1: \begin{abstract}

2:   We develop a method for sparse image reconstruction from

3:   polychromatic \gls{CT} measurements under the blind scenario where the 

4:   material of the inspected object and the incident-energy spectrum are 

5:   unknown.  We obtain a parsimonious measurement-model parameterization by 

6:   changing the integral variable from photon energy to mass attenuation, 

7:   which allows us to combine the variations brought by the   unknown 

8:   incident spectrum and mass attenuation into a single unknown 

9:   \emph{mass-attenuation spectrum} function; the resulting measurement 

10:   equation has the Laplace integral form. The mass-attenuation spectrum is 

11:   then expanded into basis functions using B-splines of order one.  We 

12:   derive a block coordinate-descent algorithm for constrained minimization 

13:   of a penalized \gls{NLL} cost function, where penalty terms ensure 

14:   nonnegativity of the spline coefficients and nonnegativity and sparsity 

15:   of the density map image.  The image sparsity is imposed using convex 

16:   \gls{TV} and $\ell_1$ norms, applied to the density-map image and its 

17:   \gls{DWT} coefficients, respectively.  This algorithm alternates between 

18:   \gls{NPG} and \gls{LBFGSB} steps for updating the image and 

19:   mass-attenuation spectrum parameters.  To accelerate convergence of the 

20:   density-map \gls{NPG} step, we apply a step-size selection scheme that 

21:   accounts for varying local Lipschitz constant of the \gls{NLL}.  We 

22:   consider lognormal and Poisson noise models and establish conditions for 

23:   biconvexity of the corresponding \glspl{NLL} with respect to the density 

24:   map and  mass-attenuation spectrum parameters.  We also prove the 

25:   Kurdyka-{\L}ojasiewicz property of the objective function, which is 

26:   important for establishing local convergence of block-coordinate schemes 

27:   in biconvex optimization problems.  Numerical experiments with simulated 

28:   and real X-ray \gls{CT} data demonstrate the performance of the proposed 

29:   scheme.  \end{abstract}

30: