Why Your Math Proofs Get Stuck After the First Few Steps

When researchers compared undergraduates with doctoral students constructing proofs in abstract algebra, the undergraduates were often aware of the facts a proof needed, and able to apply them, and still could not assemble it. What the doctoral students had was a sense of which facts and theorems to reach for^[1].

Getting started was also one of three main sources of trouble a researcher recorded when he observed and interviewed sixteen undergraduate mathematics and mathematics education majors learning to write formal proofs. The other two were concept understanding, and mathematical language and notation^[2].

So the blank page may not be showing you a content gap. It may be a selection problem^[1].

A published proof is not a record of how it came about: the scribbles, false starts and backward reasoning are gone. You see the clean route and assume you alone got stuck.

Beginning undergraduates and working mathematicians judged whether written proofs were correct while their eye movements were recorded. The beginners spent proportionately more time on what the researchers called the surface features of the arguments, a sign they attended less to logical structure. The mathematicians were more inclined to shift back and forth between consecutive lines, which the researchers took as a sign they were working out why each step was allowed^[3].

They were reading proofs somebody else wrote, not producing their own, so this is not direct evidence about the blank page^[3]. The move still works on your own draft: put your finger on line five, ask what makes it legal, and point at every earlier line it leans on.

Three mathematics educators describe a technique for the student facing a blank page: use the logical structure of the theorem and its definitions to write the frame first, then fill in the middle. They explain the method but report no test of it, so treat it as teachers' advice, not a finding^[4].

Prove: if $n$ is even, then $n^2$ is even.

"Even" means "two times an integer". That definition writes your first line, $n = 2k$ for some integer $k$. Applied to the conclusion instead, it writes your last line: you must arrive at $n^2 = 2m$ for some integer $m$. Only now does the middle have a shape to fill.

$$n^2 = (2k)^2 = 4k^2 = 2(2k^2)$$

Take $m = 2k^2$, an integer because $k$ is. What made the middle fillable was knowing the exact form the last line had to take.

If the general form will not come, run one small case first. For $n = 4$, $k = 2$ and $n^2 = 16 = 2 \times 8$; for $n = 6$, $k = 3$ and $n^2 = 36 = 2 \times 18$. The halves $8 = 2 \times 2^2$ and $18 = 2 \times 3^2$ hand you the $2k^2$.

Prove: if $a$ divides $b$ and $b$ divides $c$, then $a$ divides $c$.

Here $a \mid b$ means $b = ak$ for some integer $k$. The assumptions become $b = ak_1$ and $c = bk_2$, with $k_1$ and $k_2$ integers. The conclusion, unpacked the same way, must read $c = am$ for some integer $m$. Substituting:

$$c = bk_2 = (ak_1)k_2 = a(k_1k_2)$$

so $m = k_1k_2$, an integer. With numbers: $a = 3$, $b = 12$, $c = 60$ gives $k_1 = 4$, $k_2 = 5$, $m = 20$, and $3 \times 20 = 60$.

Writing the definition down beats trusting that you remember it: concept understanding and notation were two of the three difficulties in that study of sixteen undergraduates^[2]. The teachers behind the skeleton technique add that a definition must become what they call "operable" before you can unpack the conclusion^[4].

In one small study, thirteen undergraduates were shown an argument whose last line was true but did not follow from the lines above it. Only six saw it was not a valid proof, and only two of those for legitimate mathematical reasons. Asked directly whether the last line followed from what came before, ten of the thirteen rejected it^[5].

That ten is the hopeful number: the judgment was there, it just needed the question asked out loud^[5]. Ask it of your own page, line by line.

The divisibility proof shows a milder version. Suppose you write $b = ak_1$ and $c = bk_2$, then jump to "therefore $c = am$ for some integer $m$". That line is true, and it does follow. But the work that earns it, the substitution $c = (ak_1)k_2 = a(k_1k_2)$ and the witness $m = k_1k_2$, never appears. That is a gap in the write-up, not a false step. The argument in the study was the harder case, where the true last line did not follow at all^[5]. On the page both look alike: the destination is right either way.

Justifying lines will not catch the opposite failure, an assumption you never used. Did every hypothesis do work somewhere? An unused one usually means you proved something weaker than what was asked.

In three experiments, students worked through a short booklet that trained them to notice how the lines of a proof connect. They went on to understand proofs better than students without the training^[6].

One caution, and not a small one: those experiments measured comprehension, how well students understood a proof they were handed, not whether they wrote better proofs of their own^[6].

The habit itself is small. After each line, say what it means and why it follows from what is already on the page^[6]. If you cannot, you have found the hole.

Every learner in these studies was an undergraduate, most in abstract algebra or real analysis^[1][2][3][5]; those they were compared with were doctoral students and working mathematicians^[1][3]. The booklet experiments name their participants only as students^[6]. None of it is evidence about high-school students.

Three of the studies are about reading and judging a proof, not writing one^[3][6][5]. No study here shows a training that improved the proofs students write^[6].

The two studies about writing are small, old and purely descriptive: one observed and interviewed sixteen students^[2], the other compared a few undergraduates with doctoral students through interviews^[1]. They describe what beginners are missing, not what fixes it. The skeleton technique has no results attached at all^[4], and no review pools these studies together, so nothing above is settled consensus^{[1][2][3][6][5][4]}.

• When nothing comes, list on a separate sheet every definition, property and theorem the problem touches: a selection problem needs its candidates written down, and the missing link is often something you already knew^[1].
• Take a proof you have finished and annotate it, one sentence per line: this follows because.
• Take a theorem you have not started and write its first and last lines, unpacked from the definitions, before attempting the middle.
• Read a proof asking "does line $n$ follow from the lines above it?" rather than "is line $n$ true?", and name which earlier lines it uses.
• When a proof closes, write down the step that connected start to end. Nobody has tested whether banking those moves builds selection sense, but it is where to look.

None of this makes proofs fast. It makes the stuck moment diagnosable: you stop asking whether you are good at proofs and start asking which line you cannot justify.