Practice: The Row and Column Pictures

(a) Row picture. Two lines in the plane of the unknowns, ${\mathbb{R}}^2$: the line $2x+y=7$ and the line $x-y=2$. Their normals, $(2,1)$ and $(1,-1)$, point in different directions, so the lines are not parallel and cross at exactly one point. The row picture shows where the solution sits.

(b) Column picture. The columns are $A_1=(2,1)$ and $A_2=(1,-1)$, both in ${\mathbb{R}}^2$, and the question is which weights $x$ and $y$ satisfy

The two columns point in independent directions, so together they reach every point of ${\mathbb{R}}^2$. The column picture shows why a solution exists, and that it is unique, since no combination is wasted.

(c) Solve. Adding the equations gives $3x=9$, so $x=3$ and then $y=1$.

Row check: $2(3)+1=7$ and $3-1=2$. The point $(3,1)$ is on both lines.

Column check: $3(2,1)+1(1,-1)=(6,3)+(1,-1)=(7,2)$. The weights $3$ and $1$ do reach $b$.

Same answer, both ways. They must agree. The two pictures group the same equation differently rather than describing different problems.

(d) Which picture answers existence. The column picture. Existence asks whether $b$ can be reached as a weighted combination of the columns, that is, whether $b$ lies in their span, which is the column question stated exactly. The row picture answers it only indirectly, by asking whether the lines happen to share a point. Choosing by the question rather than by habit is what makes the two readings useful rather than decorative.

First error: the count is wrong. A $4\times 7$ system has four equations in seven unknowns. The row picture therefore consists of four hyperplanes, not seven. The seven belongs to the columns. There are seven column vectors, one per unknown.

Second error: it cannot be sketched. Those four hyperplanes live in ${\mathbb{R}}^7$, the space of the unknowns. There is nothing to draw, and no diagram will be forthcoming however carefully the attempt is made.

Third error: choosing a picture by preference. The two readings answer different questions. Picking one by familiarity rather than by the question is what makes half of this subject feel harder than it is.

What to do instead. Reduce and read the pivots. With four equations there are at most four pivots, so at least three of the seven variables are free; if the system is consistent, its solution set is a flat of dimension at least three. That is a complete structural answer and it required no picture at all.

And if the question is existence. Use the column reading: seven vectors in ${\mathbb{R}}^4$ will usually span the whole space, precisely because there are more columns than rows, so a solution typically exists, and the interesting question becomes how many rather than whether.

The general lesson. Both pictures remain useful as ways of thinking long after they stop being drawable. The column reading in particular survives into high dimensions unchanged: which vectors can be reached is a sensible question in ${\mathbb{R}}^{100}$.

Choosing by the question. The deeper error is defaulting. Each picture answers a different question directly: the row picture shows where constraints meet, the column picture shows what can be reached. For a $4\times 7$ system the column reading is available and the row reading is not drawable at all, so the choice is settled by what is being asked, not by which one the engineer finds comfortable.

The shaded region is the row picture. Each constraint is a hyperplane, a line, in two variables, and the feasible region is what survives when every allowed side is intersected. The sliding line is a contour of the objective, and the whole method is carried out in the space of the unknowns, which is exactly where the row picture lives. The only change from the equality case is that inequalities keep one side of each flat rather than the flat alone.

The column swapping is the column picture. A set of columns is being asked what it can reach; swapping one for another changes the set and therefore changes what is reachable. That is a question about mixing vectors, asked in the space of the right-hand side, which is where the column picture lives.

Why one-sided fluency hurts. A learner who only has the row picture meets the column-swapping algorithm with no way to picture what a swap does. It looks like arbitrary bookkeeping on a table of numbers, and the reason a swap is even legal becomes opaque. A learner who only has the column picture meets the shaded region with no account of why the optimum is found at a corner, because corners are a fact about intersecting flats rather than about spans.

Both are the same system. The region and the column set describe one set of constraints. Being able to move between them means the two halves of the subject are one subject rather than two unrelated techniques that happen to share notation.

Which reading each technique needs. The graphical method asks where constraints meet, so it is the row picture drawn and slid across. The simplex method asks what a new mixture of columns reaches, so it is the column picture worked algebraically. Neither is the better reading; each is chosen because it answers the question its method poses.