Mental math is pattern recognition

Good mental math is not fast calculation. It is problem rewriting.

When someone computes 94 × 98 in a few seconds, they are almost certainly not running through columns of digits. They've noticed that both numbers sit just under 100, used that structure to simplify the work, and arrived at the answer with very little arithmetic. What looks like speed is usually a better choice of representation.

This is the premise behind every strategy family in our catalog.

Raw simulation is the wrong baseline

Most people, when they try mental math, attempt to simulate written arithmetic in their heads. They carry digits, track columns, hold partial products in working memory. This is genuinely hard. Working memory is small, and written arithmetic was explicitly designed to offload its burden onto paper. Trying to run it without paper isn't a test of mental math — it's a test of how well your brain can pretend to be scratch paper.

Take 213 − 67. The written-arithmetic path means borrowing from the tens column, and most people lose track somewhere around the second digit. But there's a simpler question buried in the problem: how far is it from 67 to 213? Count up: 67 → 70 is 3, 70 → 200 is 130, 200 → 213 is 13. Add 3 + 130 + 13 = 146. No borrowing.

You could also shift both numbers by 3: (213 + 3) − (67 + 3) = 216 − 70 = 146. Or shift by 13 the other way: 200 − 54 = 146. Three reframings, same answer, all easier than the algorithm. The skill isn't computing — it's choosing.

The major moves

A handful of structural moves cover most of everyday mental math. Each is a family in our catalog, not a one-off trick.

Decomposition breaks a problem into friendlier pieces. 47 + 36 becomes 47 + 30 + 6. 18 × 25 becomes 9 × 50 = 450, because halving one factor and doubling the other preserves the product. We train this in halve-and-double.

Compensation rounds to a clean number and corrects. 49 + 38 is 50 + 38 − 1. 201 − 98 is 201 − 100 + 2. The round numbers do the work; the correction is small. Two decks live here: compensation-add and compensation-subtract.

Counting up is what subtraction often wants to be. Rather than 213 − 67, ask: what do I add to 67 to reach 213? Cashiers count up; they don't borrow. The counting-up deck biases toward the shapes where this move is the cleanest path.

Benchmarking makes percentages tractable. 10% is a decimal shift. 15% of 80 is 8 + 4 = 12. 37% of 200 is 25% (50) + 10% (20) + 2% (4) = 74. You're not computing 37% directly — you're assembling it from anchors you already know. benchmark-percentages trains exactly this assembly. The fluency deck fraction-decimal-anchors makes the anchors themselves cheap.

Near-round multiplication applies when both factors sit near a convenient anchor. 94 × 98: both near 100, deficits 6 and 2. 94 − 2 = 92, then 6 × 2 = 12, giving 9,212. One identity; the recognition is what matters. near-round-mul is the family.

Estimation first changes how you approach every problem. Commit to a range before computing: 47 × 53 is close to 50 × 50 = 2,500. When you get 2,491, you're in the right zone. When you get 24,910, you know something went wrong. We train this as its own pillar: x-near-anchors, div-near-anchors, pct-from-anchors.

There is rarely just one path

These moves are not a menu where each problem maps to one entry. Most problems yield to several, and part of fluency is noticing that.

18 × 25 has at least two clean paths. Halve-and-double: 9 × 50 = 450. Compensation: 20 × 25 − 2 × 25 = 450. Either beats the long-multiplication algorithm; what matters is that you saw a path immediately.

49 + 38 is similar. Compensate on 49: 50 + 38 − 1 = 87. Compensate on 38: 49 + 40 − 2 = 87. A fluent calculator scans for the cleanest entry and commits. We don't try to teach which transformation belongs to which problem — we try to make enough of them automatic that you always have options.

Why this looks like talent

The restructuring step is invisible to observers. You see the answer; you don't see the 49 → 50 move that preceded it. What gets attributed to natural ability is usually just a person choosing a better representation faster than you noticed they were choosing.

Which has a direct training implication: you don't get there by drilling raw computation faster. You get there by seeing patterns repeatedly, in many different number dresses, until the recognition runs ahead of the calculation. That's the whole logic behind organizing strategy training as families. Each deck in compensation-add or near-round-mul is the same move with twenty different number costumes — exposure is the mechanism.

The milestone isn't being able to multiply large numbers at a party. It's quieter: you stop reading 94 × 98 as a multiplication problem and start reading it as a near-100 problem before the question has finished forming. At that point, the work is mostly done.

This is a product-context distillation. The long-form essay lives on mingjerlee.com →.