FLIPS: Local Changes in Math & Modern Systems

• FLIPS are local, structure-preserving moves that change configurations while maintaining a global admissibility class, with applications from combinatorial tilings to birational geometry.
• They underpin convergence analyses in stochastic cooling models, flip graph studies, and reconfiguration processes in triangulations, domino tilings, and pancake sorting.
• Recent acronymic uses in federated learning, LLM fingerprinting, swarm robotics, and protein design illustrate FLIPS’ versatile impact across cutting-edge technological systems.

FLIPS is a polysemous research term whose meaning depends strongly on domain. In combinatorics, discrete geometry, tiling theory, and birational geometry, a flip is a local transformation that preserves a global admissibility class while changing local structure. In recent systems papers, FLIPS, FLIP, and FliPS also denote specific methods: Federated Learning using Intelligent Participant Selection, instance-level fingerprinting for LLMs via pseudo-random sequences, Real-Time and Resilient Formation Planning for Large-Scale Distributed Swarms via Point Cloud Registration, and a flexibility-conditioned protein backbone generator (Bhope et al., 2023, Richardeau et al., 2 Jun 2026, Zhou et al., 28 May 2026, Viliuga et al., 29 Jun 2025). Across these uses, the common theme is controlled local change under structural constraints, but the technical objects range from words, tilings, and triangulations to generalized pairs, neural generators, and robotic formations.

1. Terminological scope and core formal pattern

The literal notion of a flip is highly local. In balanced two-letter words, it is the swap $12 \leftrightarrow 21$ at adjacent unequal letters (Bodini et al., 2010). In dimer tilings of the triangular grid, it is the local transformation at a vertex belonging to exactly three tiles, equivalently a rotation by $\frac{\pi}{3}$, which in the lift picture adds or removes one cube (Fernique et al., 2011). In combinatorial triangulations, it deletes an edge $(a,b)$ shared by triangles $(a,b,c)$ and $(a,b,d)$ and inserts the other diagonal $(c,d)$, provided $(c,d)$ is not already present (Bose et al., 2011). In plane spanning trees, it removes one edge and adds another so that the result is again a plane spanning tree (Aichholzer et al., 21 Aug 2025). In domino tilings, it removes two adjacent parallel dominoes and places them back in the only other possible way (Klivans et al., 2020). In arbitrary posets, a flip reverses the cover relations crossing a flip pair $(A,B)$ while leaving the underlying undirected Hasse diagram unchanged (Nagano, 10 May 2026).

The same word also names birational surgeries. In the Minimal Model Program, a flip replaces a small contraction $(X,B)\dashrightarrow X^+$ that is $K_X+B$-negative on one side by another model where $\frac{\pi}{3}$0 becomes positive in the flipped direction (Han et al., 2022). In symplectic geometry, the relevant operation is often a reverse simple flip or a blow-up with trivial center, interpreted as a symplectic MMP transition (Charest et al., 2015).

A concise disambiguation is therefore useful.

Sense of “FLIPS”	Mathematical or algorithmic object	Representative source
Local swap or reconfiguration move	words, tilings, triangulations, spanning trees, posets	(Bodini et al., 2010, Fernique et al., 2011, Bose et al., 2011, Aichholzer et al., 21 Aug 2025, Nagano, 10 May 2026)
Birational flip	lc, klt, terminal, or generalized pairs in the MMP	(Han et al., 2022, Chen et al., 2020)
Symplectic reverse flip	surgery creating Floer-non-trivial Lagrangian tori	(Charest et al., 2015)
Acronymic FLIPS/FLIP/FliPS	FL middleware, LLM fingerprinting, swarm formation, protein design	(Bhope et al., 2023, Richardeau et al., 2 Jun 2026, Zhou et al., 28 May 2026, Viliuga et al., 29 Jun 2025)

This suggests a broad abstract schema: a flip modifies a state by a minimal admissible move, and the research questions concern reachability, invariants, complexity, convergence, or controllability.

2. Stochastic cooling and defect-eliminating flip dynamics

A particularly clear probabilistic instantiation appears in balanced two-letter words. The state space is

$\frac{\pi}{3}$1

the energy is

$\frac{\pi}{3}$2

and the absorbing states are the alternating words $\frac{\pi}{3}$3 and $\frac{\pi}{3}$4. The Markov chain chooses uniformly among flips that do not increase $\frac{\pi}{3}$5; irreversible flips decrease $\frac{\pi}{3}$6, while reversible flips leave it unchanged. The expected convergence time to $\frac{\pi}{3}$7 is bounded by $\frac{\pi}{3}$8 in the worst case and by $\frac{\pi}{3}$9 for the uniform initial distribution. The proof uses a drift argument with the Dyck-factor potential

$(a,b)$0

since the raw mismatch count may have zero expected drift on configurations admitting only reversible flips (Bodini et al., 2010).

The dimer-tiling analogue is a two-dimensional stochastic cooling process. The state space is the set $(a,b)$1 of dimer tilings of a fixed finite simply connected domain with $(a,b)$2 tiles. Energy is the total number of errors, meaning edges shared by two identical tiles. A flip changes the volume by $(a,b)$3 and the energy by $(a,b)$4, but the cooling rule allows only flips with $(a,b)$5. A key structural proposition states that whenever a tiling has an error, there exists a flip that does not increase the number of errors and can be chosen so that it decreases the volume. Hence frozen tilings are necessarily error-free. Using the potential $(a,b)$6, triconvexity, and the triconvex hull, the paper proves

$(a,b)$7

while numerical experiments suggest $(a,b)$8 and average-case behavior $(a,b)$9 (Fernique et al., 2011).

These two models are closely aligned in physical interpretation. Both treat mismatches or errors as energy, allow only energetically neutral or decreasing local moves, and prove convergence by finding a potential with negative drift even when the naive defect count is insufficient. This suggests a common “cooling-by-admissible-local-rearrangement” paradigm, although the one-dimensional and tiling settings require different variants and geometric tools.

3. Flip graphs, distances, and reconfiguration complexity

In planar graph reconfiguration, the central object is often a flip graph. For combinatorial triangulations on $(a,b,c)$0 vertices, one major result is that any triangulation can be made 4-connected using at most

$(a,b,c)$1

flips, and that this bound is tight for an infinite family. Since 4-connectivity is equivalent to the absence of separating triangles, the proof repeatedly flips an edge of a deepest separating triangle and pays for each flip with a five-coin charging scheme. The same work proves that, for $(a,b,c)$2, any 4-connected triangulation can be transformed into the canonical triangulation using at most $(a,b,c)$3 flips, matching the known lower bound. Combining the two stages yields a diameter bound of $(a,b,c)$4 for the flip graph, improving the previous $(a,b,c)$5 bound (Bose et al., 2011).

The thesis treatment of flips in triangulations broadens this picture to labelled settings. For edge-labelled triangulations of a convex polygon and for edge-labelled combinatorial triangulations, the flip-graph diameter is $(a,b,c)$6. For simultaneous flips, the paper proves $(a,b,c)$7 upper bounds for both edge-labelled convex polygons and edge-labelled combinatorial triangulations, together with an $(a,b,c)$8 lower bound. The same thesis also states $(a,b,c)$9 exchanging flips for pointed edge-labelled pseudo-triangulations and $(a,b,d)$0 flips when insertion, deletion, and exchanging flips are allowed in the general pseudo-triangulation setting (Verdonschot, 2015).

The dedicated pseudo-triangulation paper provides the local mechanics behind those bounds. In a pointed pseudo-triangulation, removing an internal edge yields a pseudo-quadrilateral with a unique other bitangent, so an exchanging flip is canonical. The proof of the $(a,b,d)$1 bound passes through the left-shelling pseudo-triangulation and relies on two global label-moving primitives, sweep and shuffle. In the general setting, free labels, indexed fans, and $(a,b,d)$2-cost shift operations yield the bound

$(a,b,d)$3

for transforming any edge-labelled pseudo-triangulation into any other (Bose et al., 2015).

Plane spanning trees on points in convex position admit two constrained flip models. A compatible flip requires that the removed and inserted edges do not cross; a rotation requires that they share a common vertex. For compatible flips, the diameter upper bound improves to

$(a,b,d)$4

and every shortest compatible flip sequence satisfies the strong happy edge property, meaning that no shortest sequence flips an edge already shared by the initial and target trees. This structural fact yields an FPT algorithm with runtime

$(a,b,d)$5

for compatible flip distance $(a,b,d)$6. For rotations, the happy edge property fails, but the rotation-graph diameter improves to

$(a,b,d)$7

(Aichholzer et al., 21 Aug 2025).

Hypertriangulations supply a higher-order generalization. A level-$(a,b,d)$8 hypertriangulation of a generic planar point set is induced by the projection of the $(a,b,d)$9-th hypersimplex, with triangle types classified as white or black. The paper introduces four local flip types: a quadrilateral flip, a triangle-to-three-triangles flip, a parallelogram reflection, and a centrally symmetric hexagon reflection. Using the aging map $(c,d)$0 and its inverse on black triangles, it proves that level-2 hypertriangulations are connected by these flips (Edelsbrunner et al., 2022).

Prefix-reversal sorting gives a different notion of flip distance. In pancake sorting, a flip reverses the top $(c,d)$1 pancakes. The average-case results include an algorithm for burnt pancakes using $(c,d)$2 flips on average and a randomized algorithm for unburnt pancakes using at most $(c,d)$3 flips on average. The same paper proves the burnt lower bound

$(c,d)$4

and exact values such as $(c,d)$5, $(c,d)$6, $(c,d)$7, and $(c,d)$8 (0901.3119).

4. Abstract flip theories, invariants, and symmetric connectivity

Some papers abstract the local move itself. In arbitrary posets, a flip is defined from a flip pair $(c,d)$9 satisfying

$(c,d)$0

The flipped poset reverses precisely the cover relations between $(c,d)$1 and $(c,d)$2. When $(c,d)$3 is a lattice, a flip is a mutation if the result is also a lattice. The main criterion states that a flip is a mutation iff it satisfies the atom-coatom condition and the $(c,d)$4-sublattice condition. The paper further defines locally mutable and mutable lattices, proves that mutable lattices are semidistributive, shows that type-A and type-B Cambrian lattices are locally mutable, and introduces Ordovician lattices as lattices obtained from Cambrian lattices by iterated mutations (Nagano, 10 May 2026).

A related but different symmetry-constrained theory arises for separated set-systems. For the involution

$(c,d)$5

the paper studies maximal symmetric strongly separated, weakly separated, and $(c,d)$6-separated collections. Classical flips of Leclerc–Zelevinsky are replaced by symmetric flips, performed together with their mirror images so that symmetry is preserved. For even $(c,d)$7, maximal symmetric strongly separated collections are connected by double hexagonal flips and big flips; maximal symmetric weakly separated collections are likewise connected by symmetric weak flips and big flips. For even $(c,d)$8 and even $(c,d)$9, maximal symmetric $(A,B)$0-separated collections form an acyclic directed graph with unique minimal and maximal elements. The paper identifies these structures with symmetric rhombus tilings, symmetric combies, and symmetric cubillages, and relates them to a type-C analogue of higher Bruhat orders (Danilov et al., 2021).

Domino tilings in dimensions $(A,B)$1 provide an invariant-based flip theory. The paper defines the twist

$(A,B)$2

through a Kasteleyn-signed determinant and proves that twist is invariant under flips, while a trit changes it. A region $(A,B)$3 is regular if equal twist implies flip-connectivity after stabilization by extra vertical layers. All boxes are regular except $(A,B)$4. For regular $(A,B)$5, there exists an even $(A,B)$6 depending only on $(A,B)$7 such that tilings of $(A,B)$8 with equal twist become flip-connected in $(A,B)$9. As a corollary, for large $(X,B)\dashrightarrow X^+$0, the flip graph of $(X,B)\dashrightarrow X^+$1 has two twin giant components, one for each twist value (Klivans et al., 2020).

Taken together, these works shift attention from raw reachability to structure-preserving local surgery under explicit invariants: symmetry, latticehood, semidistributivity, or mod-$(X,B)\dashrightarrow X^+$2 twist. This suggests that many flip theories are governed less by the move itself than by the obstruction class that survives every move.

5. Flips in birational and symplectic geometry

In birational geometry, the central question is often not distance in a flip graph but termination of flip sequences. One line of work introduces exceptionally non-canonical (enc) singularities and shows that the termination of lc flips can be reduced to the ACC conjecture for global mlds of enc pairs together with termination of terminal flips. The main reduction states that, in dimension $(X,B)\dashrightarrow X^+$3, ACC for global mlds for enc pairs with finite coefficients and termination of $(X,B)\dashrightarrow X^+$4-factorial terminal flips imply termination of lc flips in dimension $(X,B)\dashrightarrow X^+$5. A direct consequence is that ACC for global mlds of enc pairs in dimension $(X,B)\dashrightarrow X^+$6 implies termination of lc flips in dimension $(X,B)\dashrightarrow X^+$7. The same paper proves ACC for mlds of enc pairs in dimension $(X,B)\dashrightarrow X^+$8, giving a proof of termination of lc flips in dimension $(X,B)\dashrightarrow X^+$9 that does not rely on difficulty functions (Han et al., 2022).

A complementary result addresses log canonical generalized pairs. For a $K_X+B$0-dimensional NQC lc g-pair $K_X+B$1, if $K_X+B$2 is pseudo-effective over $K_X+B$3, then any sequence of flips over $K_X+B$4 terminates. The proof architecture first establishes termination for $K_X+B$5-dimensional NQC lc g-pairs, then proves that a pseudo-effective NQC lc g-pair admitting an NQC weak Zariski decomposition over $K_X+B$6 has terminating flip sequences provided lower-dimensional termination is known. Since $K_X+B$7-dimensional pseudo-effective NQC lc g-pairs admit NQC weak Zariski decompositions, this yields the four-dimensional theorem (Chen et al., 2020).

Symplectic geometry uses the same word for a different transition. Reverse simple flips or blow-ups with trivial center in compact rational symplectic manifolds are shown to create Lagrangian tori with non-trivial Floer theory. If $K_X+B$8 is obtained from $K_X+B$9 by a reverse simple flip or blow-up with trivial center, with multiplicity

$\frac{\pi}{3}$00

and sufficiently small exceptional locus, then there exists a Lagrangian torus $\frac{\pi}{3}$01 with $\frac{\pi}{3}$02 distinct local systems and Maurer–Cartan solutions such that

$\frac{\pi}{3}$03

The local model is toric near the exceptional locus, and the relevant torus is a regular Lagrangian detected by the disk potential (Charest et al., 2015).

A common misconception is that “flip” has a uniform technical meaning across these geometric literatures. It does not. In the MMP, flips are birational replacements whose admissibility is expressed through discrepancies, lc thresholds, and pseudo-effectivity. In the symplectic MMP, reverse flips are local surgeries detected by Floer-theoretic non-triviality. The shared vocabulary reflects locality and structural preservation, not identical formalism.

6. Modern acronymic systems and contemporary engineering uses

Several papers use FLIPS, FLIP, or FliPS as acronyms. In federated learning, FLIPS stands for Federated Learning using Intelligent Participant Selection. It clusters clients according to label-distribution vectors $\frac{\pi}{3}$04, chooses the number of clusters with a Davies–Bouldin-index heuristic, and performs round-robin participant selection across clusters so that each cluster is equitably represented. It supports FedAvg, FedProx, FedDyn, FedOpt, and FedYogi, incorporates over-provisioning for straggler management, and protects label distributions and cluster assignments via a trusted execution environment. The empirical study reports 17–20 percentage points higher accuracy and 20–60% lower communication costs than random selection and several “smart” baselines, with these benefits persisting in the presence of stragglers (Bhope et al., 2023).

For LLM auditing, FLIPS is a black-box, instance-level fingerprinting method. It queries an LLM with prompts that request pseudo-random binary sequences, converts the outputs to bitstrings, extracts NIST randomness-test features, and classifies instances with one XGBoost model per token pair plus soft voting. The benchmark contains 237 model instances derived from 25 open-weight LLMs, spanning temperatures, system prompts, quantization variants, and abliterated models. With 40 extraction queries and 8 verification queries, FLIPS achieves 96% closed-set accuracy and 90% open-set accuracy, versus 35% for the adapted LLMmap baseline (Richardeau et al., 2 Jun 2026).

In swarm robotics, FLIP denotes Real-Time and Resilient Formation Planning for Large-Scale DIstributed Swarms via Point Cloud Registration. The method recasts Optimal Formation Position Sequence computation as spatiotemporal point cloud registration, estimating similarity transforms $\frac{\pi}{3}$05 from other agents’ predicted positions and using RANSAC for outlier rejection. The resulting OFPS is integrated into trajectory optimization with MINCO and L-BFGS. Reported performance includes large-scale simulations of 120-drone formation, average per-agent local planning time of about 0.09 s in the 120-drone case, and maximum OFPS computation time below 0.06 s even at 1000 agents (Zhou et al., 28 May 2026).

In protein design, FliPS is an SE(3)-equivariant conditional flow matching model for generating backbones with a target per-residue flexibility profile. It is paired with BackFlip, an SE(3)-equivariant predictor of local RMSF from backbone geometry,

$\frac{\pi}{3}$06

trained on ATLAS MD data. FliPS conditions on flexibility embeddings, uses a differentiable auxiliary flexibility loss driven by BackFlip, and is validated by all-atom MD simulations. BackFlip reaches correlation around 0.80 on unseen natural proteins with MAE around 0.17 Å, while generated backbones selected by the pipeline achieve MD-derived flexibility profiles with average Pearson $\frac{\pi}{3}$07 around 0.70–0.79 on top samples (Viliuga et al., 29 Jun 2025).

Robotics also uses the literal maneuveral sense of flips. A quadrotor study models aggressive belly-flop trajectories with MPC, using a prediction horizon $\frac{\pi}{3}$08, time step $\frac{\pi}{3}$09, and flip reference

$\frac{\pi}{3}$10

The mission profile includes rally, flip, probe release, recovery, and return. In simulation, the probe’s final $\frac{\pi}{3}$11-position is approximately 21 m for a desired throw range of 20 m; in real flight, eight flip trials were successfully executed and recovered (Jain et al., 2024).

The acronymic uses are therefore substantively unrelated to the classical local-move literature. Yet they retain a recognizable design logic: a constrained intervention changes local configuration—participant sets, instance signatures, formation assignments, residue-level flexibility, or vehicle attitude—while preserving a higher-level task objective. This suggests that “FLIPS” has become a portable label for methods centered on controlled, structure-aware state transitions, even when no formal flip graph is present.