---
title: Undecimated Discrete Curvelet Coefficients
url: https://www.emergentmind.com/topics/undecimated-discrete-curvelet-coefficients
type: topic
---

# Undecimated Discrete Curvelet Coefficients

Undecimated discrete curvelet coefficients (UDCT) are a multiscale, orientation-sensitive representation of fields over discretized domains, constructed via an FFT-based transform that preserves shift-invariance by omitting spatial subsampling. In the context of fractured media, as introduced in Segura’s convex SPDE inversion framework, UDCT analysis and synthesis operators play a central role in regularizing reconstructions to capture anisotropic, fracture-aligned structures that elude isotropic priors [2601.17485].

## 1. Mathematical Definition of the Undecimated Discrete Curvelet Transform

On each piecewise-planar fracture support, the underlying scalar field is embedded on an $n \times n$ uniform grid ($M = n^2$). The UDCT analysis operator is
\[
C: \mathbb{R}^M \longrightarrow \mathbb{C}^N, \quad z \mapsto d = C z,
\]
where typically $N \approx 8M$, reflecting the transform's redundancy. The transform is constructed with reference to radial scale windows $\{ W_j(r) \}_{j=0}^{J}$ and angular windows $\{ V_{j,\ell}(\theta) \}_{\ell=0}^{L_j-1}$, with $j$ denoting the scale and $\ell$ the orientation. The generating curvelet in continuous frequency is
\[
\varphi_{j,\ell,m}(u,v) =
\frac{1}{(2\pi)^2} \int_{\mathbb{R}^2}
W_j(\|\omega\|)\, V_{j,\ell}(\arg \omega)\,
e^{i\,\omega \cdot [(u,v)-m]} \, d\omega,
\]
where $(u,v)$ enumerates the spatial grid and $m$ the offset. In the undecimated version, there is no subsampling after frequency masking; all locations are processed. For admissibility, each $L_j$ is a multiple of 3, and windows are smoothly overlapped and summed, following the constructions of Candès et al. (2006).

## 2. Enforcement of Curvelet-Sparsity via $\ell_1$ Penalty

To encourage representation of fracture-aligned ridges and edges, the framework introduces a Laplace prior on the curvelet coefficients:
\[
\lambda\,\|C\,z\|_1 = \lambda\sum_{j=0}^J \sum_{\ell=0}^{L_j-1} \sum_{m=0}^{n^2-1} |d_{j,\ell}[m]|
\]
with $d_{j,\ell}[m] = \langle z, \varphi_{j,\ell,m} \rangle$. The parabolic scaling and directional selectivity of $\varphi_{j,\ell,m}$ ensure that the $\ell_1$ penalty drives most coefficients toward zero except at the multiscale ridges and edges, enforcing sparsity preferentially along dominant fracture directions.

## 3. Algorithmic Computation of Forward and Inverse UDCT

Implementation comprises:

1. **Padding:** Fields $z$ are padded to $\tilde n \times \tilde n$ ($\approx$15% expansion per dimension) to suppress FFT wrap-around.
2. **Forward UDCT:**
   - Compute 2D FFT of padded $z$.
   - For each $(j, \ell)$, mask spectrum: $\hat d_{j,\ell}(\omega) = W_j(\|\omega\|)\, V_{j,\ell}(\arg \omega)\, \hat z(\omega)$.
   - Inverse FFT to obtain $d_{j,\ell}(u,v)$.
   - Crop to $n \times n$ if needed; vectorize and stack.
3. **Inverse UDCT:** 
   - Pad coefficient slices and FFT.
   - Multiply by conjugated window masks, sum over $(j, \ell)$, inverse FFT, crop.

Memory use is dominated by window storage and the coefficient array ($N \approx 8 M$). The transform is $O(M \log M)$ per plane due to FFT efficiency and sharing of buffers.

## 4. Integration with Convex MAP Reconstruction and ADMM Splitting

The total per-plane optimization problem is
\[
\hat z = \arg\min_{z} \frac12\|W^{1/2}(H z-y)\|_2^2 + \frac12 z^\top Q z + \lambda \|C z\|_1,
\]
where $H$ is the interpolation matrix for observed points, $W$ encodes noise weights, and $Q$ is the Matérn-type GMRF precision. Introducing $d = C z$ and recasting as constrained minimization, the problem is solved by ADMM:

- **$z$-update:** Solve $(H^\top W H + Q + \rho I)\,z^{k+1} = H^\top W y + \rho C^*(d^k-u^k)$ via sparse Cholesky or CG.
- **$d$-update:** Apply soft-thresholding shrinkage to $C z^{k+1} + u^k$ with $\tau = \lambda/\rho$.
- **Dual update:** $u^{k+1} = u^k + (C z^{k+1} - d^{k+1})$.

ADMM convergence is guaranteed under convexity, with termination upon reduction of primal and dual residuals below set tolerances. Parseval-tightness ($C^*C \approx I$) ensures efficient block decoupling.

## 5. Soft-Thresholding (Shrinkage) Operator for Complex Coefficients

For coefficient $w$ and threshold $\tau$, the operator is
\[
\mathrm{shrink}(w, \tau) = \max\left(0, 1 - \frac{\tau}{|w|}\right) w
\]
yielding zero when $|w| \le \tau$ and $w - \tau w/|w|$ otherwise. For real $w$, this becomes $\mathrm{sign}(w)\max(|w|-\tau,0)$. $\tau$ is tied to the curvelet sparsity parameter and the ADMM penalty, with its selection guided by discrepancy principles or cross-validation.

## 6. Numerical Benchmarks Demonstrating Effect of UDCT Regularization

Section 9 presents synthetic “ridge + blob” experiments isolating the curvelet penalty, with summary metrics:

| Method                       | RMSE (hold-out)      | $R^2$               |
|------------------------------|----------------------|---------------------|
| GMRF/SPDE (no curvelets)     | $1.83 \pm 0.33$      | $-2.97 \pm 4.28$    |
| + Curvelets via ADMM         | $1.15 \pm 0.30$      | $-0.12 \pm 0.35$    |

Thus, inclusion of $\ell_1$ UDCT penalty yields a $37\%$ decrease in RMSE and pushes $R^2$ near zero, demonstrating directional sparsity value. L-curve analysis substantiates classical residual–sparsity trade-offs and justifies $\lambda$ selection strategies. Robustness is demonstrated across grid resolutions $(n = 64, 128, 256)$, and additional studies detail sensitivity to smoothness and sparsity weights.

## 7. Context, Implications, and Significance

In the referenced framework, the UDCT block introduces anisotropic, multiscale sparse regularization that is efficiently split from conventional quadratic blocks via ADMM. The transform’s undecimated structure ensures shift-invariance and effective localization of ridge/edge features typical of fractured media. Empirical studies establish the superiority of curvelet sparsity over isotropic priors alone, particularly in reconstructing curvilinear anomalies. A plausible implication is that similar undecimated, directional transforms may serve as essential regularization tools in other inverse problems where anisotropic structures dominate [2601.17485].

Source: https://www.emergentmind.com/topics/undecimated-discrete-curvelet-coefficients