Years ago, after a casual conversation with my friend Chai about aging and powers of 2 (specifically $2^5$), I stumbled on a neat observation:

$$33 = 3 \cdot 11,\qquad 34 = 2 \cdot 17,\qquad 35 = 5 \cdot 7.$$

The numbers $33, 34, 35$ were all consecutive, distinct semiprimes. And it begged the question: what is the longest possible chain of consecutive semiprimes? (Spoiler alert: it’s 3).

Ok before we continue, some vocab to job our memories.

A semiprime is a product of exactly two primes. The two primes can be equal ($25 = 5 \cdot 5$) or different ($15 = 3 \cdot 5$).

A distinct semiprime is the second case: two different primes. So is 15 a distinct semiprime? Yes! Is 25 a distinct semiprime? No. For the purposes of this section, I say “semiprime” without qualifying it I mean the distinct kind. The non-distinct kind (squares of primes) breaks half the patterns we’re about to consider, so I’ll mostly ignore it.

A semiprime sandwich 🥪 is what I call sequences such as $(33, 34, 35)$. Consecutive, distinct semiprimes, of length 3 (which, as we will show soon enough, is the longest possible chain of consecutive semiprimes. And the first occurance of such a sequence happens to be: $(33, 34, 35)$.

why three is the ceiling

Time to actually show why 3 is the max. Note: I’m not labeling this a proof. The reasoning below stays pretty informal, but it’s tight enough to convince.

The claim: there’s no four consecutive integers $(n, n+1, n+2, n+3)$ where every one of them is a distinct semiprime.

Two cases to consider, depending on whether we anchor on an odd or even distinct semiprime.

Case 1: $n$ is an odd distinct semiprime. Both neighbors $n - 1$ and $n + 1$ are even.

Now suppose $n - 1$ is also a distinct semiprime. The only way an even number is a distinct semiprime is for its factorization to look like $2 \cdot a$ for some odd prime $a$. So $n - 1 = 2a$, which means $n = 2a + 1$.

Look at $n + 1$. It’s $2a + 2 = 2(a + 1)$. But $a$ is an odd prime, so $a + 1$ is even, which means $n + 1 = 2 \cdot (\text{even})$, and that’s divisible by 4. A distinct semiprime can’t be divisible by 4: that puts $2^2$ in the factorization (the prime 2 appearing twice), which breaks “distinct.”

So $n + 1$ is not a distinct semiprime. The takeaway: given $n$ an odd distinct semiprime, either $n - 1$ or $n + 1$ is a distinct semiprime, but not both.

Case 2: $n$ is an even distinct semiprime. Same logic as above: the only way an even number is a distinct semiprime is $n = 2k$ where $k$ is an odd prime.

Look at $n - 2 = 2k - 2 = 2(k - 1)$. Since $k$ is an odd prime, $k - 1$ is even, so $n - 2$ is divisible by 4. Same kill, can’t be a distinct semiprime.

By the same logic, $n + 2 = 2(k + 1)$ is divisible by 4 too. So neither $n - 2$ nor $n + 2$ is a distinct semiprime.

Putting it together. Among any four consecutive integers, exactly one of them is divisible by 4 (residues mod 4 cycle through $\{0, 1, 2, 3\}$, hitting each one once). That divisible-by-4 slot can’t be a distinct semiprime. Case 1 catches it if we anchored on an odd, Case 2 catches it if we anchored on an even. Either way the chain breaks at that slot.

So no four-in-a-row, and the max chain length is 3. ∎

By the way: a useful side effect of the proof is that the middle slice of any sandwich is always the even one. The two outside slices are odd (each a product of two different odd primes), and the middle is the unique even distinct semiprime in the triple.

That smallest sandwich $(33, 34, 35)$ also happens to land in a real chunk of human ages. So if you’re 33, 34, or 35 right now: you’re literally inside a sandwich. (Next sandwich birthday after that? See Puzzle 2 below. Heads up: it’s a long wait.)

Five puzzles. They get harder.

Where does the factor of 3 live in a sandwich? Let me check the first three:

$$\begin{aligned} (33, 34, 35) &= (3 \cdot 11,\ 2 \cdot 17,\ 5 \cdot 7) \\ (85, 86, 87) &= (5 \cdot 17,\ 2 \cdot 43,\ 3 \cdot 29) \\ (93, 94, 95) &= (3 \cdot 31,\ 2 \cdot 47,\ 5 \cdot 19) \end{aligned}$$

In $(33, 34, 35)$ it’s the bottom. In $(85, 86, 87)$ it’s the top. In $(93, 94, 95)$ back in the bottom. The factor of 3 jumps around between the two outer slices.

Why never the middle? We showed in Puzzle 3 of §1: the middle of any sandwich is never divisible by 3. So the factor of 3 has exactly two places it can live: the bottom or the top. That gives us the binary classification this section is about.

A semiprime sandwich is Class A if 3 divides the bottom slice, and Class B if 3 divides the top slice. Every sandwich is exactly one of these. Going by the table above: $(33, 34, 35)$ is Class A, $(85, 86, 87)$ is Class B, $(93, 94, 95)$ is Class A.

the linear couplings

Each class has a clean algebraic shape. The middle is always $2p$ for $p$ an odd prime (Puzzle 1 of §1). The factor-of-3 slice is $3b$ for $b$ an odd prime $\neq 3$. Lining the two facts up:

Class	bottom	middle	top
A	$3b$	$2p$	$3b + 2$
B	$3b - 2$	$2p$	$3b$

Class A: the middle is exactly one more than the bottom, giving $2p = 3b + 1$. Class B: the middle is exactly one less than the top, giving $3b = 2p + 1$.

These are the linear couplings. They tie the two auxiliary primes ($p$ in the middle, $b$ in the factor-of-3 slice) together with one easy equation each. Pretty much every modular constraint we’ll see falls out by reducing one of these mod something small.

detect the class from p alone

Small but useful observation. If you know just the middle prime $p$, you can decide the class by looking at $p \pmod 3$.

square sandwiches are always Class B

A square sandwich is one whose bottom slice is the square of a prime: $(r^2, r^2 + 1, r^2 + 2)$ for $r$ prime. The motivating example $(121, 122, 123)$ is one. Most of the rest of this pamphlet is about square sandwiches specifically.

Quick fact about them: every square sandwich is Class B.

If it were Class A, then 3 would divide the bottom $r^2$, which means 3 divides $r$ (since 3 is prime). The only prime divisible by 3 is 3 itself, giving the triple $(9, 10, 11)$. But 11 is prime, not a semiprime. So no Class A square sandwich exists.

Square sandwiches are always Team Class B.

you can’t pack two sandwiches back-to-back

How tightly can two sandwiches sit? Could you have $(n, n+1, n+2)$ and $(n+3, n+4, n+5)$? Six distinct semiprimes in a row?

No. The middles would be $n+1$ and $n+4$, both even (Puzzle 1 of §1). But they differ by 3, which is odd. Two evens can’t differ by an odd number. Contradiction.

So sandwiches always have at least one non-semiprime integer between them. The tightest possible configuration is the twin sandwich: $(n, n+1, n+2)$ and $(n+4, n+5, n+6)$, separated by the lone integer $n+3$. Six distinct semiprimes among seven consecutive integers.

We met the smallest twin sandwich in Puzzle 5 of §1: $(213, 214, 215)$ and $(217, 218, 219)$, with separator $216 = 2^3 \cdot 3^3$.

twin sandwiches are at least as hard as twin primes

Punchline of this section. Proving twin sandwich infinitude is at least as hard as the famous twin prime conjecture (open since 1849).

Suppose $(n, n+1, n+2)$ and $(n+4, n+5, n+6)$ are both sandwiches. Each carries a factor of 3 in one of its outer slices (Class A or Class B). The two factor-of-3 slices land at one of these distances apart on the integer line:

sandwich 1	sandwich 2	factor-of-3 distance
Class A (3 in bottom $n$)	Class A (3 in bottom $n+4$)	4
Class A (3 in bottom $n$)	Class B (3 in top $n+6$)	6
Class B (3 in top $n+2$)	Class A (3 in bottom $n+4$)	2
Class B (3 in top $n+2$)	Class B (3 in top $n+6$)	4

The two factor-of-3 slices look like $3b_1$ and $3b_2$ with $b_1, b_2$ primes. So $3(b_2 - b_1)$ has to be one of $\{2, 4, 6\}$. Only $3(b_2 - b_1) = 6$ has an integer solution, giving $b_2 - b_1 = 2$.

So the two $b$-primes form a twin prime pair. An infinite family of twin sandwiches gives you an infinite family of twin primes for free.

The twin prime conjecture is open. So twin sandwich infinitude is open in a strictly harder way.

but square sandwiches don’t twin

Surprise: square sandwiches never appear in twin configuration. Empirically, zero clusters across all 4,668 square sandwiches with $r < 10^7$.

For the rest of this pamphlet I’ll narrow focus to square sandwiches: sandwiches whose bottom slice is the square of a prime. Like $(121, 122, 123)$ where $121 = 11^2$.

Two facts about square sandwiches we already have:

1. The bottom is $r^2 = r \cdot r$, a non-distinct semiprime. (That’s OK. The original “distinct” condition is really about the middle and top, and we’ll keep enforcing it there.)
2. The sandwich is Class B (we proved this in §2). The factor of 3 lives in the top.

So for any prime $r$, the candidate sandwich looks like

$$(r^2,\ r^2 + 1,\ r^2 + 2)$$

with $r^2 + 1 = 2p$ ($p$ prime, since the middle is $2p$ from Puzzle 1 of §1) and $r^2 + 2 = 3b$ ($b$ prime, Class B). Solving:

$$p = \frac{r^2 + 1}{2}, \qquad b = \frac{r^2 + 2}{3}.$$

Subtract the equations: $(r^2 + 2) - (r^2 + 1) = 3b - 2p$, giving $3b = 2p + 1$. That’s the Class B linear coupling we already met in §2.

The whole sandwich is determined by $r$. The sandwich exists iff $r$, $p$, and $b$ are all prime.

the smallest examples by hand

Let me walk through small primes $r$ and compute $p, b$:

$r$	$r^2$	$p$	$b$	sandwich?
3	9	5	$11/3$ (not integer)	✗
5	25	13	9 (not prime)	✗
7	49	25 = 5² (not prime)	17	✗
11	121	61	41	✓
13	169	85 = 5·17 (not prime)	57 = 3·19 (not prime)	✗
17	289	145 = 5·29 (not prime)	97	✗
19	361	181	121 = 11² (not prime)	✗
23	529	265 = 5·53 (not prime)	177 = 3·59 (not prime)	✗
29	841	421	281	✓

Sparse. Among $r$ up to 29, only 11 and 29 produce sandwiches. So 121 and 841 are the first two prime squares that center a sandwich.

modular constraints, slowly

Some patterns are visible in the table. Several $r$ values fail in similar ways. Which $r$ can possibly produce a sandwich at all?

The trick is to take each defining equation and reduce it modulo a small prime. Three reductions are enough to pin down the picture: mod 3, mod 4, mod 5.

mod 4

$r$ is an odd prime, so $r \equiv 1$ or $3 \pmod 4$. In either case $r^2 \equiv 1 \pmod 8$ (just check $r = 1, 3, 5, 7$: their squares are $1, 9, 25, 49$, all $\equiv 1 \pmod 8$).

So $r^2 + 1 \equiv 2 \pmod 8$, which means $p = (r^2 + 1)/2 \equiv 1 \pmod 4$.

The middle prime $p$ is always one more than a multiple of 4.

mod 3

$r$ is prime and $r \neq 3$, so $r \equiv 1$ or $2 \pmod 3$. Squaring kills the sign: $r^2 \equiv 1 \pmod 3$.

So $r^2 + 1 \equiv 2 \pmod 3$, and $p = (r^2 + 1)/2$. To divide by 2 mod 3, multiply by 2’s inverse, which is 2 (since $2 \cdot 2 \equiv 1 \pmod 3$). So $p \equiv 4 \equiv 1 \pmod 3$.

This also matches what §2 said: square sandwiches are Class B, and Class B forces $p \equiv 1 \pmod 3$.

mod 5

$r$ is prime and $r \neq 5$, so $r \equiv 1, 2, 3,$ or $4 \pmod 5$. Squaring: $r^2 \in \{1, 4\} \pmod 5$.

If $r^2 \equiv 4 \pmod 5$, then $r^2 + 1 \equiv 0 \pmod 5$, which means $5 \mid 2p$, so $5 \mid p$. But $p$ is prime $> 5$, contradiction. So $r^2 \equiv 1 \pmod 5$.

Then $r^2 + 1 \equiv 2 \pmod 5$. Inverse of 2 mod 5 is 3 (since $2 \cdot 3 = 6 \equiv 1 \pmod 5$). So $p \equiv 6 \equiv 1 \pmod 5$.

gluing them together

We’ve shown $p \equiv 1 \pmod 4$, $p \equiv 1 \pmod 3$, $p \equiv 1 \pmod 5$. The Chinese Remainder Theorem (CRT) says: residue conditions modulo coprime numbers can be glued. If $m_1, m_2, m_3$ are pairwise coprime, knowing $a$ mod each pins down $a$ mod $m_1 m_2 m_3$.

Here 3, 4, 5 are pairwise coprime, with product 60. Each condition is "$p \equiv 1$", which glues to:

the four lanes mod 30

The mod-5 step pinned $r^2 \equiv 1 \pmod 5$, which means $r \equiv 1$ or $4 \pmod 5$. Combine with the obvious ”$r$ is prime $> 5$” (so $r$ coprime to 2, 3, 5), and we can read off all surviving residues modulo $30 = 2 \cdot 3 \cdot 5$.

Primes $> 5$ live in residue classes coprime to 30, which are $\{1, 7, 11, 13, 17, 19, 23, 29\} \pmod{30}$. Eight classes. Demanding $r \equiv 1$ or $4 \pmod 5$ kills four of them, leaving:

the constraint on b

A similar analysis on $b$ (using $r^2 + 2 = 3b$, which rearranges to $r^2 \equiv -2 \pmod b$) gives:

$$b \equiv 1 \text{ or } 17 \pmod{24}.$$

The key ingredient is that $-2$ has to be a quadratic residue modulo $b$: there has to be some integer $r$ with $r^2 \equiv -2 \pmod b$. By the supplementary laws of quadratic reciprocity (don’t worry, you don’t need to know the proof to use the result), this happens iff $b \equiv 1$ or $3 \pmod 8$. Combined with the mod-3 piece (Class B forces $b \not\equiv 0 \pmod 3$), the answer mod 24 is exactly $\{1, 17\}$.

We’ll come back to this in §6 with a much cleaner explanation in the language of the ring $\mathbb{Z}[\sqrt{-2}]$.

the brute force says

Searching primes $r < 10{,}000$ in the four valid lanes turns up 35 sandwiches. The split between $b \equiv 1$ and $b \equiv 17 \pmod{24}$ runs 18 to 17. Nearly even. About 5.7% of primes in valid lanes actually produce a sandwich. Sparse but not vanishingly so.

The two clean small examples again:

• $r = 11$: $p = 61$, $b = 41$. Triple $(121, 122, 123)$.
• $r = 29$: $p = 421$, $b = 281$. Triple $(841, 842, 843)$.

This is the part that surprised me. To say what’s going on we need a quick detour through the Gaussian integers.

a number system one step beyond ℤ

The Gaussian integers, written $\mathbb{Z}[i]$, are the complex numbers of the form

$$a + bi, \qquad a, b \in \mathbb{Z},$$

where $i$ is the imaginary unit ($i^2 = -1$). You can add and multiply Gaussian integers using the usual rules, treating $i$ as a formal symbol that squares to $-1$. Examples:

$$(2 + 3i) + (1 - i) = 3 + 2i, \qquad (2 + 3i)(1 + i) = 2 + 2i + 3i + 3i^2 = -1 + 5i.$$

So $\mathbb{Z}[i]$ is a number system you can add, multiply, and (with care) factor in. Just like the ordinary integers, but in two dimensions instead of one. Geometrically: a lattice of points in the complex plane.

The role that “absolute value” plays for $\mathbb{Z}$ is played here by the norm:

$$N(a + bi) = (a + bi)(a - bi) = a^2 + b^2.$$

The norm is always a non-negative integer, and crucially it’s multiplicative:

$$N(\alpha \beta) = N(\alpha) N(\beta).$$

So if $\alpha \in \mathbb{Z}[i]$ factors as $\alpha = \beta \gamma$, then $N(\alpha) = N(\beta) N(\gamma)$. The norm factors too. That’s the lever for everything below: knowing how integer primes factor in $\mathbb{Z}[i]$ tells you about norms, and knowing about norms tells you about sums of squares.

Gaussian primes and Fermat’s two-squares theorem

A Gaussian prime is the analogue of an ordinary prime: a Gaussian integer $\pi$ that can’t be written as $\pi = \alpha \beta$ for non-trivial $\alpha, \beta$ (where “non-trivial” means not multiples of $\pm 1$ or $\pm i$, the four units in $\mathbb{Z}[i]$).

The structure of Gaussian primes is governed by Fermat’s old theorem on sums of two squares.

So whenever you spot a prime $p \equiv 1 \pmod 4$, you also spot a Gaussian prime $\pi$ of norm $p$ sitting in the lattice. Now back to sandwiches.

the centered-square identity

Set $r = 2k + 1$ (every odd $r > 2$ has this form, with $k = (r - 1)/2$), and look at the middle prime of the sandwich:

$$p = \frac{r^2 + 1}{2} = \frac{(2k + 1)^2 + 1}{2} = \frac{4k^2 + 4k + 2}{2} = 2k^2 + 2k + 1.$$

Now watch this:

$$2k^2 + 2k + 1 = k^2 + (k^2 + 2k + 1) = k^2 + (k + 1)^2.$$

So the middle prime is always a sum of two consecutive squares.

Spot check. $r = 11$, so $k = 5$, so $p = 5^2 + 6^2 = 25 + 36 = 61$. ✓ $r = 29$, so $k = 14$, so $p = 14^2 + 15^2 = 196 + 225 = 421$. ✓

the Gaussian integer π_k

Whenever $p$ is a sum of two squares, $p$ shows up as a norm in $\mathbb{Z}[i]$. In our case:

$$N(k + (k + 1)i) = k^2 + (k + 1)^2 = p.$$

So we have a Gaussian integer

$$\pi_k = k + (k + 1)i \in \mathbb{Z}[i]$$

whose norm is the middle prime. And whenever $p$ is actually prime in $\mathbb{Z}$, the element $\pi_k$ is a Gaussian prime. (Why $p \equiv 1 \pmod 4$, which Fermat’s theorem requires: exactly one of $k, k+1$ is even, so $p$ is even-square + odd-square $\equiv 0 + 1 \equiv 1 \pmod 4$.)

Geometrically, $\pi_k$ sits one step above the diagonal in the lattice. On the line $y = x + 1$.

the (1+i) reformulation

Here’s a small but beautiful identity. In $\mathbb{Z}[i]$:

$$(1 + i)\pi_k = (1 + i)(k + (k+1)i) = k + (k+1)i + ki + (k+1)i^2 = (k - (k+1)) + (k + (k+1))i = -1 + (2k+1)i.$$

But $2k + 1 = r$. So:

A one-parameter Gaussian-arithmetic encoding of the entire square-sandwich structure. Every valid $r$ corresponds to a Gaussian prime $\pi_k$ such that $(1 + i)\pi_k = -1 + ri$, with $r = 2k + 1$.

the geometric picture

The argument (angle) of $\pi_k = k + (k+1)i$ is

$$\arg(\pi_k) = \arctan\!\left(\frac{k+1}{k}\right) \to \arctan(1) = \frac{\pi}{4} = 45° \quad\text{as } k \to \infty.$$

So sandwich-generating Gaussian primes don’t appear randomly across the lattice. They spiral toward the diagonal $y = x$ as $k$ grows, hugging the line $y = x + 1$ from above and landing closer and closer to a 45° angle.

$r$	$k$	$\pi_k$	$\arg(\pi_k)$
11	5	$5 + 6i$	$50.19°$
29	14	$14 + 15i$	$46.97°$
79	39	$39 + 40i$	$45.73°$
271	135	$135 + 136i$	$45.21°$

Picture: an infinite sequence of dots on the line $y = x + 1$, starting visibly above the diagonal and slowly settling onto it, picking out the centered-square primes as you go. The places where the dot isn’t a Gaussian prime (like $\pi_3 = 3 + 4i$) are exactly where the would-be middle prime $p$ turns out to be composite.

why r = 7 fails, Gaussian-style

At $r = 7$, we have $k = 3$, so $\pi_3 = 3 + 4i$. The norm is $N(\pi_3) = 9 + 16 = 25 = 5^2$. By Puzzle 1 above, since $25 = 5^2$ isn’t a rational prime, $\pi_3$ isn’t a Gaussian prime either.

Concretely, $\pi_3$ factors as

$$\pi_3 = (2 + i)^2.$$

(Quick check: $(2 + i)^2 = 4 + 4i + i^2 = 4 + 4i - 1 = 3 + 4i$. ✓) The factor $(2 + i)$ is itself a Gaussian prime (norm $5$, prime), so $\pi_3$ is a Gaussian prime squared, not a Gaussian prime itself.

The would-be middle prime at $r = 7$ is $p = N(\pi_3) = 25$, which is composite. The $\mathbb{Z}[i]$ side and the $\mathbb{Z}$ side fail in lockstep: $\pi_3$ doesn’t stay irreducible in $\mathbb{Z}[i]$, $p$ doesn’t stay prime in $\mathbb{Z}$. There’s a deeper reason this happens at $r = 7$ specifically (it’s a solution to the negative Pell equation $r^2 - 2m^2 = -1$ with $m = 5$). More on that in §6.

the small-k failures

The construction starts working at $k = 5$ (giving $r = 11$). For $k < 5$, at least one of the three sandwich conditions fails.

The first ten valid $k$ values:

Corresponding middle primes: $61, 421, 3121, 36721, 71821, \ldots$. All centered square primes (forced by the construction).