,
\qquad
J_\alpha = c_\alpha^2.
\]
The former is surplus-based in the hierarchical Newton basis; the latter is the orthonormal polynomial variance contribution.

Anisotropy enters through weights \(\gamma=(\gamma_1,\dots,\gamma_d)\), with priority score
\[
\text{priority}(\alpha) \propto \Big[\prod_{i=1}^{d}\gamma_i^{\alpha_i}\Big]\cdot f(\Delta_\alpha),
\]
where \(f\) is monotone. The resulting schedule is therefore “scale-adaptive” in two senses: it resolves polynomial degree and physical dimension simultaneously, and it can preferentially refine influential dimensions. The procedure begins with \(\mathcal{I}=\{0\}\), constructs the admissible frontier, computes indicators, selects the index maximizing priority, appends the corresponding Leja node, maintains downward-closedness, and terminates when a global criterion such as frontier-indicator decay, simulation budget, or cross-validation stabilization is reached [1911.08312].

Two properties are central. First, each collocation point adds exactly one polynomial term, yielding “one polynomial term per collocation point.” Second, the resulting PCE is exact on the collocation nodes,
\[
I[u](\xi^{(m)}) = u(\xi^{(m)}) \quad \text{for all } m.
\]
This distinguishes the schedule from pseudo-spectral projection and regression-based collocation, which often require oversampling. A plausible implication is that scale-adaptive interpolation schedules are especially natural when the approximation architecture supports a one-to-one relation between refinement units and basis terms.

## 3. Representation-capacity schedules in optimization and control

A second major family of scale-adaptive interpolation schedules appears in optimization of parameterized trajectories. In QAOA, the schedule is the ordered set of \(2p\) parameters \(\{\gamma_1,\dots,\gamma_p,\beta_1,\dots,\beta_p\}\), but the iterative interpolation method replaces direct optimization over all \(2p\) angles by optimization over \(2m\) basis coefficients with \(m\ll p\) [2504.01694]. The construction treats the layer angles as smooth functions of \(t\in[0,1]\), expands them in orthogonal-function bases such as Chebyshev, Legendre, or Fourier, and discretizes by
\[
\gamma_k = \gamma(k/p), \qquad \beta_k = \beta(k/p).
\]

The adaptive schedule is not only over depth \(p\), but jointly over depth \(p\) and representation capacity \(m\). The algorithm begins from small \(p_0\) and \(m_0\), optimizes the \(2m\) coefficients, measures relative performance improvement
\[
\delta_{\text{perf}} = \frac{\text{Perf}_{\text{new}} - \text{Perf}_{\text{prev}}}{|\text{Perf}_{\text{prev}}|},
\]
and increases \(m\) only after \(\tau\) consecutive rounds with \(\delta_{\text{perf}}<\varepsilon\) [2504.01694]. Depth is increased each round, with warm-starting from the previous coefficients, and newly added basis modes are initialized at zero. The paper reports that this reaches schedules exceeding 1000 layers on LABS instances and yields better performance with fewer optimization steps than current approaches on SK, portfolio optimization, and LABS [2504.01694].

This schedule is “scale-adaptive” because it allocates optimization capacity in accordance with the observed smoothness of the optimal schedule. The paper notes that coefficient spectra decay rapidly, and that reconstructing a \(p=100\) schedule with only approximately the first 12 coefficients yields similar performance. This suggests that the adaptive mechanism is exploiting a low-dimensional manifold structure in high-depth schedules rather than merely compressing parameters.

Adiabatic state preparation supplies a related but physically distinct notion of schedule adaptivity. There the schedule is a strictly increasing function \(u(s)\in[0,1]\) in
\[
H(u(s)) = (1-u(s))H_0 + u(s)H_1,
\]
and the adaptive rule slows the evolution in small-gap regions and speeds it up elsewhere [2512.10329]. The paper introduces the power-law ODE
\[
u'(s) = c_p \Delta^p(u(s)), \qquad p\in(1,2),
\]
with
\[
c_p = \int_0^1 \Delta^{-p}(u)\,du.
\]
Under the spectral gap measure condition
\[
\mu(\{s\in[0,1]: \Delta(s)\le x\}) \le Cx,
\]
this improves the runtime scaling for constant accuracy from \(T=O(\Delta_*^{-2})\) under linear scheduling to \(T=O(\Delta_*^{-1})\) [2512.10329]. For linear or piecewise linear gaps, \(p=3/2\) satisfies the Euler–Lagrange optimality condition for the derived adiabatic-error functional. The paper also states that the linear schedule is never optimal unless the gap is constant.

Both QAOA and adiabatic scheduling therefore instantiate the same principle in different variables: refinement should follow the latent complexity of the path. In one case, the complexity is empirical schedule bandwidth; in the other, it is instantaneous spectral gap.

## 4. Routing, gating, and discrete scale-pool refinement

In multiscale representation learning, scale-adaptive interpolation schedules often appear as mixtures of experts with learnable interpolation weights. The molecular force-prediction framework on a NaCl aqueous ionic system is explicit about this interpretation [2606.09480]. It treats two predefined scales as initial anchors, \(S_0=\{0,1\}\), where \(a=0\) denotes the long-range expert and \(a=1\) denotes the short-scale expert. The prediction for atom \(i\) is
\[
\hat{F}_i = \alpha_i f_{\text{short}}(x_i) + (1-\alpha_i) f_{\text{long}}(x_i),
\]
or, in residual form,
\[
\hat{F}_i = f_{\text{long}}(x_i) + \alpha_i \big[f_{\text{short}}(x_i)-f_{\text{long}}(x_i)\big].
\]

The schedule variable \(\alpha_i\) is learned or optimized under the component-wise force MAE
\[
L_{\text{force}} = \frac{1}{3N}\sum_{i=1}^{N}\sum_{c\in\{x,y,z\}} |F_{i,c}-\hat{F}_{i,c}|.
\]
The paper derives the subgradient with respect to \(\alpha_i\),
\[
\frac{\partial L_{\text{force}}}{\partial \alpha_i}
= \frac{1}{3N}\sum_{c\in\{x,y,z\}}
\mathrm{sgn}(\hat{F}_{i,c}-F_{i,c})
\big[f_{\text{short},c}(x_i)-f_{\text{long},c}(x_i)\big],
\]
and uses this to update either gate parameters or the interpolation weights directly [2606.09480].

A distinctive contribution is scale-pool refinement. Starting from endpoint anchors \(\{0,1\}\), the method defines interpolated experts between anchor pairs,
\[
f_{ij}(x;\alpha)=\alpha f_{s_i}(x)+(1-\alpha)f_{s_j}(x),
\]
chooses \(\alpha_{ij}^\star\) minimizing the task loss, and inserts the resulting scale if the improvement \(\Delta L\) exceeds a threshold \(\varepsilon\) [2606.09480]. The reported update trajectory is
\[
\{0,1\} \to \{0,0.5,1\} \to \{0,0.25,0.5,1\} \to \{0,0.25,0.5,0.75,1\}
\]
followed by insertion of \(0.125\) and \(0.375\), yielding the final scale pool
\[
\{0,0.125,0.25,0.375,0.5,0.75,1\}.
\]
This final updated pool achieves an overall MAE of \(381.23\), compared with \(399.65\) for the long-only baseline, \(382.67\) for oracle hard routing, and \(380.96\) for continuous oracle interpolation [2606.09480]. In the close-contact regime with nearest-ion distance below \(0.6\) nm, the close-contact MAE decreases from \(327.22\) to \(260.51\) under continuous oracle interpolation, and to \(262.04\) under the final scale-pool update.

The main technical significance is that the scale schedule is discovered rather than fixed. Endpoint anchors dominate globally, but intermediate scales become important in difficult transitional regimes. The reported anchor-usage statistics quantify this: over all atoms, \(a=0\) accounts for \(48.41\%\), \(a=1\) for \(42.78\%\), and all intermediates for \(8.82\%\); in the close-contact regime, intermediates rise to \(29.02\%\), with \(a=0.75\) increasing to \(9.23\%\) [2606.09480]. This directly supports the claim that adaptive schedules are most valuable where regime boundaries are sharp but not binary.

## 5. Path discretization, tempering, and generative schedules

A large and mathematically diverse literature studies scale-adaptive interpolation schedules for paths of probability measures. In adaptive tempering for filtering, the canonical path is
\[
p_\beta(z) \propto p(z\mid y_{1:k-1})\, p(y_k\mid z)^\beta,\qquad \beta\in[0,1],
\]
and the schedule is the choice of intermediate tempering levels [2405.14408]. The paper introduces a finite schedule with \(T=2\) steps, \(\tau_0=0,\tau_1=\alpha,\tau_2=1\), so that the likelihood is split into \(\alpha\) and \(1-\alpha\) components, and different filters are applied to different substeps. Adaptation is driven either by effective sample size,
\[
\mathrm{ESS}(\beta)=\frac{(\sum_i w_i)^2}{\sum_i w_i^2},
\]
or by an interquartile-range criterion in observation space. If one chooses to adapt \(\Delta\beta\), the step is selected so that \(\mathrm{ESS}(\Delta\beta)\ge \kappa N_{\text{ens}}\), using bisection or line search [2405.14408]. The resulting schedule is explicitly scale-sensitive to likelihood curvature, dimensionality, prior spread, and observation mismatch.

Annealed importance sampling provides another path-based formulation. There, a geometric path
\[
\tilde p_t(x)=\tilde p_0(x)^{1-t}\tilde p_1(x)^t
\]
is paired with a constant-rate discretization schedule based on local divergence curvature [2306.15283]. For the geometric path, the local KL approximation yields
\[
\mathrm{KL}(p_t\Vert p_{t+\Delta t}) \approx \frac{(\Delta t)^2}{2}\,
\mathrm{Var}_{p_t}[\Delta(X)],
\qquad
\Delta(X)=\log \tilde p_1(X)-\log \tilde p_0(X),
\]
hence
\[
\Delta t \approx \sqrt{\frac{2\varepsilon}{\mathrm{Var}_{p_t}[\Delta(X)]}}.
\]
This is a direct example of scale-adaptive interpolation scheduling: smaller steps are allocated where the log-density difference has larger variance.

Diffusion models yield yet another version. In “Score-Optimal Diffusion Schedules,” the adaptive object is the discretization schedule along the reverse diffusion path, and the local cost is defined as
\[
\mathcal{L}(t,t') = v(t')^2 D(p_{t'} \| F_{t,t'}^\sharp p_t),
\]
with the corrector-optimized version
\[
\mathcal{L}_c(t,t') = v(t')^2 D(p_{t'}\|p_t).
\]
The paper proves a dense-limit geometry with local metric density \(\delta(t)\) and shows that the optimal schedule equalizes the cumulative length
\[
\Lambda(t)=\int_0^t \sqrt{\delta(u)}\,du,
\]
leading to the constant-speed geodesic allocation
\[
t_i^\star = \hat\Lambda^{-1}\!\left(\hat\Lambda\,\frac{i}{T}\right)
\]
after interpolation of empirical cumulative costs [2412.07877]. The method is hyperparameter-free in its core UpdateSchedule routine and uses only score evaluations. On CIFAR-10 with NFE 35 and Heun solver, the paper

Source: https://www.emergentmind.com/topics/scale-adaptive-interpolation-schedules