## Flexible empirical Bayes estimation for wavelets (2000)

### Cached

### Download Links

Venue: | Journal of the Royal Statistics Society, Series B |

Citations: | 74 - 14 self |

### BibTeX

@ARTICLE{Clyde00flexibleempirical,

author = {Merlise Clyde and Edward I. George},

title = {Flexible empirical Bayes estimation for wavelets},

journal = {Journal of the Royal Statistics Society, Series B},

year = {2000}

}

### Years of Citing Articles

### OpenURL

### Abstract

Wavelet shrinkage estimation is an increasingly popular method for signal denoising and compression. Although Bayes estimators can provide excellent mean squared error (MSE) properties, selection of an effective prior is a difficult task. To address this problem, we propose Empirical Bayes (EB) prior selection methods for various error distributions including the normal and the heavier tailed Student t distributions. Under such EB prior distributions, we obtain threshold shrinkage estimators based on model selection, and multiple shrinkage estimators based on model averaging. These EB estimators are seen to be computationally competitive with standard classical thresholding methods, and to be robust to outliers in both the data and wavelet domains. Simulated and real examples are used to illustrate the flexibility and improved MSE performance of these methods in a wide variety of settings.

### Citations

929 | Ideal spatial adaptation by wavelet shrinkage - Donoho, Johnstone - 1994 |

768 | Adapting to unknown smoothness via wavelet shrinkage - Donoho, Johnstone - 1995 |

451 | Maximum likelihood estimation from incomplete data via the EM algorithm - Dempster, Laird, et al. - 1977 |

256 | Wavelet Shrinkage: Asymptopia
- Donoho, Johnstone, et al.
- 1996
(Show Context)
Citation Context ...neral not positive definite for all values of cj, ωj, and σ 2 , convergence may only be to a local mode. However, we have achieved reasonable success using the MAD estimate ˆσ = Median(|D1k|)/0.6745 (=-=Donoho et al. 1995-=-) as an initial value for σ, and ˆωj equal to the number of observed wavelet coefficients that exceed √ 2 log n ˆσ based on hard thresholding. 3.2 Maximum Likelihood Estimation using the EM Algorithm ... |

221 | Wavelet thresholding via a Bayesian approach
- Abramovich, Sapatinas, et al.
- 1998
(Show Context)
Citation Context ...orm (IDWT) to obtain an estimate of the unknown function f . Bayesian methods provide a natural and coherent approach to adaptive data dependent shrinkage and thresholding, and several recent papers (=-=Abramovich et al. 1998-=-, Chipman et al. 1997, Clyde et al. 1998) show that they can obtain superior performance over early shrinkage methods. The above Bayesian methods involve taking the standard linear model (1) with inde... |

204 | Wavelet threshold estimators for data with correlated noise - Johnstone, Silverman - 1997 |

183 | Adaptive bayesian wavelet shrinkage - Chipman, Kolaczyk, et al. - 1997 |

160 |
Scale mixtures of normal distributions
- Andrews, Mallows
- 1974
(Show Context)
Citation Context ... conditionally independent, the λjk’s are independent, and h() is a scale mixing distribution on (0, ∞). Scale mixtures of normals have been widely used in robustness studies and in outlier analysis (=-=Andrews and Mallows 1974-=-, West 1984, 1987, O’Hagan 1979, 1988) and include as special cases the normal, 2 (1)sLaplace, exponential power, and Student t distributions. The normal model is obtained when λjk ≡ 1, whereas indepe... |

130 | Calibration and empirical Bayes variable selection - George, Foster - 2000 |

127 | Multiple shrinkage and subset selection in wavelets
- Clyde, Parmigiani, et al.
- 1998
(Show Context)
Citation Context ...ators for non-normal error distributions that are robust to outliers. 2.1 Hierarchical Model The cornerstone of our Bayesian approach is an extension of the hierarchical normal mixture prior used by (=-=Clyde et al. 1998-=-). Based on the natural multilevel grouping of the wavelet coefficients, this distribution for the βjk’s at level j is βjk | λ ∗ jk , γjk ∼ N(0, σ 2 cjγjk/λ ∗ jk ) (5) λ ∗ jk ∼ h ∗ γjk ∼ Bernoulli(ωj)... |

74 |
Elements of statistical computing
- Thisted
- 1988
(Show Context)
Citation Context ...mates ĉj and ˆωj are generally unavailable, standard iterative methods can be used for direct maximization of the log likelihood (15). In 8sparticular, we have found nonlinear Gauss-Seidel iteration (=-=Thisted 1988-=-, pp 187-188) to work well. This entails iterating between finding the maximizing ĉj given (ˆωj, ˆσ 2 ), the maximizing ˆωj given (ĉj, ˆσ 2 ) for each j, and the maximizing ˆσ 2 given (ĉ1, ˆω1), . . .... |

38 |
Outlier models and prior distributions in Bayesian linear regression
- West
- 1984
(Show Context)
Citation Context ..., the λjk’s are independent, and h() is a scale mixing distribution on (0, ∞). Scale mixtures of normals have been widely used in robustness studies and in outlier analysis (Andrews and Mallows 1974, =-=West 1984-=-, 1987, O’Hagan 1979, 1988) and include as special cases the normal, 2 (1)sLaplace, exponential power, and Student t distributions. The normal model is obtained when λjk ≡ 1, whereas independent error... |

26 |
Minimax multiple shrinkage estimation
- George
- 1986
(Show Context)
Citation Context ...in the hyperparameter estimates) for n = 1024. The posterior mean of f is obtained by applying the IDWT to the multiple shrinkage estimator of β. The estimator (24) is a multiple shrinkage estimator (=-=George 1986-=-, Clyde et al. 1998) and corresponds to model averaging. Because of the form of the conditional distribution (21), the usual posterior weighted sum 12sof conditional expectations reduces here to the s... |

23 | Empirical Bayes approaches to mixture problems and wavelet regression (Tech - Johnstone, Silverman - 1998 |

17 | On outlier rejection phenomena in Bayes inference - O’Hagan - 1979 |

12 | Denoising and Robust Nonlinear Wavelet Analysis - Bruce, Donoho, et al. - 1994 |

3 |
Exact convolution of t distributions, with applications to Bayesian inference for a normal mean with t prior distributions
- Fan, Berger
- 1990
(Show Context)
Citation Context ...utations for this case must be done by numerical integration. From a Bayesian prior robustness perspective, the tails of the prior should typically be at least as flat as the tails of the likelihood (=-=Fan and Berger 1990-=-). When the errors have a tν distribution, this is achieved when the βjk’s are iid tν ∗ with ν∗ ≤ ν, and are independent of the ɛjk’s. This corresponds to taking λ ∗ jk iid Gamma(ν ∗ /2, 2/ν ∗ ) with ... |

1 | Modelling with heavy tails. Bayesian Statistics 3 - O’Hagan - 1988 |