Claude Code transcript

I've read instr.md, decomp.tex (esp. §on décalage and §on edgewise subdivision), from_hsd.tex, from_coskel.tex, spheres.tex/higher_segal_spheres.tex. Here's where I've landed plus what I'd want to pin down before digging in.

What I take the target statement to be

Conjecture. For $k \geq 1$: $X$ is $2k$-Segal (upper and lower) $\iff$ $\operatorname{sd} X$ is lower $(2k{-}1)$-Segal. ($k=1$ recovers thm edgewise.)

Two things support your instinct here:

1. Duality forces it. Since $\operatorname{sd} X = \operatorname{sd}(X^\op)$, any condition of the form "$\operatorname{sd}X$ satisfies $P$" is invariant under $X \mapsto X^\op$, so it can only ever characterize a self-dual condition on $X$. By lem opposites, the self-dual conditions are lower/upper $(2k{\pm}1)$-Segal and "upper and lower $2k$-Segal". So a one-sided $d$-Segal criterion via $\operatorname{sd}$ alone is impossible, and the even case is the natural candidate — exactly your reasoning about $\operatorname{sd}$ containing both décalages.

2. A gappedness count that comes out right. With $Z = \operatorname{sd}X$, $Z_n = X_{2n+1}$ and $d_i^Z = d_{n-i}d_{n+i+1}$, a gapped $I \subset [n]$ of cardinality $k+1$ gives the "doubled" set $\tilde I = {n-i,\, n+i+1 : i \in I} \subset [2n+1]$, and $Z\llbracket I\rrbracket$ is the diagonal of $X\llbracket \tilde I \rrbracket$ — i.e. a composite of $(k{+}1)$-cubes indexed by the choices $\varepsilon_i \in {n-i, n+i+1}$. I checked: $\tilde I$ itself is not gapped when $0 \in I$ (the pair $n, n{+}1$ is adjacent), but every section ${\varepsilon_i}$ is gapped in $[2n+1]$ — the cross-distance is $i+j+1 \geq 2$ whenever $i \neq j$. Moreover such a section contains $0$ or $2n{+}1$ only when $n \in I$, and never both. So the generalized pasting law (generalized pasting law, from_hsd) should give the forward direction from upper + lower $2k$-Segality and nothing more — precisely mirroring how lem decomp implies sd Segal uses an active–inert square rather than a bare gapped one. The converse should be the cube version of lem ldec square as retract ($Y = \ldec X$ cube a retract of a $Z$ cube via $s_\bot^n$ / $d_0d_2^\bullet$), then prop path space criterion (PSC ldec) + the $X^\op$ trick.

So my working guess is that it's true and that both directions are adaptations rather than new ideas — with the real work in the forward direction being bookkeeping about which intermediate $S' \subsetneq [2n+1]$ the pasting steps live over (deleting elements can create adjacencies, though the ones I spot-checked are protected by gappedness of $I$).

Questions

1. Setting. Simplicial objects in an $\infty$-category with finite limits (as in from_hsd.tex), or simplicial sets? I'd default to the general setting since the retract/pasting lemmas are already stated there.

2. Conventions. from_coskel.tex ($X(S)$, $e_i$, $\llbracket I \subset S\rrbracket$ covariant on $\mathcal P(I)$) vs from_hsd.tex (intersection cubes on $\mathcal P(I)^\op$). Which do you want the writeup in? I lean from_coskel since it's the most recent.

3. Odd case. Do you want me to also chase the other self-dual family — is there a $\operatorname{sd}$-criterion for upper $(2k{+}1)$-Segal? The obvious guess "$\operatorname{sd}X$ lower $2k$-Segal" is immediately suspect (not self-dual as a condition on $Z$... though it is when combined with upper), so maybe "$\operatorname{sd} X$ is $2k$-Segal". Worth a look or a distraction?

4. Iterated subdivision. Any interest in $\operatorname{sd}_r$ ($[n] \mapsto$ $r$-fold analogue), which might shift $d$ by more than one step? Or strictly the classical $\operatorname{sd}$?

5. Deliverable + method. Do you want (a) a short sketch/verdict first, (b) a full LaTeX note with proofs, or (c) me to try to break it computationally first (finite simplicial sets like $\Delta^n/\partial\Delta^n$, $\partial\Delta^{n+1}$, where your coskeletality theorem makes $d$-Segality a finite check)? I'd suggest a quick computational sanity check on $k=2$ against a known non-$4$-Segal example before writing anything up.

6. Literature. Should I look up BOORS ESC and Poguntke directly (web), or do you want me working only from these files?

Here's where things stand.

Answer: yes, twice — and the odd case is the fun one

Both conjectures survive everything I threw at them. Writing $Z = \operatorname{sd}X$, for every $k \ge 1$:

• (A) $X$ is $2k$-Segal (upper and lower) $\iff$ $Z$ is lower $(2k{-}1)$-Segal. (k=1 is the classical criterion.)
• (B) $X$ is upper $(2k{+}1)$-Segal $\iff$ $Z$ is upper $2k$-Segal.

(B) is asymmetric, which surprised me until I noticed that $\operatorname{sd}X$ is not self-opposite — $\operatorname{sd}(X^\op) = \operatorname{sd}X$, but $(\operatorname{sd}X)^\op$ genuinely differs (verified numerically). So the condition on $X$ must be self-dual, but the condition on $\operatorname{sd}X$ need not be. That's what leaves room for an odd criterion at all.

The forward direction of both, from one lemma

$D(S) = S^\op \star S$ satisfies $D(S_i) = D(S)\setminus{i^-,i^+}$, so $Z\llbracket I \subset S\rrbracket$ is the diagonal of the $(2k{+}2)$-cube $X\llbracket I^\pm \subset D(S)\rrbracket$ — each $Z$-direction is the composite of two $X$-directions. Iterating your generalized pasting law one direction at a time reduces cartesianness to the $2^{k+1}$ leaves $X\llbracket E_\varepsilon \subset A_\varepsilon\rrbracket$, one per section $\varepsilon_i \in {i^-,i^+}$, with $A_\varepsilon = D(S)\setminus{i^+ : \varepsilon_i = i^-}$.

Then two combinatorial facts do all the work:

1. Every $E_\varepsilon$ is gapped in $A_\varepsilon$ (cross-pairs $i^-, j^+$, $i\neq j$, are separated by $\min(S)^-$; the degenerate case $i=\min(S)$ is separated by $s^+$ for the minimal $s\in S\setminus I$ between them).
2. No leaf contains both endpoints of its ambient: $\min(A_\varepsilon)=\max(S)^-$, and if $E_\varepsilon$ contains it then $\max(S)^+$ was deleted and the new max is $t^+$ with $t\notin I$ by gappedness.

So $2k$-Segality of $X$ suffices — and that's exactly why the level is $2k$: the two blocks of the subdivision demand the two one-sided conditions, never both at once. If additionally $\max(S)\notin I$ (i.e. you're checking upper $2k$-Segality of $Z$), every leaf avoids both endpoints, so only upper $(2k{+}1)$-Segality is needed — that's (B). And imposing $\min(S)\notin I$ instead changes nothing, which explains an odd pattern in the data: on $\operatorname{sd}X$, lower $(2k{-}1)$ $\iff$ lower $2k$, and upper $2k$ $\iff$ upper $(2k{+}1)$. Only three profiles ever occur.

Remaining gap: the converses. Both should come from a cube-level version of lem ldec square as retract plus easy psc (the "$\max\notin I$" restriction matches on both sides, which is what (B) needs). The comparison map comes from the natural collapse $c\colon D(S)\to S^\triangleleft$; the retraction $r(\bot)=\max(S)^-$, $r(s)=s^+$ has $cr=\mathrm{id}$ but isn't natural — same corner-map adjustment as your $d_1^n$ vs $d_0d_2^{n-1}$. I haven't checked those assemble into a map of $(k{+}1)$-cubes.

Evidence

No counterexample in ~350 finite simplicial sets ($\Delta^n/\partial\Delta^n$, $\partial\Delta^N$, all quotients $\Delta^N/A$ by subcomplexes, $N\le 3$) at $k=1,2$, plus spheres at $k=3,4$. The code reproduces your published sphere theorems exactly ($\Delta^4/\partial\Delta^4$ is $8$-Segal, not lower $7$-Segal), which is my main confidence in the implementation.

Sharpest data point, $X=\Delta^3/\partial\Delta^3$ at $k=2$: $X$ is upper $5$-Segal but not $4$-Segal, and $\operatorname{sd}X$ is upper $4$-Segal but not lower $4$-Segal. That single row separates (B) from (A) and shows the asymmetry is real, not an artifact.

Written up in notes/edgewise-higher.md, with notes/segal.py (higher Segal conditions + $\operatorname{sd}$ for finite simplicial sets) and notes/search.py. Left untracked — say the word and I'll commit them.

Want me to push on the retract lemma next, or hunt harder for counterexamples first (e.g. non-coskeletal or asymmetric examples at $k=2$, where my pool is thinnest)?

Both directions are now done. Summary of where the investigation landed:

Result

For every $k\ge 1$, with $Z=\operatorname{sd}X$:

• (A) $X$ is $2k$-Segal (upper and lower) $\iff$ $Z$ is lower $(2k{-}1)$-Segal.
• (B) $X$ is upper $(2k{+}1)$-Segal $\iff$ $Z$ is upper $2k$-Segal.

So your instinct was right on both counts: the even case is where the criterion lives, and the odd case does have an analogue — but only after allowing the condition on $\operatorname{sd}X$ to be one-sided, which is legitimate because $\operatorname{sd}X$ is not self-opposite even though $\operatorname{sd}$ can't distinguish $X$ from $X^\op$.

The retract lemma

The clean version turned out to be stronger and simpler than the $k=1$ case in decomp.tex:

Lemma. For $S\in\Delta$ and any proper $I\subset S$, the cube $(\decbot X)\llbracket I\subset S\rrbracket$ is a retract of $(\operatorname{sd}X)\llbracket I\subset S\rrbracket$.

The comparison map is $X(c)$ for the natural collapse $c\colon D(S)\to S^\triangleleft$. For the retraction, the point is to fix one $t\in S\setminus I$ globally and set $r_U(\bot)=t^-$, $r_U(s)=s^+$ for every corner $U\subseteq I$. Then $c r=\id$, and the corner squares commute on the nose because $t\notin I$ guarantees $t$ survives in $S\setminus V$ for every $V\subseteq I$ — no corner-by-corner renormalization. Such a $t$ exists precisely because gapped subsets are proper.

This explains the awkward asymmetry in lem ldec square as retract: your $d_1^n$ and $d_0d_2^{n-1}$ are the same map $r$ (with $t=\max S$) read at two corners whose $[n]$-relabelings differ. I checked combinatorially — it's a statement about maps in $\Delta$, independent of $X$ — over 22,638 corner pairs with no failures, and confirmed that the naive corner-local choice $t=\min(S\setminus U)$ really does break (140 non-commuting squares), so the fixed-$t$ trick is doing genuine work.

Both converses follow immediately, since retracts of cartesian cubes are cartesian: (A) from easy psc (1) plus $\operatorname{sd}(X^\op)=\operatorname{sd}X$ for the upper half; (B) because the lemma carries the restriction "$\max(S)\notin I$" across verbatim, landing exactly on easy psc (3b).

Caveat on what's proved to what standard

The converses I'd now call proved. The forward directions are a solid sketch, not a finished proof: the leaf analysis (gappedness, and that no leaf contains both endpoints of its ambient) I checked carefully by hand, but the iterated-pasting induction that reduces the diagonal cube to those $2^{k+1}$ leaves is written as a recursion over "mixed double/single direction" cubes and deserves to be written out properly before you'd trust it in a paper. That's the one place I'd want a second pass.

Committed to notes/edgewise-higher-segal (two commits: bf4b43b, 8b54b9b), with notes/edgewise-higher.md, segal.py, search.py, retract.py. Merge with git merge notes/edgewise-higher-segal from main.

Next most useful thing, if you want it: write the pasting induction out in full LaTeX in your from_coskel conventions — at which point (A) and (B) are a short paper-section, since the two lemmas are each about a page.

Done — notes/edgewise.tex, 6 pages, compiles clean with no undefined references. Committed as f7033ff on notes/edgewise-higher-segal.

What the write-up contains

The Splitting Lemma is the piece that was only a sketch before, and writing it out clarified it. The trick is to prove something slightly more general than needed: cubes $W(A;(B_\lambda))$ indexed by a set of directions $\Lambda$, where each direction $\lambda$ deletes a subset $B_\lambda \subseteq A$ with $1 \le |B_\lambda| \le 2$, the $B_\lambda$ pairwise disjoint. Induction on the number of two-element directions, splitting one at a time via the generalized pasting law; the base case $d=0$ is literally a higher-Segal cube $X\llbracket E \subset A\rrbracket$.

Stating it that way makes the bookkeeping evaporate: the "which partner got deleted first" asymmetry is absorbed into the definition $A_\varepsilon = A \setminus {c_\lambda : \varepsilon_\lambda = b_\lambda}$, and the inductive step is just the observation that sections of the two halves are exactly the sections of the whole with $\varepsilon_\mu$ fixed, with the same $E_\varepsilon$ and $A_\varepsilon$. No ordering choices leak into the statement.

Then two short combinatorial lemmas ($E_\varepsilon$ gapped; no leaf contains both endpoints), both resting on one observation I isolated as equation \eqref{eq mij}: for $i<j$ in $I$, the least $s \in S$ with $i<s<j$ satisfies $s \notin I$, so $s^\pm$ is never deleted. That single fact does all three cases of gappedness and the endpoint argument.

Two things I'd flag for your review

• The $|S| \ge 2k+1$ bound is used twice (properness of $E_\varepsilon \subset A_\varepsilon$, and existence of a predecessor of $\max(S)$). It's immediate from $I$ gapped of cardinality $k+1$, but it's the kind of thing worth a second look since the lemmas would be false without it in edge cases.
• The $k=0$ case is excluded throughout ($k \ge 1$), matching your conventions. I didn't check whether either theorem degenerates sensibly there.

I also added a short final section recording the sharpness example ($\Delta^3/\partial\Delta^3$ at $k=2$: $\operatorname{sd}X$ is upper $4$-Segal but neither lower $3$- nor lower $4$-Segal, so (B)'s "upper" can't be swapped for "lower"), and stated the leftover collapse phenomenon as a Question — with the note that the Wiggle Lemma reduces its first half to the single cube $\lowod^k_{2k}$ on $\operatorname{sd}X$. That's the obvious next thing to try if you want it closed.

notes/.gitignore covers the LaTeX aux files and the PDF. There's an untracked .DS_Store at the repo root I left alone.