Solving Multi-Step Linear Equations

A multi-step linear equation has variables on both sides, parentheses to distribute, and like terms to combine. Mastering this is the prerequisite to every algorithm in machine learning that involves “isolating a parameter” — which is most of them.

The procedure

There’s a single, mechanical procedure that solves every multi-step linear equation. Follow it in order; do not skip steps mentally.

1. Distribute every coefficient across its parentheses. Watch the sign on the coefficient carefully — a negative outside the parentheses changes the sign of every term inside.
2. Combine like terms on each side of the equation independently.
3. Move all variable terms to one side by adding or subtracting.
4. Move all constant terms to the other side by adding or subtracting.
5. Divide by the coefficient on the variable.

Worked example

Solve for $x$:

Step 1 — distribute. The two coefficients on the left are $3$ and $-2$; on the right, the implicit coefficient on $-(x-6)$ is $-1$.

• $3(2x-4)=6x-12$
• $-2(x+5)=-2x-10$ ← the negative carries to both terms
• $-(x-6)=-x+6$ ← same: the negative flips both signs

After distribution:

Step 2 — combine like terms.

Step 3 — variable to one side. Subtract $3x$:

Step 4 — constant to the other side. Add $22$:

The classic trap

The single most common error in this kind of problem is failing to distribute a negative sign. There are two specific places it bites:

• $-2(x+5)$ is $-2x-10$, not $-2x+10$.
• $-(x-6)$ is $-x+6$, not $-x-6$.

If you wrote the wrong distribution and proceeded, you might land at $x=7$ or some other plausible-looking number. The math is internally consistent — it’s just solving the wrong equation.

Why this matters for ML

1. Algebraic structure learned. Distribute, combine, isolate — a deterministic procedure for solving any linear equation, with sign discipline as the non-negotiable habit.
2. ML object with the same structure. Backpropagation cascades the chain rule across layers; each step distributes partial derivatives across sums and products — the same algebra, scaled to thousands of terms per layer.
3. Computational payoff. A single dropped negative sign flips a gradient and pushes weights away from the optimum: training diverges silently. The sign-discipline muscle you train here is what keeps a from-scratch two-layer network actually learning instead of drifting.