In sparse recovery one aims to leverage the sparsity of some vector to get a good reconstruction of this vector from a noisy and indirect measurement . Sparsity can be expressed in two different ways: The analysis way and the synthesis way. For both of them you need a system of other vectors .

**The analysis way:**Here you consider the inner products and the sparsity assumption is, that this sequence is sparse.**The synthesis way:**Here you assume that there is sparse sequence of coefficients such that .

The first approach is called “analysis” since here one *analyzes* the vector with the system of vectors. In the second approach, you use the vectors to synthesize the vector .

We will keep things very simple here and only consider a noisy observation (and not an indirect one). The recovery approach in the analysis way is then to find as a solution of

while for the synthesis way you write (where we collected the vectors as the columns of the matrix ) and solve

The functional should enforce the sparsity of the vector of coefficients in the former case and the sparsity of in the latter case. With the matrix the analysis way reads as

One important observation is, that both approaches coincide if the matrix is orthonormal: Since in this case, is equivalent to , i.e. . For other sets of vectors , this is not true and it is not so clear, how the analysis and the synthesis approach are related. In this post I’d like to show one instance where both approaches do lead to results that are not even close to being similar.

**1. The “ROF analysis” approach to denoising **

If is an image, the famous Rudin-Osher-Fatemi denoising model fits under the “analysis” framework: The vectors are all the finite differences of neighboring pixels such that with the discrete gradient . As sparsifying functional you take a mixed – norm, i.e. . The denoised is the solution to

Thus, the model will lead to some denoised with a sparse gradient. This sparse gradient is partly the reason for the success of the model in that it keeps edges, but is also responsible for the drawback of this approach: The denoised image will be mostly piesewise constant.

**2. The “ROF synthesis” approach **

As , the corresponding synthesis approach would be to solve

and the set . In this approach, the denoised image will be a sparse linear combination of particular two-pixel images. If you think about that, it does not really sound like a good idea to approximate an image by a sparse linear combination of two-pixel images. (One technicality: The operator is not onto – has a non-trivial kernel, consisting of the constant images and hence, all images in the range of have zero mean. Thus, to get something remotely close to one should subtract the mean of before the reconstruction and add it back afterwards.)

Here are some results:

Note that the results actually fit to what we would have predicted: For larger we get a “sparser gradient” (less edges, less variation) for the analysis approach, and “sparser two-pixel combinations” for the synthesis approach. In fact, for the synthesis approach to give something that is remotely close to the image, one needs quite a lot two-pixel images, i.e. a very low sparsity.

In conclusion, one should not consider the analysis and synthesis approach to be something similar.

Incidentally, the dual of the analysis approach

can be written as

which looks related to the synthesis approach (one basically replaces the -norm by its conjugate function). Actually, a similar relation between the dual of the analysis approach and the synthesis approach always holds.