Goursat Bundles: Structure and Invariants

• Goursat bundles are rank‑2 distributions defined by a strict growth pattern where the derived flag increases sequentially from 2 up to the manifold dimension.
• They are modeled via the universal monster tower and analyzed using RVT code words, Puiseux characteristics, and small-growth invariants, allowing explicit algorithmic computation.
• This framework bridges singularity theory and control theory, offering a robust method for classifying nonholonomic structures in both complex and real geometries.

Searching arXiv for papers on Goursat bundles and closely related Goursat distributions. Goursat bundles are rank‑2 Goursat distributions, equivalently rank‑2 Pfaffian systems satisfying the Goursat condition: on a manifold $M$ of dimension $m$, a distribution $D\subset TM$ is Goursat when its Lie-square sequence consists of genuine subbundles, the rank increases by exactly one at each step, and the process terminates at the full tangent bundle. In the rank‑2 case this means a bracket-generating distribution whose derived flag has ranks $2,3,4,\dots,m$ (Colley et al., 3 Dec 2025). The modern theory places these objects in a universal geometric model, the monster tower over a surface, and studies them through two interacting families of local invariants: structural invariants of curve-singularity type and small-growth invariants of nonholonomic type (Colley et al., 2023, Colley et al., 3 Dec 2025).

1. Definition and basic geometric structure

Let $M$ be a manifold of dimension $m\ge 2$, and let $D\subset TM$ be a distribution with sheaf of sections $\mathcal D$. The Lie square is

$\mathcal D_2=[\mathcal D,\mathcal D],$

generated by local sections of $\mathcal D$ together with their Lie brackets. Inductively,

$m$0

A distribution is Goursat if each $m$1 is locally free, the ranks satisfy

$m$2

and

$m$3

where $m$4 (Colley et al., 3 Dec 2025, Colley et al., 2023).

For rank $m$5, the essential case emphasized in the recent literature, the growth vector is

$m$6

If $m$7, then $m$8 is the corank, and the derived flag takes the form

$m$9

(Colley et al., 3 Dec 2025). This is the defining growth pattern of a Goursat bundle.

A closely related formulation appears in the literature on unbendable rational curves: a rank‑2 distribution $D\subset TM$0 on an $D\subset TM$1-dimensional complex manifold is Goursat if

$D\subset TM$2

so the growth vector is $D\subset TM$3 (Hwang et al., 2021). The equivalence of these formulations is a matter of notation: both encode bracket generation with minimal stepwise growth.

Historically, the standard local model is the Cartan contact distribution on jet spaces. In coordinates $D\subset TM$4 on $D\subset TM$5, the classical Pfaffian system is

$D\subset TM$6

(Castro et al., 2013). This model expresses the characteristic “slow growth” that distinguishes Goursat bundles among bracket-generating systems.

2. Monster tower and universal local model

The fundamental geometric model is the monster tower, also called the Semple tower in the algebraic setting. Starting from a surface $D\subset TM$7, one forms a tower

$D\subset TM$8

where $D\subset TM$9 has dimension $2,3,4,\dots,m$0, and each level carries a canonical rank‑2 distribution $2,3,4,\dots,m$1, the focal distribution (Colley et al., 3 Dec 2025). The construction is by iterated Cartan prolongation of directions.

This tower is universal for rank‑2 Goursat germs: given any rank‑2 Goursat germ of corank $2,3,4,\dots,m$2, there exists $2,3,4,\dots,m$3 such that the germ is locally equivalent to the germ of $2,3,4,\dots,m$4 at $2,3,4,\dots,m$5 (Colley et al., 3 Dec 2025). The same universality is described in the structural-invariants paper by repeated deprolongation to the contact distribution on a 3‑fold and re-embedding into the monster tower over a surface (Colley et al., 2023).

On a standard chart over a coordinate patch $2,3,4,\dots,m$6 with coordinates $2,3,4,\dots,m$7, one has coordinates

$2,3,4,\dots,m$8

and the focal distribution is defined by the annihilation of a Pfaffian system

$2,3,4,\dots,m$9

in suitable local names (Colley et al., 3 Dec 2025). The paper further introduces vertical fields $M$0 and recursively defined focal fields $M$1, with either the ordinary choice

$M$2

or the inverted choice

$M$3

Then

$M$4

(Colley et al., 3 Dec 2025).

The derived Lie-square sequence of $M$5 admits an explicit basis description: $M$6 (Colley et al., 3 Dec 2025). This explicit chart-level structure is central to the later computation of invariants.

The same tower perspective underlies the classification of Goursat multi-flags. For $M$7, one obtains a tower of fibrations with $M$8-fibers, and the problem of classifying Goursat 2-flags up to local equivalence becomes the classification of points in the tower up to symmetry (Castro et al., 2011, Castro et al., 2013). In the spatial case $M$9, the first four levels contain $m\ge 2$0, $m\ge 2$1, $m\ge 2$2, and $m\ge 2$3 orbits, respectively (Castro et al., 2013, Castro et al., 2011).

3. Codes, singularities, and structural invariants

A major development is the translation between Goursat bundles and singularities of curves on surfaces. Each point of the monster tower may be viewed as a multidirection, and a curve germ on the base surface lifts to a focal curve through the tower. Singular positions relative to divisors at infinity produce discrete symbolic invariants (Colley et al., 3 Dec 2025, Colley et al., 2023).

The principal coding device is the RVT code word in the alphabet $m\ge 2$4, recording whether the lifted direction is regular, vertical, or tangential relative to divisors at infinity and their prolongations (Colley et al., 3 Dec 2025, Colley et al., 2023). A parallel “Goursat code word” is defined from the sandwich structure of the derived flag and the Cauchy characteristics (Colley et al., 2023). These code words satisfy an admissibility rule: $m\ge 2$5 may appear only after $m\ge 2$6 or $m\ge 2$7 (Colley et al., 2023).

The first structural-invariants paper associates to a point $m\ge 2$8 several invariants (Colley et al., 2023):

• a Goursat code word or RVT code word $m\ge 2$9,
• a Puiseux characteristic $D\subset TM$0,
• a multiplicity sequence $D\subset TM$1,
• a multiplicity vector $D\subset TM$2,
• a vertical orders vector $D\subset TM$3,
• a restricted Puiseux characteristic.

The Puiseux characteristic is defined exactly as in plane curve singularity theory, after choosing local coordinates $D\subset TM$4 adapted to the focal plane and writing $D\subset TM$5 with $D\subset TM$6, so that $D\subset TM$7 is a convergent fractional power series and the essential exponents produce

$D\subset TM$8

(Colley et al., 2023). The multiplicity sequence records the orders of the successive lifts of the focal curve (Colley et al., 2023).

The vertical orders $D\subset TM$9 are local intersection multiplicities with the divisors at infinity $\mathcal D$0. They satisfy a difference formula with the multiplicities: $\mathcal D$1 (Colley et al., 2023). This already indicates that the singularity-theoretic invariants are tightly organized by the tower geometry.

The same paper establishes recursive algorithms relating RVT words and Puiseux characteristics (Colley et al., 2023). One central device is the lifted word $\mathcal D$2, obtained by dropping the first symbol and replacing a leading $\mathcal D$3, together with any immediately following $\mathcal D$4's, by $\mathcal D$5's (Colley et al., 2023). Front-end and back-end recursions then compute Puiseux data from the code, and conversely reconstruct the code from the Puiseux characteristic (Colley et al., 2023). This makes the structural invariants effectively computable from the combinatorics of the tower.

A common misconception is that the monster-tower code is merely a bookkeeping device for singular directions. The 2023 and 2025 papers show that it determines substantive local invariants, including Puiseux-type data and small-growth data, not just orbit strata (Colley et al., 2023, Colley et al., 3 Dec 2025).

4. Small growth sequence and nonholonomic invariants

The second paper in the pair turns to invariants coming from the small growth sequence, defined using repeated brackets with the original distribution rather than the full derived bundles. For a rank‑2 Goursat distribution $\mathcal D$6 with sheaf $\mathcal D$7,

$\mathcal D$8

(Colley et al., 3 Dec 2025).

At a point $\mathcal D$9, the ranks

$\mathcal D_2=[\mathcal D,\mathcal D],$0

form the small growth vector $\mathcal D_2=[\mathcal D,\mathcal D],$1 (Colley et al., 3 Dec 2025). For any rank‑2 Goursat distribution,

$\mathcal D_2=[\mathcal D,\mathcal D],$2

so the small growth vector always begins

$\mathcal D_2=[\mathcal D,\mathcal D],$3

(Colley et al., 3 Dec 2025). For $\mathcal D_2=[\mathcal D,\mathcal D],$4, the inclusions $\mathcal D_2=[\mathcal D,\mathcal D],$5 may be strict (Colley et al., 3 Dec 2025).

Jean’s $\mathcal D_2=[\mathcal D,\mathcal D],$6-vector is then defined by

$\mathcal D_2=[\mathcal D,\mathcal D],$7

(Colley et al., 3 Dec 2025). Its last entry $\mathcal D_2=[\mathcal D,\mathcal D],$8 is the degree of nonholonomy (Colley et al., 3 Dec 2025). From $\mathcal D_2=[\mathcal D,\mathcal D],$9, the paper defines the derived vector and second derived vector: $\mathcal D$0 (Colley et al., 3 Dec 2025).

The paper groups together as “small growth invariants” (Colley et al., 3 Dec 2025):

• the small growth vector $\mathcal D$1,
• Jean’s $\mathcal D$2-vector,
• the derived vector $\mathcal D$3,
• the second derived vector $\mathcal D$4.

These invariants are local invariants of the germ and depend only on the corresponding RVT or Goursat word (Colley et al., 3 Dec 2025).

A technically central construction is a table of integers $\mathcal D$5, determined by the level $\mathcal D$6, the vertical orders vector $\mathcal D$7, the height $\mathcal D$8, and the index $\mathcal D$9, via

$m$00

The associated $m$01-vector is

$m$02

(Colley et al., 3 Dec 2025).

The structural theorem for the small-growth sheaves states that on a standard chart, every local section of $m$03 can be written as

$m$04

where $m$05 is a section of $m$06, and the coefficient $m$07 has focal order at least $m$08 (Colley et al., 3 Dec 2025). The theorem is sharp: explicit sections attain these lower bounds (Colley et al., 3 Dec 2025).

From this, the paper obtains

$m$09

and

$m$10

(Colley et al., 3 Dec 2025). The most significant comparison results are then:

$m$11

$m$12

$m$13

(Colley et al., 3 Dec 2025).

This identifies structural invariants and small-growth invariants as different presentations of the same local data. A plausible implication is that the dichotomy sometimes drawn between “curve-singularity invariants” and “control-theoretic invariants” is too sharp; for Goursat bundles, the 2025 paper shows they are algorithmically equivalent (Colley et al., 3 Dec 2025).

5. Classification by words and recursive algorithms

The local type of a rank‑2 Goursat germ is encoded by an RVT code word or equivalently a Goursat code word, in the tradition of Montgomery–Zhitomirskii (Colley et al., 3 Dec 2025). The 2025 paper develops two recursive computational frameworks.

The front-end recursion uses the lifted Goursat word $m$14, obtained by dropping the first symbol and normalizing an initial $m$15 and any following $m$16's to $m$17's until the next $m$18, $m$19, or the end (Colley et al., 3 Dec 2025). Via proximity diagrams, one computes the multiplicity sequence, then the multiplicity vector, and from the relations above one gets $m$20, $m$21, and $m$22 (Colley et al., 3 Dec 2025).

The back-end recursion extends Jean’s original recursion from the car-with-$m$23-trailers setting to all Goursat distributions in the smooth, complex, and algebraic settings (Colley et al., 3 Dec 2025). If $m$24 is a Goursat word and $m$25, then for the $m$26-vector:

1. For any nonempty word, $m$27 and $m$28.
2. If $m$29, then

$m$30

1. If $m$31, then

$m$32

2. If $m$33, then

$m$34

Equivalent recursions are written for $m$35 and $m$36 (Colley et al., 3 Dec 2025). The paper shows the front-end and back-end procedures are equivalent (Colley et al., 3 Dec 2025).

An explicit example is the Goursat word $m$37 in a chart of $m$38, for which the paper computes

$m$39

$m$40

$m$41

and the restricted structural invariants

$m$42

$m$43

$m$44

(Colley et al., 3 Dec 2025).

6. Related geometric contexts and broader interpretations

Beyond the core rank‑2 real/smooth theory, Goursat bundles appear in several adjacent domains.

In complex geometry, a bracket-generating family of rational curves is of Goursat type when the associated rank‑2 distribution on the deformation space has growth vector $m$45 (Hwang, 2024). In that setting the normal bundle of a member is

$m$46

(Hwang, 2024). The paper proves that for a family of rational curves of Goursat type, a general member satisfies the formal principle with convergence (Hwang, 2024). The proof uses natural ODEs of the form

$m$47

and canonical Cartan connections due to Doubrov–Komrakov–Morimoto (Hwang, 2024).

A related low-dimensional classification appears in the study of unbendable rational curves. For anti-canonical degree $m$48 and $m$49, the associated rank‑2 distribution is either of Goursat type, with growth $m$50, or of Cartan type, with growth $m$51 in dimension $m$52 (Hwang et al., 2021). In dimensions $m$53, any bracket-generating family of such curves is of Goursat type (Hwang et al., 2021). Lines on a smooth cubic 4‑fold furnish a concrete example (Hwang et al., 2021, Hwang, 2024).

In the multi-flag literature, Goursat $m$54-flags generalize the rank‑2 case by requiring a rank increase of $m$55 at each derived step (Castro et al., 2013, Castro et al., 2011). The spatial $m$56 case is treated via a Semple/Monster tower with $m$57-fibers, where orbits of the prolonged diffeomorphism group classify local types (Castro et al., 2013, Castro et al., 2011). This is not the same object as a rank‑2 Goursat bundle, but it is part of the same prolongation-based paradigm.

By contrast, “Goursat category” in category theory refers to a regular category with 3‑permutable internal equivalence relations (Gran et al., 2017, Gran et al., 2015, Gran et al., 2019). Although some of those papers discuss internal groupoids and bundle-like constructions, this is terminologically unrelated to Goursat bundles as rank‑2 distributions. Confusing the two usages is a common error.

Another separate usage is the “Goursat problem” for hyperbolic PDEs, meaning characteristic boundary data prescribed on intersecting null or characteristic hypersurfaces. The signature kernel solves a hyperbolic PDE of Goursat type

$m$58

(Salvi et al., 2020), and the Klein–Gordon equation on De Sitter–Kerr admits a Goursat problem at the horizons (Millet, 2020). These are again distinct from Goursat bundles as distributions, though they share the historical name.

7. Significance, limits, and current picture

The contemporary picture is that Goursat bundles are among the most rigid nontrivial rank‑2 nonholonomic structures. Their universality through the monster tower reduces local equivalence to the study of canonical distributions $m$59 at points of $m$60 (Colley et al., 3 Dec 2025, Colley et al., 2023). The RVT/Goursat word captures the combinatorics of the germ, while Puiseux characteristics, multiplicity data, vertical orders, and small-growth invariants provide interchangeable numerical realizations of that combinatorics (Colley et al., 2023, Colley et al., 3 Dec 2025).

Three features are particularly notable.

First, the theory unifies two traditions. Structural invariants arise from singularity theory of curves on surfaces; small-growth invariants arise from nonholonomic geometry and control theory. The 2025 comparison theorems show that, for Goursat bundles, these are not parallel descriptions but equivalent ones (Colley et al., 3 Dec 2025).

Second, the monster tower is not merely a convenient model but a universal one. Rank‑2 Goursat germs, and more generally Goursat germs of arbitrary rank after reduction, are locally equivalent to canonical focal bundles on levels of the tower (Colley et al., 2023, Colley et al., 3 Dec 2025).

Third, explicit computation is now algorithmic. Front-end and back-end recursions determine $m$61-vectors, derived vectors, second derived vectors, multiplicity vectors, and vertical orders from the code word, and conversely recover the code from Puiseux data (Colley et al., 2023, Colley et al., 3 Dec 2025).

This suggests a mature local classification theory for rank‑2 Goursat bundles. What remains more open, as the broader literature indicates, is the extension of this precision to higher multi-flags, higher levels of the monster tower, and geometric settings where Goursat-type distributions arise as deformation systems of rational curves or as singular configurations in control systems (Castro et al., 2013, Castro et al., 2011, Hwang, 2024, Hwang et al., 2021).