TITLE: Hahn-Banach From Systems of Linear Equations
QUESTION [48 upvotes]: In this paper1 on the history of functional analysis, the author mentions the following example of an infinite system of linear equations in an infinite number of variables $c_i = A_{ij} x_j$:
\begin{align*}
\begin{array}{ccccccccc}
1 & = & x_1 & + & x_2 & + & x_3 & + & \dots  \\
1 & = &     &   & x_2 & + & x_3 & + & \dots \\
1 & = &     &   &     &   & x_3 & + & \dots \\
  & \vdots  &   &     &   &     &   &   &   \ddots 
\end{array} \to \begin{bmatrix} 1 \\ 1 \\ 1 \\ \vdots \end{bmatrix} = \begin{bmatrix}
1 & 1 & 1 & \dots \\
  & 1 & 1 & \dots \\
  &   & 1 & \dots \\
  &   &   & \ddots
\end{bmatrix} \begin{bmatrix} x_1 \\ x_2 \\ x_3 \\ \vdots \end{bmatrix} 
\end{align*}
as an example of a system such that any finite truncation of the system down to an $n \times n$ system has a unique solution $x_1 = \dots = x_{n=1} = 0, x_n = 1$ but for which the full system has no solution. 
This book2 has the following passage on systems such as this one:

The Hahn-Banach theorem arose from attempts to solve infinite systems of linear equations... The key to the solvability is determining "compatibility" of the system of equations. For example, the system $x + y = 2$ and $x + y = 4$ cannot be solved because it requires contradictory things and so are "incompatible". The first attempts to determine compatibility for infinite systems of linear equations extended known determinant and row-reduction techniques. It was a classical analysis - almost solve the problem in a finite situation, then take a limit. A fatal defect of these approaches was the need for the (very rare) convergence of infinite products."

and then mentions a theorem about these systems that motivates Hahn-Banach:

Theorem 7.10.1 shows that to solve a certain system of linear equations, 
  it is necessary and sufficient that a continuity-type condition be satisfied. 
Theorem 7.10.1 (The Functional Problem): Let $X$ be a normed space over $\mathbb{F} = \mathbb{R}$ or $\mathbb{C}$, let $\{x_s \ : \ s  \in S \}$ and $\{ c_s \ : \ s \in S \}$ be sets of vectors and scalars, respectively. Then there is a continuous linear functional $f$ on $X$ such 
  that $f(x_s) = c_s$ for each $s \in S$ iff there exists $K > 0$ such that 
  \begin{equation}
\left|\sum_{s \in S} a_s c_s \right| \leq K  \left\| \sum_{s \in S} a_s x_S \right\| \tag{1}, 
\end{equation}
  for any choice of scalars $\{a_s \ : \ s \in S \}$ for which $a_s = 0$ for all but finitely many $s \in S$ ("almost all" the $a_s = 0$). 
Banach used the Hahn-Banach theorem to prove Theorem 7.10.1 but Theorem 7.10.1 implies the Hahn-Banach theorem: Assuming that Theorem 7.10.1 holds, let $\{ x_s \}$ be the vectors of a subspace $M$, let $f$ be a continuous linear functional on $M$; for each $s \in S$, let $c_s = f(x_s)$. Since $f$ is continuous, $(1)$ is satisfied and $f$ possesses a continuous extension to $X$.

My question is: 

If you knew none of the theorems just mentioned, how would one begin from the system $c_i = A_{ij} x_j$ at the beginning of this post and think of setting up the conditions of theorem 7.10.1 as a way to test whether this system has a solution?
How does this test show the system has no solution? 
How do we re-formulate this process as though we were applying the Hahn-Banach theorem?
Does anybody know of a reference for the classical analysis of systems in terms of infinite products?


1Neal L. Carothers: A Brief History of Functional Analysis.
2Lawrence Narici, Edward Beckenstein: Topological Vector Spaces, 2nd Edition.

REPLY [4 votes]: I will try to provide some rather incomplete answers to your questions. 

Let us for simplicity, consider the finite case. Let us say we want to solve a $n\times n$ linear system of equations, i.e $Ax = b$, with all quantities in question being real. That is we are solving:
$$
\begin{matrix}
a_{11}x_{11} +\cdots + a_{1n}x_{1n} = b_1 \\ 
\vdots \\
a_{n1}x_{n1} +\cdots + a_{nn}x_{nn} = b_n
\end{matrix}
$$
Recall that $\mathbb{R}^n$ is self dual, with any linear functional (all which are continuous) in $\mathbb{R}^n$ given by $f_a(x) = \langle a , x\rangle = \sum_{k =1}^n a_kx_k$. Thus, in the case we consider here, solving a linear system amounts to finding a solution to:
$$
f_{a_1}(x) = b_1 \\ 
\vdots 
\\ 
f_{a_n}(x) = b_n
$$ 
Where $a_i = (a_{i1},\cdots a_{in}$). Thus, solving a simultaneous system of linear equations is equivalent to finding $x \in \mathbb{R}^n$ solving the above vector problem. Because $\mathbb{R}^n$ is reflexive, this is equivalent to picking $x \in \left(\mathbb{R}^n\right)^*$, and then solving the equation if given functionals $f_{a_1}, \cdots , f_{a_n}$.  With this framework in mind, it is useful to discuss when inconsistency occurs in a linear system. Let $X$ be a vector space over $\mathbb{R}$ and let $f \in X^*$, $\{x_s: s\in S\}\subset X$ and $c_s = f(x_s)$ for all $s$. Let $M = \mathrm{span}(x_s : s \in S)$. If the $x_s$ are linearly independent there are no issues. For any finite collection $x_{s_1}, \cdots x_{s_n}$, we can define $f(x_{s_k}) = c_{s_k}$, and  then extend this to $M$ by linearity. Everything is well defined on all of $M$, thanks to the uniqueness of representation given by linear dependence. However, we may have issues when the $x_s$ are not linearly independent. Specifically, let us say we have $x\in M$, with two representations:
$$
x = \sum_{s \in J} \alpha_s x_s \\ 
x = \sum_{t \in I} \beta_t x_t
$$
where $I,J \subset S$, finite. (The above sum can be indexed by new set, say $K = I \cup J$, and with scalars indexed by things not in $I\cap J$, we set them equal to $0$) We have:
$$
f(x) = \sum_{k \in K}\alpha_kc_k  = \sum_{k \in K}\beta_kc_k = f(x)
$$
Thus we must have:
$$
\left| \sum_{k \in J}(\alpha_k - \beta_k)c_k\right| = 0
$$
Now, we have no guarantees that this is true, as the values of $c_k$ are chosen a-priori.Thus, we need to constrain $f$ in some fashion in order to ensure well-definedness. The condition that gives us this is the specification:
$$
\left|\sum_{k \in K}\gamma_kc_k  \right | \leq K \left\|\sum_{k \in K} \gamma_k x_k\right\|
$$
for any finite set of scalars $\gamma_k$. In this case, we set $\gamma_k = \alpha_k - \beta_k$. In finite dimensions, this $K$ always exists, as we can express all vectors in terms of a common basis. In the general case, this cannot be done. Thus, in summary, the only condition which gives us well-definedness/consistency is indeed the above condition. 

The above result also implies Hahn Banach as the author says.