1: \begin{abstract}

2: We consider the linear inverse problem of estimating an unknown signal $f$ from

3: noisy measurements on $Kf$ where the linear operator $K$ admits a

4: wavelet-vaguelette decomposition (WVD).

5: %This refers to a setup which can be described by the following sequence model:

6: %$$%y_{jk} = \lambda_j \theta_{jk} + \epsilon z_{jk},\qquad k=0,1,\ldots

7: %2^j-1, \quad  j=1,2,\ldots

8: %$$%where $\theta = (\theta_{jk})$ is the vector of wavelet coefficients

9: %of the unknown function $f$. The weights $\lambda_j$

10: %satisfy $\lambda_j \sim 2^{-\alpha j}$ for some $\alpha > 0$. And the noise

11: %$z_{jk}$ are assumed to be Gaussian with 0 mean, and their covariance matrix

12: %has bounded condition number. We proceed to estimate $\theta$ (equivalently

13: %$f$)

14: We formulate the problem in the Gaussian sequence model and propose

15: estimation based on complexity penalized regression

16: %scheme for the vaguelette coefficients of the signal

17: on a level-by-level basis.

18: We adopt squared error loss and show that the estimator achieves exact

19: rate-adaptive optimality

20: % is used as performance measure for estimation of $f$, then the

21: % proposed estimator achieves the optimal rate of convergence in the small noise

22: % limit

23: as $f$ varies over a wide range of Besov function classes.

24: \end{abstract}