1. The compressed sensing model
We start with:
where:
: measurement vector.
: sensing matrix.
: sparse signal.
We can express the multiplication as:
Here, represents the -th column of matrix .
This is the key point: Each column represents the contribution of one signal coefficient to the measurement vector.
2. The measurement matrix (sensing matrix) in Orthogonal Matching Pursuit (OMP) for compressed sensing.
In general, a matrix has dimensions:
where:
= number of rows (measurements).
= number of columns (signal coefficients).
In OMP, each column represents one possible signal atom or dictionary element.
For example:
OMP works by selecting the column that has the highest correlation with the current residual:
Then it adds that selected column to the support set.
Important distinction:
The original sensing matrix contains columns.
OMP selects only a few columns (say, columns).
These selected columns form an equivalent or reduced sensing matrix:
So, OMP does not mean the original matrix becomes a single column. Rather, it selects columns iteratively to construct a reduced matrix.
3. Why does OMP select columns?
Suppose the original signal is sparse:
Only the second and fifth coefficients are nonzero.
Therefore:
OMP tries to identify which columns contributed to the measurement vector.
It selects columns and , ideally recovering the sparse signal.
4. Why not rows?
Rows represent individual measurement equations, not individual signal coefficients.
For example, each row tells us how all signal coefficients contribute to one particular measurement.
Selecting rows would mean selecting measurements, rather than identifying which signal coefficients are nonzero.
5. Remember this concept
Matrix component | Meaning |
|---|---|
Columns | Signal atoms / coefficients |
Rows | Measurement equations |
Selected columns | Estimated nonzero signal support |
In one sentence: OMP selects columns because its objective is to identify the nonzero elements of the sparse signal , and each column of corresponds to one element of .