Recent studies have demonstrated the potential of dictionary learning for painter style analysis. The main idea behind these approaches is to train a dictionary of image atoms on a set of drawings/paintings of the same artist and to test how well this dictionary can represent a painting with disputed authorship, by evaluating the sparsity of representation. We extend this approach such that it can evaluate the goodness of fit between the trained dictionary and the test image both in the synthesis direction (similar to existing approaches) and in the opposite, analysis direction. We evaluate these approaches on oil paintings, focusing on a case study on the Ghent Altarpiece. The results give new insights into the potential of dictionary learning for painter style characterisation and suggest advantages of the proposed analysis based approach irrespective of the scale and redundancy of trained dictionaries.