The cheese cut (one plane through a cube of cubes)

Edition 3 · On this site since · last revised · Verify & co-author
Kind
Computer-assisted theorem for every size from 3 upward
Status
Proved for every n≥3n \ge 3: Lean-verified for n=3n = 3–3232 and n≥700n \ge 700; checked by certificates for n=33n = 33–699699; unrefereed
Prior work
The formula is Bárány and Frankl's own construction; the theorem that it cannot be beaten was not found in the searches logged
On this page
  1. The statement
  2. Check it yourself
  3. Worked examples
  4. Status and who checked it
  5. Prior work
  6. Further reading

The statement

Stack n3n^3 unit cubes into an n×n×nn \times n \times n cube and cut it once with a plane. A small cube counts as met when the plane passes through its inside: its 8 corners lie strictly on both sides of the plane. Let m(n)m(n) be the most small cubes one plane can meet, and let

F(n)=⌊9n2+4n−174⌋.F(n) = \left\lfloor \frac{9n^2 + 4n - 17}{4} \right\rfloor .

Theorem. m(n)=F(n)m(n) = F(n) for every n≥3n \ge 3. (At n=2n = 2, FF gives 6 but m(2)=7m(2) = 7.)

Verification split. The theorem is proved for every n≥3n \ge 3: it is Lean-verified for n=3n = 3–3232 (including n=10n = 10 at 230) and n≥700n \ge 700 in Challenge.lean, and checked by certificates for n=33n = 33–699699 by two independent exact-integer programs in the verification certificates (ZIP), for which Lean verifies the sufficiency criterion.

What is Bárány and Frankl's, and what is claimed. In the American Mathematical Monthly (2021) Bárány and Frankl proved m(n)≤9n2/4+2n+1m(n) \le 9n^2/4 + 2n + 1 and built a plane, perpendicular to the direction (n−1,n−1,n)(n - 1,\allowbreak n - 1,\allowbreak n), that meets exactly F(n)F(n) cubes. So the formula is their construction, not a fit. The claim is that their construction cannot be beaten, for any nn. It follows that their remark, made without proof, that the construction "can be further improved by 2" for large nn is false for every nn.

The printed question it answers. Of the 10×10×1010 \times 10 \times 10 box they wrote: "at least 230 and at most 246 pieces ... The exact number is unknown." It is 230, by exhaustive computation, and the theorem gives the exact answer F(n)F(n) for every box from n=3n = 3 up.

Check it yourself

  • n=3n = 3 by hand. Number the small cubes (i,j,k)(i,\allowbreak j,\allowbreak k) with i,j,ki,\allowbreak j,\allowbreak k in {0,1,2}\{0,\allowbreak 1,\allowbreak 2\} and let s=i+j+ks = i + j + k. Along cube (i,j,k)(i,\allowbreak j,\allowbreak k), x+y+zx + y + z runs from ss to s+3s + 3, so the plane x+y+z=4.5x + y + z = 4.5 passes through its inside exactly when ss is 2, 3 or 4. The 27 cubes split by ss as 1, 3, 6, 7, 6, 3, 1, so the cut meets 6+7+6=196 + 7 + 6 = 19 cubes, and (81+12−17)/4=19(81 + 12 - 17)/4 = 19.
  • n=6n = 6: the tilt. The plain diagonal cut x+y+z=8.5x + y + z = 8.5 meets 25+27+27=7925 + 27 + 27 = 79 cubes. Bárány and Frankl's tilted cut 5x+5y+6z=45.55x + 5y + 6z = 45.5 keeps all 79 and adds (0,0,5)(0,\allowbreak 0,\allowbreak 5), (4,5,0)(4,\allowbreak 5,\allowbreak 0) and (5,4,0)(5,\allowbreak 4,\allowbreak 0): 82 =F(6)= F(6).
  • n=4n = 4: the lines of cubes. Every small cube lies on one line of cubes in the direction (1,1,1)(1,\allowbreak 1,\allowbreak 1), and a cut meets at most one cube of each line. There are 3n2−3n+1=373n^2 - 3n + 1 = 37 lines at n=4n = 4. The cut 3x+3y+4z=18.53x + 3y + 4z = 18.5 misses exactly 2 of them, the single cubes (0,3,3)(0,\allowbreak 3,\allowbreak 3) and (3,0,3)(3,\allowbreak 0,\allowbreak 3), whose values start at 3⋅0+3⋅3+4⋅3=213 \cdot 0 + 3 \cdot 3 + 4 \cdot 3 = 21, above 18.5. It meets 37−2=35=F(4)37 - 2 = 35 = F(4) cubes.
  • The exact values, m(n)=F(n)m(n) = F(n) for n=3n = 3 to 24: 19, 35, 57, 82, 113, 147, 187, 230, 279, 331, 389, 450, 517, 587, 663, 742, 827, 915, 1009, 1106, 1209, 1315.

Worked examples

Figure 1. n=3n = 3. Left: the cut x+y+z=4.5x + y + z = 4.5 is a regular hexagon, and the grid planes split it into 19 pieces, one for each cube met. Right: the three height layers seen from above, each cube labelled with s=i+j+ks = i + j + k; the shaded cubes are the ones met, 6, 7 and 6 per layer.Full size
Figure 2. n=6n = 6, one grid per height layer, each cube met labelled with ss. Top: the diagonal cut takes the three whole diagonal layers s=6,7,8s = 6,\allowbreak 7,\allowbreak 8, 79 cubes. Bottom: the tilted cut 5x+5y+6z=45.55x + 5y + 6z = 45.5 keeps those and adds 3 cubes from the layers beside them (dark), 82 in all.Full size

How the proof goes. Since a cut meets at most one cube of each line of cubes, the cubes it meets are 3n2−3n+13n^2 - 3n + 1 minus the lines it misses, and the theorem says every cut misses at least D(n)=3n2−3n+1−F(n)D(n) = 3n^2 - 3n + 1 - F(n) lines. For a cut near the diagonal the missed lines are the whole-number points of six right triangles, two for each pair of axes, and a pairing argument bounds each pair of triangles from below. Earlier proved steps (on this page and the upper bound's) confine a better cut to directions near the diagonal. Then the proof splits nn into three ranges:

  • n=3n = 3 to 15: exhaustive computer search over every really different cut.
  • n=16n = 16 to 700: one computer certificate per nn, 685 in all. The directions still in play form a small triangle; the pairing bound changes only across finitely many lines in it, so it is enough to check every corner where two lines cross. At n=16n = 16 there are 174 lines and 4,315 corners, and D(16)=721−587=134D(16) = 721 - 587 = 134. At the corner for the direction (16,16,17)(16,\allowbreak 16,\allowbreak 17) the three pairs give 49+49+36=13449 + 49 + 36 = 134, exactly enough. Every corner reaches 134, and the certificate passes.
  • n≥700n \ge 700: a hand proof, which estimates the same three pair bounds by formulas instead of a computer.

Status and who checked it

  • Proved, computer-assisted, for every n≥3n \ge 3. An agent sent to prove or break the conjecture wrote the proof. A second agent, whose job was to refute it, re-derived every step, re-derived every constant of the hand proof for n≥700n \ge 700 and tested its steps at 18 sizes from 700 to 1,000,001, and re-ran all 685 certificates with a program of its own (685 PASS from each program). Its bottom line: "15 claims: 13 CONFIRMED, 0 REFUTED, 0 GAP, 0 MISLABELLED, 2 UNCHECKED", with the theorem confirmed; the two unchecked claims are a sketch and an open conjecture, labelled as such.
  • Current proof coverage. Lean-verified for n=3n = 3–3232 and n≥700n \ge 700 (see Challenge.lean); checked by certificates for n=33n = 33–699699 by two separately written exact-integer programs whose sources and per-size outputs ship in the verification archive (ZIP).
  • Earlier steps it uses (hand proofs by an agent, re-derived line by line by a second agent): m(n)≥F(n)m(n) \ge F(n) for every n≥3n \ge 3, odd nn included (the paper writes out even nn and leaves odd nn "to the reader"); a cut through every cube of three neighbouring diagonal layers meets at most F(n)F(n); no cut whose direction is within 1/61/6 of (n−1,n−1,n)(n - 1,\allowbreak n - 1,\allowbreak n) in each entry beats F(n)F(n); and the upper bound m(n)≤⌊(9n2+6n−20)/4⌋m(n) \le \lfloor (9n^2 + 6n - 20)/4 \rfloor, on its own page.
  • Checks outside the proof: exhaustive search gives m(n)=F(n)m(n) = F(n) for n=16n = 16 to 24 by two or three programs, and m(25)=1427m(25) = 1427 and m(26)=1542m(26) = 1542, both F(n)F(n); certificates for 16 sizes past 700, up to n=2001n = 2001, all pass; searches for a better cut (every direction in a cone around (1,1,1)(1,\allowbreak 1,\allowbreak 1) for n=16n = 16 to 27, the tilts Bárány and Frankl describe up to n=61n = 61, and a random walk among directions up to n=64n = 64) found none.
  • The "+2" remark is false for every nn, by the theorem. It was read in the manuscript deposited at UCL, not in the printed Monthly pages, which no agent could open.
  • Open: which cuts reach F(n)F(n) (whether every best cut passes through three full diagonal layers); the four-dimensional version; a proof covering the intervening finite range entirely in Lean.

Prior work

Searches run. The agents searched OEIS by terms (no entry holds 19, 35, 57, 82, … or the other windows tried) and by words, arXiv, zbMATH, the works citing the Monthly paper and its companion paper, Zenodo, DataCite, HAL, OSF, the formal-conjectures collection and the Open Problem Garden. On 2026-11-19 a keyed CORE search made 7 requests and got 17 records: 3 relatives, none with the same claim. On 2026-11-20 the agents that proved and checked the bounds and the theorem searched again with their own logs, including GitHub and the web. No source found gives an exact m(n)m(n) beyond n=4n = 4 or tests the "+2" remark: the theorem was not found in the searches logged. On 2026-11-22, a Google Scholar sweep (6 queries) found no prior work stating this claim.

Known parts and relatives, credited to their authors.

  • Bárány and Frankl (2021): the question, both bounds, m(2),m(3),m(4)=7,19,35m(2),\allowbreak m(3),\allowbreak m(4) = 7,\allowbreak 19,\allowbreak 35, the construction that gives F(n)F(n), and the "+2" remark.
  • Math Stack Exchange question 3762021 (July 2020): one answer computes the best cut among one family of central cuts for n=3n = 3 to 9 (19, 35, 57, 81, 113, 145, 187), proves only n=3n = 3, and has "no idea how to solve it in general". Its values equal F(n)F(n) at n=3,4,5,7,9n = 3,\allowbreak 4,\allowbreak 5,\allowbreak 7,\allowbreak 9 and fall short at 6 and 8.
  • The 1995 AHSME, problem 30: the case n=3n = 3 (19 of 27).
  • Bárány and Frankl's companion paper "Cells in the box and a hyperplane" (every dimension, asymptotics only); the Nayar–Tkocz survey; Kalarickal, Rotunno, Singh and Tkocz (2024), which still quotes the two printed bounds; and digital geometry, where the cubes a cut meets form a standard arithmetic plane (Reveillès, Andres).

Still owed: the printed Monthly pages and MathSciNet; Croft, Falconer and Guy, Unsolved Problems in Geometry, and Brass, Moser and Pach, Research Problems in Discrete Geometry.

Further reading