Burnt Pancake Graphs: Structure & Sorting

• Burnt Pancake Graphs are Cayley graphs constructed on signed permutations that use signed prefix reversals to model sorting problems and interconnection networks.
• They reveal precise cycle structures, recursive cluster decompositions, and Hamiltonian properties with exact invariants for order, degree, girth, and connectivity.
• They play a key role in algorithmic complexity and network design by linking group theory with practical models for sorting, fault tolerance, and diagnosability.

Burnt pancake graphs, usually denoted $BP_n$ or $\mathbb{BP}_n$, are Cayley graphs on the group of signed permutations, also called the hyperoctahedral group $B_n$, in which adjacency is defined by signed prefix reversals. They encode the burnt pancake sorting problem: each vertex represents a stack of $n$ distinct pancakes with orientations, and each edge corresponds to flipping the top $k$ pancakes, reversing their order and inverting their signs. In the generalized prefix-reversal framework, the burnt pancake graph is the case $m=2$ of the family $P_m(n)$ or $\mathbb{P}_m(n)$, and it is studied simultaneously as a combinatorial object, a metric space for prefix-reversal sorting, and an interconnection-network topology (Blanco et al., 2022, Blanco et al., 2024).

1. Algebraic model and recursive structure

Formally, the burnt pancake graph $BP_n$ is the Cayley graph over the set of signed permutations $\mathfrak{S}_n^\pm$. If

$\mathbb{BP}_n$0

with $\mathbb{BP}_n$1 and $\mathbb{BP}_n$2 distinct, then a burnt prefix reversal of length $\mathbb{BP}_n$3 maps $\mathbb{BP}_n$4 to

$\mathbb{BP}_n$5

Equivalently, $\mathbb{BP}_n$6, where $\mathbb{BP}_n$7 is the set of signed prefix reversals (Bulteau et al., 2011, Blanco et al., 2024).

Several structural invariants are explicit. The graph has $\mathbb{BP}_n$8 vertices, $\mathbb{BP}_n$9 edges, and degree $B_n$0; it is vertex-transitive and a Cayley graph (Blanco et al., 2023, Wang et al., 2023, Blanco et al., 2024). The generators are involutions (Blanco et al., 2023). A recurring decomposition fixes the last signed symbol: $B_n$1 splits into $B_n$2 vertex-disjoint clusters $B_n$3, each isomorphic to $B_n$4, and each vertex has a unique out-neighbor in another cluster via a cross edge (Wang et al., 2023, Wang et al., 2022).

Invariant	Value for $B_n$5	Source
Order	$B_n$6	(Blanco et al., 2023)
Size	$B_n$7	(Blanco et al., 2023)
Degree	$B_n$8	(Wang et al., 2023)
Connectivity	$B_n$9	(Wang et al., 2023)
Girth	$n$0	(Blanco et al., 2022)

This recursive cluster structure is not merely descriptive. It is the mechanism used in proofs of cycle embedding, generalized connectivity, Hamiltonian fault tolerance, genus bounds, and disjoint path covers (Blanco et al., 2022, Wang et al., 2023, Dvořák et al., 2023, Zhu et al., 2024).

2. Cycle structure, girth, and Hamiltonicity

The shortest cycles in $n$1 have length $n$2. In the generalized pancake notation, the girth of $n$3 is $n$4, whereas the pancake graph $n$5 has girth $n$6, and for $n$7 the girth is $n$8 (Blanco et al., 2022). The same girth value is used in later genus estimates and local cycle classifications (Blanco et al., 2023).

The global cycle spectrum is unusually rich. $n$9 contains every cycle length

$k$0

so it is Hamiltonian and $k$1-pancyclic in the sense that it contains cycles of every length from $k$2 to its order (Blanco et al., 2022, Blanco et al., 2018). The constructive proof uses the recursive structure of $k$3: explicit base cycles traverse multiple copies of $k$4, and selected edges are replaced by longer cycles supplied inductively (Blanco et al., 2018).

Local cycle structure has also been classified. All $k$5-cycles in $k$6 have canonical forms as products of the burnt generators $k$7, and one complete characterization lists four families: $k$8

$k$9

$m=2$0

$m=2$1

with the corresponding index restrictions stated in the classification theorem (Blanco et al., 2019). The same paper gives the first full characterization of all $m=2$2-cycles in $m=2$3, showing that they fall into two canonical forms (Blanco et al., 2019).

Within the generalized family $m=2$4, the burnt pancake case serves as a base case for broader cycle-length theorems. The results for $m=2$5 and $m=2$6 are used inductively to prove that odd $m=2$7 generalized pancake graphs have all cycle lengths from their girth to a Hamiltonian cycle, while even $m=2$8 have all even cycle lengths from girth to a Hamiltonian cycle (Blanco et al., 2022).

3. Distances, sorting, and computational complexity

In graph-theoretic terms, burnt pancake sorting asks for shortest paths from a signed permutation to the identity. The diameter $m=2$9 of $P_m(n)$0 is the worst-case number of burnt prefix reversals needed to sort a stack. Exact values are known for $P_m(n)$1, and the classical bounds

$P_m(n)$2

are summarized in later complexity work (Bulteau et al., 2011).

For the specific stack $P_m(n)$3, denoted in the literature by $P_m(n)$4 or $P_m(n)$5, the distance is often written $P_m(n)$6. Recent work proves that for $P_m(n)$7 and $P_m(n)$8,

$P_m(n)$9

so for all odd $\mathbb{P}_m(n)$0,

$\mathbb{P}_m(n)$1

For even $\mathbb{P}_m(n)$2, only two possibilities remain: $\mathbb{P}_m(n)$3 These results refine older work in which $\mathbb{P}_m(n)$4 had already been shown exact for $\mathbb{P}_m(n)$5 (Jäger et al., 14 Jan 2026). Earlier papers had also established exact values of $\mathbb{P}_m(n)$6 for large congruence classes via fortuitous sequences and generalized fortuitous sequences (Pierre, 2016), and had shown that $\mathbb{P}_m(n)$7 is not always the worst case by giving counterexamples at $\mathbb{P}_m(n)$8 (0901.3119).

Average-case behavior has likewise been studied. There is an algorithm whose average number of flips on $\mathbb{P}_m(n)$9 burnt pancakes is $BP_n$0, while every algorithm must use, on average, at least

$BP_n$1

flips (0901.3119). For restricted input classes, the metric problem is tractable: if $BP_n$2 is a simple signed permutation, then its burnt pancake distance $BP_n$3 is given exactly by

$BP_n$4

and an optimal sorting algorithm runs in $BP_n$5 time, with the distance computable in $BP_n$6 time (Labarre et al., 2010).

The central complexity question remains unresolved for the burnt case. The unburnt pancake problem is $BP_n$7-hard, but for the burnt variant the computational complexity status remains open (Bulteau et al., 2011). Known algorithmic results include $BP_n$8-approximation algorithms for both burnt and unburnt variants, and polynomial-time optimal sorting for simple permutations via the breakpoint graph technique (Bulteau et al., 2011, Labarre et al., 2010).

4. Connectivity, diagnosability, and fault tolerance

Burnt pancake graphs have been studied extensively as interconnection networks, and several higher-order reliability parameters are known exactly. Beyond classical connectivity $BP_n$9, the generalized $\mathfrak{S}_n^\pm$0-connectivity and generalized $\mathfrak{S}_n^\pm$1-connectivity are both

$\mathfrak{S}_n^\pm$2

meaning that for any three, respectively four, vertices there exist $\mathfrak{S}_n^\pm$3 internally edge-disjoint Steiner trees connecting them (Wang et al., 2022, Wang et al., 2023). The proofs use the decomposition into clusters $\mathfrak{S}_n^\pm$4, together with fan constructions, cross-edge structure, and induction on $\mathfrak{S}_n^\pm$5 (Wang et al., 2022, Wang et al., 2023).

Extra connectivity and diagnosability have also been determined. For $\mathfrak{S}_n^\pm$6,

$\mathfrak{S}_n^\pm$7

and for $\mathfrak{S}_n^\pm$8,

$\mathfrak{S}_n^\pm$9

where $\mathbb{BP}_n$00 is the pessimistic diagnosability under the PMC model (Gu et al., 2017). The supporting arguments use explicit lower bounds on neighborhoods of small vertex sets and structural descriptions of small vertex cuts (Gu et al., 2017).

Fault-tolerant Hamiltonian structure is particularly strong. If $\mathbb{BP}_n$01 is a hybrid faulty set consisting of pairs of end-vertices of matching edges together with faulty edges, then $\mathbb{BP}_n$02 contains a Hamiltonian cycle whenever $\mathbb{BP}_n$03, and a Hamiltonian path whenever $\mathbb{BP}_n$04. Equivalently, $\mathbb{BP}_n$05 is $\mathbb{BP}_n$06-hybrid fault Hamiltonian and $\mathbb{BP}_n$07-hybrid fault Hamiltonian connected for $\mathbb{BP}_n$08, and the bounds are tight (Zhu et al., 2024).

A related path-cover result states that for $\mathbb{BP}_n$09, if $\mathbb{BP}_n$10 has at most $\mathbb{BP}_n$11 faulty elements, then for any two disjoint terminal pairs $\mathbb{BP}_n$12 and $\mathbb{BP}_n$13, there exist vertex-disjoint fault-free $\mathbb{BP}_n$14-$\mathbb{BP}_n$15 and $\mathbb{BP}_n$16-$\mathbb{BP}_n$17 paths whose vertices partition the surviving graph. Conversely, for every $\mathbb{BP}_n$18, there exists a set of $\mathbb{BP}_n$19 faulty edges or vertices for which such a paired $\mathbb{BP}_n$20-disjoint path cover does not exist (Dvořák et al., 2023).

These exact formulas place the graph among Cayley-network models with unusually well-resolved reliability parameters. A plausible implication is that the recursive cluster structure, together with vertex-transitivity and high regularity, is the common combinatorial source of these fault-tolerance theorems.

5. Spectral and topological invariants

Recent work has begun to determine the adjacency spectra of burnt pancake graphs explicitly. One theorem states that the spectrum of $\mathbb{BP}_n$21 contains all integer values in

$\mathbb{BP}_n$22

obtained via an equitable partition and an explicit quotient matrix

$\mathbb{BP}_n$23

with $\mathbb{BP}_n$24, $\mathbb{BP}_n$25, and $\mathbb{BP}_n$26 specified entrywise (Blanco et al., 2024). Broader work on prefix-reversal graphs states that, under the specialization $\mathbb{BP}_n$27, the directed spectrum contains

$\mathbb{BP}_n$28

and the undirected formulation contains all even integers in

$\mathbb{BP}_n$29

(Blanco et al., 10 Jun 2025).

The spectral gap has been a focal point. Earlier results gave $\mathbb{BP}_n$30 and computational evidence that the gap approaches $\mathbb{BP}_n$31 as $\mathbb{BP}_n$32 (Blanco et al., 10 Jun 2025). A subsequent note proves that for all $\mathbb{BP}_n$33 the spectral gap of the burnt pancake graph is strictly less than $\mathbb{BP}_n$34, settling a conjecture of Blanco and Buehrle (Greaves et al., 11 Sep 2025).

Topological graph theory has produced the first orientable genus bounds for $\mathbb{BP}_n$35. For $\mathbb{BP}_n$36,

$\mathbb{BP}_n$37

where $\mathbb{BP}_n$38 is the minimum genus of a $\mathbb{BP}_n$39-cell embedding (Blanco et al., 2023). The proof is constructive: a rotation system is defined after a recursive labeling algorithm, ALGRA, labels vertices as $\mathbb{BP}_n$40 or $\mathbb{BP}_n$41 so that every cycle of the form $\mathbb{BP}_n$42 becomes alternating. There are exactly $\mathbb{BP}_n$43 such base cycles, each of length $\mathbb{BP}_n$44, and they serve as facial boundaries in the embedding (Blanco et al., 2023). The resulting bounds are asymptotically tight, and for generalized pancake graphs the genus is $\mathbb{BP}_n$45; specializing to $\mathbb{BP}_n$46 gives the corresponding order for burnt pancake graphs (Blanco et al., 2023).

6. Position within generalized pancake graphs and current research directions

Within the family of generalized pancake graphs $\mathbb{BP}_n$47, the burnt pancake graph is exactly the case $\mathbb{BP}_n$48: each pancake has two sides, so the model coincides with the signed setting (Blanco et al., 2022). The recursive property that induced subgraphs with a fixed last signed symbol are isomorphic to $\mathbb{BP}_n$49 is shared across the family, and the $\mathbb{BP}_n$50 case functions as a base case in the induction schemes used for cycle-length theorems and related constructions (Blanco et al., 2022).

Several contrasts with the ordinary pancake graph are now precise. The unsigned graph has girth $\mathbb{BP}_n$51, whereas the burnt pancake graph has girth $\mathbb{BP}_n$52; the unburnt sorting problem is $\mathbb{BP}_n$53-hard, whereas the burnt case remains open; and enumeration at fixed distance appears algebraically cleaner in the signed setting (Blanco et al., 2022, Bulteau et al., 2011, Blanco et al., 2019). For example, the number of signed permutations requiring exactly four burnt flips is

$\mathbb{BP}_n$54

and numerical evidence suggests polynomial formulas for $\mathbb{BP}_n$55 for $\mathbb{BP}_n$56 (Blanco et al., 2019).

The most persistent open problems in the supplied literature concern optimal sorting and exact distance. The computational complexity of the general burnt pancake problem remains open (Bulteau et al., 2011). For the specific stack $\mathbb{BP}_n$57, the odd case is settled by

$\mathbb{BP}_n$58

while for even $\mathbb{BP}_n$59 only the two values

$\mathbb{BP}_n$60

remain possible (Jäger et al., 14 Jan 2026). This suggests that the coarse geometry of $\mathbb{BP}_n$61 is now much better understood than the full algorithmic complexity of shortest-path computation on it.

Taken together, the current literature presents burnt pancake graphs as a technically mature class of Cayley graphs: their order, degree, girth, cycle spectrum, Hamiltonian structure, generalized connectivities, diagnosability, genus bounds, and several spectral features are all explicit, while the central shortest-path complexity problem remains unresolved (Blanco et al., 2022, Bulteau et al., 2011, Blanco et al., 2023, Greaves et al., 11 Sep 2025).