Suppose I have a matrix

$$A= \begin{pmatrix} 1 & 0 & 1 & 0\\ 0 & 1 & 1& 0\\ \end{pmatrix} $$

where the variables are channel information like assume $X_1$, $X_2$, $X_3$ and $X_4$. Considering the typical least-squares problem

$$Y= AX+b$$

However, matrix $A$ is fat and, thus, $A^{T} A$ is not invertible. I cannot get $X_3$ or $X_4$ and they will be zero. I can use regularization and optimize joint least-squares and least-norm problem but that does not give me the same $X_1$ and $X_2$. Is there any way to solve this problem?

I tried to stack all the received vectors into matrix $Y$ and do the same for $A$ and $b$ to make matrix $A$ fat. However, due to the structure of $A$, it is still not invertible. Does anyone know how to solve this problem? Can I recover the solution for parameters based on what I get from regularized solution?

  • $\begingroup$ There might be a answer if we can assume the solution is sparse (is mostly zero) $\endgroup$
    – MimSaad
    Commented Jul 24, 2017 at 5:15
  • 1
    $\begingroup$ Shouldn't $X_k$ be in a discrete set, say, $\{\pm 1\}$, or $\{0,1\}$? $\endgroup$ Commented Jul 24, 2017 at 8:42
  • $\begingroup$ Do you have any further, e.g. quadratic constraints? $\endgroup$ Commented Aug 26, 2017 at 18:23
  • $\begingroup$ Could you please review the answers and mark one of them? $\endgroup$
    – Royi
    Commented Mar 8, 2022 at 18:21

6 Answers 6


It means you will have a non-unique solution or redundancy. In your formulation, $X_4$ is completely free and $X_3$ is an offset parameter. You can deflate your matrix and obtain the full rank part to eliminate the unconstrained part.

In your example the solution is the family of

$$ \begin{bmatrix} X_1\\X_2 \end{bmatrix} = b - \begin{bmatrix} X_3\\X_3\end{bmatrix}$$

You have a degenerate case in which one of the variables is completely free but similar problems are pretty common in robotics with overactuated manipulators and so on for which you optimize within all the possible solutions.


Is there anyway to solve this problem

No, not really. Having such a matrix of reason-observation values being ill-formed simply is that you don't have more information. You cannot spawn information from nowhere.

  • $\begingroup$ The Least Squares solution (Which is also the Laast Norm Solution) always exists for any Linear Equation System. $\endgroup$
    – Royi
    Commented Aug 1, 2017 at 6:02

As it can be observed, if $b=0$ , since $A$ is non square matrix. Thus inverse cannot exist.

Observe that the pseudo-inverse $A^TA$ does not exist, due to non independent columns.

The only way that the problem can be addressed is to introduce additional conditions on the signal such as sparsity. If $X$ has 2 or fewer non zero coefficients then there is a possible solution.

  • $\begingroup$ Your answer has a contradiction with the fact that a least squares solution minimizes squared Euclidean norm of the residual error Ax+b-y (which always has a solution). $\endgroup$
    – msm
    Commented Jul 23, 2017 at 11:23
  • 1
    $\begingroup$ The pseudoinverse of $\rm A^\top A$ certainly does exist. $\endgroup$ Commented Jul 24, 2017 at 8:18
  • $\begingroup$ As Rodrigo says, the Moore-Penrose pseudo inverse will always exist. $\endgroup$
    – Peter K.
    Commented Jul 24, 2017 at 14:06

Since this question has an information-theoretic flavor, let us work over finite field $\mathbb F_2$. Rephrasing:

Given $$\mathrm A = \begin{bmatrix} 1 & 0 & 1 & 0\\ 0 & 1 & 1 & 0\end{bmatrix}$$ and vector $\mathrm y \in \mathbb F_2^2$, find a vector $\mathrm x \in \mathbb F_2^4$ of minimal Hamming weight such that $\rm A x = y$.


Suppose we are given

$$\mathrm y = \begin{bmatrix} 1\\ 1\end{bmatrix}$$

which is the 3rd column of $\rm A$ and also the sum of the 1st and 2nd columns. The solution set then has only two elements

$$\mathrm x \in \left\{ \begin{bmatrix} 0\\ 0\\ 1\\ 0\end{bmatrix}, \begin{bmatrix} 1\\ 1\\ 0\\ 0\end{bmatrix} \right\}$$

The former has Hamming weight $1$, while the latter has Hamming weight $2$.

Fortunately, matrix $\rm A$ is already in reduced row echelon form (RREF)

$$\rm A = \begin{bmatrix} \mathrm I_2 & \mathrm F\end{bmatrix}$$

and, thus, the solution set of the linear system $\rm A x = y$ is parameterized as follows

$$\mathrm x \in \left\{ \begin{bmatrix} \mathrm y\\ 0_2\end{bmatrix} + \begin{bmatrix} \mathrm F\\ \mathrm I_2\end{bmatrix} \mathrm z : \mathrm z \in \mathbb F_2^2 \right\}$$

Only $x_4$ depends on $z_2$. In order to avoid increasing the Hamming weight of $\mathrm x$ unnecessarily, let us choose $x_4 = z_2 = 0$. The solution set of the linear system $\rm A x = y$ is given by

$$\mathrm x \in \left\{ \begin{bmatrix} y_1\\ y_2\\ 0\\ 0\end{bmatrix} + z_1 \begin{bmatrix} 1\\ 1\\ 1\\ 0\end{bmatrix} : z_1 \in \mathbb F_2 \right\} = \left\{ \begin{bmatrix} y_1\\ y_2\\ 0\\ 0\end{bmatrix}, \begin{bmatrix} y_1+1\\ y_2+1\\ 1\\ 0\end{bmatrix} \right\} =: \mathcal X (\mathrm y)$$

Thus, the solution of minimal Hamming weight is

$$\mathrm x^* := \arg \min_{\mathrm x \in \mathcal X (\mathrm y)} w (\mathrm x)$$

where function $w : \mathbb F_2^4 \to \mathbb N$ returns the Hamming weight of its input. To summarize, we find the optimal solution by comparing the Hamming weights of two $4$-dimensional vectors. That is all!


Least Squares solution is always well defined for Linear System of Equations.

In your case, which is under determined it means there are many solutions to the Linear Equations.
The Least Squares solution has nice property, it also minimizes the $ {L}_{2} $ norm of the solution (Least Norm Solution) hence it is well defined.

In practice, in order to solve this equation you need to use the pinv() function in MATLAB.
This function uses the Moore Penrose Pseudo Inverse Matrix of $ A $ to solve the equation:

x = pinv(A) * b

It basically decompose the Matrix using the Singular Value Decomposition (SVD) such that $ U S {V}^{T} = A $ and the rebuild the matrix $ {A}^{\dagger} = U {S}^{\dagger} {V}^{T} $ where $ {S}^{\dagger} $ is the result of inverting any non zero element of the diagonal matrix $ S $.


The user @percusse said the following:

You just need A\b

You shouldn't use pinv for linear system solutions that's what I mean. Not only inefficient but also not accurate. There is no unique solution hence you can't talk about a least.

Backslash operator is optimized for all possible linear eq sets. Try it.

There is no least squares solutions for underdetermined systems because of the nontrivial nullspace. Anyways nevermind.

So, to make things clear.
The Least Squares solution always exists and it is also the Least Norm solution.
Namely, from the set $ \mathcal{B} = \left\{ x \mid \min_{x} \left\| A x - b \right\| \right\} $ there is one unique $ \hat{x} $ such that $ \hat{x} = \arg \min_{x} \left\| x \right\|_{2} \text{ subject to } \left\{ x \mid x \in \mathcal{B} \right\} $.

For instance:

for ii = 1:1e12
    numRows = 3;
    numCols = 27; %<! Must be bigger than 'numRows' for Under Determined System
    mA   = randn([numRows, numCols]);
    vB   = randn([numRows, 1]);
    vXLs = pinv(mA) * vB;
    vX   = mA \ vB;
    if(norm(vXLs) >= norm(vX))
    if(mod(ii, 1e4) == 0)
        disp(['Iteration Number - ', num2str(ii)]);


  • The difference arises for Under Determined System. For over / well determined system the solutions collide (Namely Tall Matrices which are the "Classic" case in LS problems).
  • The Psuedo Inverse is much slower to use then the \ operator. If you're looking for any solution, use \. If you're looking for the Least Squares solution (The Least Squares solution is always the least norm solution) for Under Determined System you need to use the pinv() function.
  • Proof and deeper analysis can be found in my Singular Value Decomposition (SVD) Presentation. Specifically this issue is on pages 50-60.
  • Cleve Moler's (Inventor of MATLAB) post on Pseudo Inverse vs. \ - The World’s Simplest Impossible Problem.
  • $\begingroup$ You just need A\b $\endgroup$
    – percusse
    Commented Aug 1, 2017 at 7:53
  • $\begingroup$ @percusse, No. That's wrong. This is not the Least Norm solution (At least not guaranteed to be) and hence not the Least Squares solution. This is a common mistake people do. $\endgroup$
    – Royi
    Commented Aug 1, 2017 at 15:26
  • $\begingroup$ You shouldn't use pinv for linear system solutions that's what I mean. Not only inefficient but also not accurate. There is no unique solution hence you can't talk about a least. $\endgroup$
    – percusse
    Commented Aug 1, 2017 at 15:33
  • $\begingroup$ In the case you look for LS solution in under determined system this is the only way to go. This is what the OP asked for. Of course that for full rank system you should use \. $\endgroup$
    – Royi
    Commented Aug 1, 2017 at 15:38
  • $\begingroup$ Backslash operator is optimized for all possible linear eq sets. Try it. $\endgroup$
    – percusse
    Commented Aug 1, 2017 at 15:44

The given matrix A is of rank 2, the third and the fourth columns are linear combinations of the first and second columns. So if the data matrix X is of rank lesser than or equal to the rank of A (i.e 2), you can recover the matrix X.

To know more about problems like this one, please read up on Orthogonal Matching Pursuit and Compressed sensing Matching Pursuit algorithms.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.