Optimally damped rank-reduction method

A shortcoming of the traditional rank-eduction and damped rank-reduction methods is that the output perfomance is quite sensitive to the input parameter $N$. The rank parameter varies greatly for different datasets, especially for field datasets. In order to alleviate the influence of the rank parameter, we take advantage of an adaptive singular-value weighting algorithm developed by Benaych-Georges and Nadakuditi (2012) and Nadakuditi (2013). An adaptive singular-value weighting matrix can be obtained by solving the following problem:

$\displaystyle \hat{\boldsymbol{\omega}}=\arg\min_{\boldsymbol{\omega}} \paralle...
...i^S(\mathbf{v}_i^S)^H - \omega_i\sigma_i\mathbf{u}_i\mathbf{v}_i^H \parallel_F,$ (11)

where $\mathbf{u}_i^S$ and $\mathbf{v}_i^S$ denote $i$th left and right singular vectors corresponding to the estimated signal component, $\sigma_i$ denotes the $i$th singular value in $\boldsymbol{\Sigma}$. $\boldsymbol{\omega}$ is the adaptive weight vector, which can be used to construct the adaptive weighting matrix $\hat{\mathbf{W}}=$diag$(\hat{\boldsymbol{\omega}})$. The solution to the optimization problem can be obtained as:

$\displaystyle \hat{w}_i=\left(-\frac{2}{\sigma_i}\frac{\mathcal{D}(\sigma_i;\boldsymbol{\Sigma})}{\mathcal{D}'(\sigma_i;\boldsymbol{\Sigma})}\right).$ (12)

where $\mathcal{D}$ represents a transform expressed as:

$\displaystyle \mathcal{D}(\sigma;\boldsymbol{\Sigma})= \frac{1}{N^2}$   Tr$\displaystyle \left(\sigma(\sigma^2\mathbf{I}-\boldsymbol{\Sigma}\boldsymbol{\Sigma}^H)^{-1}\right)$   Tr$\displaystyle \left(\sigma(\sigma^2\mathbf{I}-\boldsymbol{\Sigma}^H\boldsymbol{\Sigma})^{-1}\right),$ (13)

and $\mathcal{D}'$ denotes its derivative. The expression of $\mathcal{D}'$ can be expressed as:

\begin{displaymath}\begin{split}
D'(\sigma;\boldsymbol{\Sigma})&= 2\left[\frac{1...
...athbf{I}-\boldsymbol{\Sigma}^2)^{-2}\right)\right].
\end{split}\end{displaymath} (14)

The symbol Tr$(\cdot)$ denotes the trace of the input. The trace of a square matrix $\mathbf{X}$ is defined to be the sum of elements on the main diagonal of $\mathbf{X}$.

Tr$\displaystyle (\mathbf{X}) = \sum_{i=1}^{N}X_{i,i},$ (15)

where $X_{i,i}$ denotes the main diagonal elements of $\mathbf{X}$.

The adaptive weighting operator can be applied to make the traditional rank-reduction method adaptive:

$\displaystyle \hat{\mathbf{S}}=\mathbf{U}_N\hat{\mathbf{W}}\boldsymbol{\Sigma}_N\mathbf{V}_N^H.$ (16)

To incorporate the adaptive weighting operator into the damped rank-reduction framework, we can introduce intermediate variables as

$\displaystyle \mathbf{U}_N^{Q} = \mathbf{U}_N,$ (17)
$\displaystyle \boldsymbol{\Sigma}_N^{Q} = \hat{\mathbf{W}}\boldsymbol{\Sigma}_N,$ (18)
$\displaystyle \mathbf{V}_N^{Q} = \mathbf{V}_N,$ (19)

then equation 16 turns into

$\displaystyle \hat{\mathbf{S}}=\mathbf{U}_N^{Q}\boldsymbol{\Sigma}_N^{Q}\mathbf{V}_N^{Q}.$ (20)

It can be derived that the damping formula (equation 8) also holds for equation 20 but with the damping threshold matrix expressed as

$\displaystyle \boldsymbol{\Gamma} = \hat{\sigma}^K\left(\boldsymbol{\Sigma_N^Q}\right)^{-K},$ (21)

where the subscript $N$ denotes a sufficiently large rank parameter, as required by the derivations detailed in Chen et al. (2019b). The resulted final form of the optimally damped rank-reduction method then can be expressed as

$\displaystyle \hat{\mathbf{S}} = \mathbf{U}_N\mathbf{P}\hat{\mathbf{W}}\boldsymbol{\Sigma}_N\mathbf{V}_N^H.$ (22)

The algorithm workflow for the optimally damped rank-reduction method (ODRR) is outlined in Algorithm 3. Due to the existence of a weighting matrix in the proposed method, it is more convenient to choose the input rank parameter $N$ empirically in practice. As mentioned previously, the rank parameter $N$ is in practice set to a relatively large value to forestall damaging the signal. But in the proposed algorithm, even if a large $N$ is used, the algorithm can adaptively shrink the singular-values. Thus, its performance is not sensitive to the input rank parameter, in contrast to other related approaches (Oropeza and Sacchi, 2011). Because of the insensitivity, one can use a sufficiently large $N$ in processing complicated datasets without leaving strong residual noise in the result. The convenience in tuning parameters makes the rank-reduction related methods more computationally feasible in large-scale data processing, e.g., the 5D reconstruction problem (Chen et al., 2019b), since one no longer needs to tune the parameters many times while one trial is already computationally demanding. A glossary describing the main mathematical notations are presented in Table 1. The three methods are all variations of how the singular-values are thresholded. Simply speaking, the DRR method improves the RR method by introducing a damping operation and the ODRR method improves DRR method by further introducing a weighting operation.


Table 1: Glossary of main mathematical notations used in this paper.
Symbols Meanings
$\mathbf{D}(t/w,x,y)$ 3D noisy seismic data
$\mathbf{D}(w_0)$ or $\mathbf{D}$ abbreviated notation of 2D frequency slice
$\mathbf{R}_i$ $i$th Hankel matrix
$\mathbf{M}$ block Hankel matrix
$\mathbf{S}$ signal component
$\hat{\mathbf{S}}$ estimated signal component
$\mathbf{U}$ left singular vector matrix
$\boldsymbol{\Sigma}$ singular-value matrix
$\mathbf{V}$ right singular vector matrix
$\hat{\mathbf{W}}$ weighting matrix
$\mathbf{P}$ damping matrix
$\boldsymbol{\Gamma}$ damping threshold matrix
$\mathbf{Q}$ temporary variable
$\mathcal{D}$ D-transform
$\mathcal{H}$ Hankelization operator
$\mathcal{T}$ thresholding operator
$\mathcal{A}$ averaging operator
$N$ rank
$K$ damping factor


Table 2: Comparison of SNRs in dB for different rank-reduction based methods. RR denotes the rank-reduction method. DRR denotes the damped rank-reduction method. ODRR denotes the optimally damped rank-reduction method.
Tests Noisy (dB) RR (dB) DRR (dB) ODRR (dB)
Linear synthetic (N=3) -8.39 6.57 10.29 11.27
Linear synthetic (N=6) -8.39 3.45 8.35 10.86
Hyperbolic synthetic (N=10) -2.17 8.27 9.58 9.65
Hyperbolic synthetic (N=20) -2.17 7.04 10.08 11.00


Table 3: Comparison of computational costs in seconds for different rank-reduction based methods.
Tests FK (s) RR (s) DRR (s) ODRR (s)    
Linear synthetic (N=3) 0.13 2.27 2.45 2.43    
Linear synthetic (N=6) 0.17 3.02 3.04 3.17    
Hyperbolic synthetic (N=10) 0.43 1.69 1.71 1.83    
Hyperbolic synthetic (N=20) 0.49 2.75 2.69 2.83    


2020-12-06