Shearlets are multiscale filters that are long and thin and tilted by shearing. They represent edges in images and surfaces in volumes far more compactly than ordinary wavelets. The theory justifying them is asymptotic: it says how fast approximation error falls as more coefficients are kept, and it hides every multiplicative constant. In two dimensions this rarely matters. In three dimensions it does. Each scale needs a two-parameter family of shear directions in each of three frequency cones, so the number of subbands grows quickly. Every extra subband adds memory, raises the noise threshold, and can worsen how evenly the filters cover frequency space. This proposal argues that at practical volume sizes these constants, not the asymptotic rate, decide which 3D shearlet system works best. It treats the choice of how many shear directions to use at each scale as an optimization of an explicit, constant-tracking error bound, solved under a memory budget.
The practical stakes are concrete. Volumetric shearlet denoising of retinal optical coherence tomography (OCT) already uses 99 subbands and is tuned by hand. Universal digital filtering for denoising volumetric retinal OCT and OCT angiography in 3D shearlet domain reports 21.81 dB peak signal-to-noise ratio (PSNR) against 21.46 dB for BM4D, a standard non-local 3D denoiser. The ShearLab 3D authors note that performance is sensitive to the number of scales and shear levels, and that no principled rule exists for setting them (ShearLab 3D). We propose to supply that rule, check it against data-driven tuning, and explain how the right choice changes when moving from 2D to 3D.
Core Claim
For FFT-based digital shearlet systems, the frame operator is diagonal in the Fourier domain. That operator is the map that sums each filter times its own response. Being diagonal, its lower and upper frame bounds $A$ and $B$ are simply the minimum and maximum, over the frequency grid, of $\sum_i |\hat\psi_i(\xi)|^2$. Here $\hat\psi_i$ is the Fourier transform of filter $i$, and $A$ and $B$ measure how unevenly the filters cover frequency space. These bounds can therefore be computed exactly for any candidate design, not merely asserted to exist.
Combining this with a standard oracle argument for thresholding gives a dimension-explicit risk bound. Assume reconstruction uses the canonical dual frame, which is what ShearLab-style inversion does by dividing by that Fourier sum, so its frame bound is $1/A$. Then thresholding coefficients corrupted by white noise of level $\sigma$ satisfies, up to an absolute constant $C$:
$$\mathbb{E}\|f-\hat f\|^2 \;\le\; \frac{C}{A}\sum_{i=1}^{M}\min\!\big(|c_i|^2,\ \lambda_i^2\big),\qquad \lambda_i=\sigma\|\psi_i\|\sqrt{2\ln M}.$$
In this bound, $c_i=\langle f,\psi_i\rangle$ are the clean coefficients and $M$ is the total coefficient count. The design $\mathbf{s}=(s_1,\dots,s_J)$, the shear level at each of the $J$ scales, enters in three places. It sets $M$, which grows roughly like $3(2^{s_j}+1)^2$ directions per scale in 3D versus $2(2^{s_j}+1)$ in 2D. It sets $A$. And it sets how many coefficients of a surface-bearing volume exceed threshold, since finer shears align better with surfaces and leave fewer large coefficients.
The asymptotic theory of The Continuous Shearlet Transform in Arbitrary Space Dimensions prescribes parabolic growth of $s_j$ with scale. The claim is that the bound-minimizing design departs from parabolic scaling much more in 3D than in 2D. In 3D it should use fewer shears at fine scales and accept slightly worse alignment for a much smaller $M$ and a better $A$. The claim also holds that choosing the design by minimizing the bound under a memory budget matches or beats both the hand-tuned defaults and per-volume data-driven tuning.
Relation to Prior Work
The proposal draws on three roles of prior work.
Theory that motivates the gap. The group-theoretic line gives existence results without constants. Dahlke et al. build the $n$-variate shearlet group and prove that Banach frames exist, but they leave frame constants and discretization parameters unspecified for $n\ge 3$. Continuous Shearlet Tight Frames obtains exact tightness ($A=B=1$) only in the continuous 2D setting. Dualizable Shearlet Frames and Sparse Approximation tightens the logarithmic factor in the 2D approximation rate, again only asymptotically. We do not prove new existence results. We compute and optimize the constants these papers leave open, for the discrete systems people actually run.
Software that supplies the method. The digital line provides the object to optimize. ShearLab: A Rational Design of a Digital Parabolic Scaling Algorithm shows that tightness can be engineered to machine precision ($M_{alg}\approx 6.6\times10^{-16}$) in 2D, but only with radial oversampling $R=8$. That price in redundancy is exactly what becomes prohibitive in 3D. ShearLab 3D supplies compactly supported universal shearlets whose anisotropy can vary by scale, which makes non-parabolic designs expressible. We reuse its filter construction and change only how its parameters are chosen.
Constant-tracking analysis that supplies the attitude. Several papers show that constants hidden by asymptotic proofs are routinely loose and that tracking them changes conclusions. Offset Rademacher localization deliberately leaves its constant at $c=1/18$ (Learning with Square Loss). Explicit Hardy-operator constants on product spaces show that earlier literature values were not sharp (Sharp bounds for Hardy type operators). The CNN deformation bound $C_K\|\tau\|_\infty^{1/2}$ on cartoon functions in $\mathbb{R}^d$ never makes the $d$-dependence of $C_K$ explicit (Deep Convolutional Neural Networks on Cartoon Functions).
Constraint on the claim. The dimension-gap result of Near-Minimax Optimal Estimation With Shallow ReLU Neural Networks shows that all linear estimators degrade badly with dimension on ridge-structured functions. Fixed-dictionary thresholding is nonlinear, so that result does not rule it out. It does warn that any advantage we find is a statement about constants at moderate $d$, not a claim that the approach escapes dimensional scaling.
Research Program
Step 1: certified frame bounds. We will write a routine that, for any per-scale shear vector $\mathbf{s}$ and anisotropy choice in a ShearLab 3D-style system, computes $A$, $B$, $M$, memory footprint and filter norms exactly on the target grid. We will tabulate how $B/A$ and $M$ scale with $\mathbf{s}$ in $d=2$ and $d=3$.
Step 2: a predictive risk curve. We will model the sum in the bound for cartoon-like volumes, meaning piecewise smooth functions with $C^2$ surface discontinuities and, separately, with filaments. At scale $j$ and shear level $s_j$, a surface of area $S$ meets about $S\cdot 2^{j}\cdot 2^{\,s_j}$ shearlets per aligned direction. Coefficient magnitude decays away from tangency, following the hyperplane-decay theorem in Dahlke et al.. This yields a closed-form predicted risk $R(\mathbf{s};\sigma,S,N)$, where $N$ is the side length of the volume.
Step 3: budgeted design. We will solve $\min_{\mathbf{s}} R(\mathbf{s})$ subject to $M(\mathbf{s})\le M_{\max}$ by exhaustive search over the small discrete design space. The output is a design rule that depends only on $\sigma$, $N$, $d$ and a rough surface-density estimate, not on the volume itself.
Step 4: dimension analysis. We will compare the optimal $\mathbf{s}$ in 2D and 3D and measure the gap to the parabolic default. The prediction is that this gap is small in 2D and large in 3D.
Baselines
Do-nothing reference. The first reference point is the ShearLab 3D default parameterization with hard thresholding at the universal threshold. This is the naive choice a practitioner makes. Beating it at equal or lower memory is the minimum bar.
Strongest existing 3D shearlet result. The Yang et al. 3D shearlet filter is reported at 21.81 dB PSNR on clinical OCT, versus 21.46 dB for BM4D, with 99 subbands and 19 s versus 332 s per volume (Yang et al.). If their volumes are obtainable, we rerun their pipeline with their subband count and our designed subband allocation at the same count. Otherwise we compare on comparable OCT volumes and report their numbers only as context. Beating it means higher PSNR at equal $M$, or equal PSNR at a materially smaller $M$.
BM4D. BM4D is run on the same volumes and noise levels. It is the non-shearlet reference.
Separable 3D wavelets. Separable wavelets with thresholding serve as the isotropic floor.
Per-volume SURE tuning (most serious competitor). Stein's unbiased risk estimate (SURE) gives an unbiased estimate of mean squared error computed from the noisy data alone. Here it is used with soft thresholding to grid-search $\mathbf{s}$ separately for each volume. It is given the same design space and memory cap as our method, and it gets oracle knowledge of $\sigma$.
Ablations isolating the key ingredient. The first ablation is an asymptotic-only rule: minimize the same bound with $A$ fixed to 1 and $\ln M$ treated as constant. This is what the rate theory alone would prescribe. The second ablation keeps $M$ but drops $A$. If our designs beat the first ablation but not the second, the frame-bound term is doing no work, and we will say so.
Why not just…?
Why not just tune shear levels per volume with SURE or a held-out grid search? This is the real threat. On a single volume of known noise level, SURE may match our choice. The proposal's practical value then rests on three things SURE does not offer. First, our rule selects a design before seeing data, which matters when one design must be fixed for a scanner or hardware pipeline. Second, the rule transfers across volume sizes $N$, where SURE must be rerun. Third, it explains why the optimum moves with dimension. The experiments compare the two at equal memory. If SURE matches us on every volume, we will report the contribution as explanatory and predictive rather than as a better denoiser.
Why not just compute $A$ and $B$ numerically for the default and stop? Computing them is part of our method, but bounds alone do not choose a design. The decision needs the trade-off between alignment gain, coefficient count and frame quality, and it is the risk curve that couples them. The second ablation, dropping $A$, tests directly whether the certified bounds change decisions.
Why not just denoise slice by slice in 2D, where defaults are known to work? Yang et al. report that 2D B-scan filtering fails to separate capillaries from speckle, and that the 3D decomposition is what yields 21.81 dB (Yang et al.). We include a slice-wise 2D shearlet baseline at matched total memory to confirm this on our data.
Why not just train a 3D CNN denoiser? Learned denoisers need clean training volumes and give no certified reconstruction bound. No reported result was found on Lacuna for a learned 3D denoiser on these OCT volumes. This proposal targets the regime of fixed, data-free transforms, which is the regime in which LDAN also argues fixed shearlet filters remain competitive. We treat learned denoisers as out of scope rather than claim to beat them.
Proposed Experiments
Synthetic tests of the prediction. On synthetic 3D cartoon volumes with known surfaces and filaments, at sizes $N\in\{64,128,256\}$ and several noise levels, we will measure PSNR, coefficient count, runtime and certified $B/A$ for every design. We then check whether the predicted risk curve ranks designs in the same order as measured error. The decisive check is a rank correlation between predicted and measured risk across designs. A second check is whether the predicted 3D optimum is non-parabolic and appears in measured error.
2D sanity check. On the standard grayscale images lenna, barbara, boat and peppers used by ShearLab 3D, the rule should approximately recover the defaults. Failure here would discredit the risk model.
Real volumes. On retinal OCT and OCTA volumes, we compare our budgeted design against the default, the Yang et al. configuration, BM4D, separable wavelets, slice-wise 2D shearlets and per-volume SURE, all at matched memory caps. The headline result would be one of two outcomes. Either the bound-designed system, fixed before seeing any test volume, matches per-volume SURE and beats the default and the 21.81 dB configuration at equal $M$. Or it reaches that configuration's quality with substantially fewer subbands.
Falsifying outcome. If the asymptotic-only ablation performs as well as the full rule, the claim that constants decide 3D design is false. We would report it as such.
References
- Universal digital filtering for denoising volumetric retinal OCT and OCT angiography in 3D shearlet domain
- ShearLab 3D: Faithful Digital Shearlet Transforms Based on Compactly Supported Shearlets
- The Continuous Shearlet Transform in Arbitrary Space Dimensions
- Continuous Shearlet Tight Frames
- Dualizable Shearlet Frames and Sparse Approximation
- ShearLab: A Rational Design of a Digital Parabolic Scaling Algorithm
- Learning with Square Loss: Localization through Offset Rademacher Complexity
- Sharp bounds for Hardy type operators on higher-dimensional product spaces
- Deep convolutional neural networks on cartoon functions
- Near-Minimax Optimal Estimation With Shallow ReLU Neural Networks
- Revisiting non-learned operators based deep learning for image classification: a lightweight directional-aware network