WorksEngineering
Control Invariance on Manifolds through Tangent Bundle Splitting
Abstract
Control invariance for interconnected non-linear systems on manifolds requires conditions under which subsystem safety constraints can be combined to form an invariant set. This paper develops local product type control invariant sets on manifolds using tangent bundle splitting, adapted product coordinates, and subsystem tangency conditions satisfied simultaneously by one admissible input at each state. The resulting conditions establish control invariance of product sets without requiring independent subsystem dynamics, with separate control choices possible when inputs are independently constrained and each acts only on its subsystem. This provides a way to verify subsystem safety without discarding dynamic coupling and is illustrated by a torus example in which adapted coordinates give a curved constraint region a product representation.
Introduction
The state of a network of interconnected systems can be described in Euclidean space by the partition , with each being the state of the subsystem. In this paper, we assume each subsystem can be controlled independently, using with compact, and the dynamics evolve by
where collects the states of the other subsystems on . Given this partition, a natural question is whether there exist sets such that an admissible control keeps for every subsystem and all , whenever for all . The main tools for answering this question lie in set invariance theory [1] and viability theory [2]. For linear systems, [3] gives conditions for the existence of such sets based on the coupling strength, i.e., , and [4] shows the maximum coupling that allows these sets to exist. These ideas were exploited in distributed control for linear networks in [5] and [6] by allowing a time-varying state partition. As a result, partitioning the state is key to obtaining distributed guarantees from a network.
Set invariance is closely linked to safety, as trajectories remain “trapped” within the set. For example, consider a scalar function, known as a Control Barrier Function (CBF), ; its zero superlevel set serves as the safety certificate. The authors in [7] and [8] employ quadratic programmes and Lie derivatives to ensure this level set encloses the system trajectories. The stochastic setting was handled by [9], and probabilistic results were obtained in [10]. An alternative view of safety certification can be achieved via reachability analysis using the Hamilton-Jacobi-Bellman equation [11]. For networks, existing work, see [10] and [12], examines how subsystem constraints can be combined into an invariant set while accounting for dynamic coupling. Recent work in [13] proposes global certification via only local certificates. This implies that the certificates have a product-like structure.
Geometric control describes dynamics intrinsically on manifolds [14], [15], allowing system structure to be expressed independently of coordinates.Safety filters based on CBFs on manifolds have been explored using backstepping techniques [16]. Ibuki et al. Ibuki et al. [17] formulated CBFs on for collision avoidance. Networks evolving on manifolds have also been an active research area, as shown in power converters [18] and attitude tracking [19]. For interconnected systems on manifolds, a global state partition may not exist, coordinates are no longer globally defined, and identifying coordinates and subsystems is no longer straightforward. In recent work, Abdelgadil et al. Abdelgadil et al. [20] show how to model multi-agent systems on a manifold by splitting the tangent bundle via distributions. This type of splitting identifies subsystem velocity directions, but does not by itself provide compatible coordinates or a safety guarantee. In this paper, we construct invariant sets with a local product structure for interconnected systems evolving on manifolds. Following [20], subsystem directions are represented by distributions that are direct sums of the tangent bundle of the ambient manifold . These distributions generalise coordinates in curved spaces and can be assigned to each subsystem; we show that the distributions we build are involutive. This does not, by itself, provide compatible subsystem state coordinates. On these distributions, we introduce the notion of “adapted chart” which satisfies . As such, each subsystem distribution corresponds to its own coordinate block and, within these coordinates, one admissible input must satisfy all subsystem tangency conditions simultaneously. Furthermore, we give conditions for which a splitting supports local product representations of constraints and allows their control invariance to be verified through subsystem conditions.
For the problem formulated in Section II, this paper:
The paper establishes necessary and sufficient conditions for the existence of coordinates adapted to the given subsystem distributions in Proposition 1, and Section III. These coordinates allow subsystem constraints to be represented as a product without assuming the manifold is, globally, a Cartesian product.
Establishes the control invariance of compact product sets through subsystem tangency conditions satisfied by
a common input in Theorem 5, and describes the set and conditions intrinsically on in Section IV.
iii) Separates feasible subsystem input choices under Cartesian product input constraints and input alignment, without requiring independent subsystem dynamics in Corollary 1.
The existence of adapted charts and their associated distributions are conditions guaranteeing the existence of an agent interpretation of the non-linear dynamics and pave the way for decentralised and distributed safety certificates.
Notation: We use standard geometric notation from [21], [22]. Let be a smooth -dimensional manifold without boundary, with tangent spaces and tangent bundle . The differential of a smooth map at is , and is the set of smooth vector fields. For , the Lie bracket is for , where . A smooth distribution of constant rank is a smooth vector subbundle whose fibres are -dimensional. It is involutive if its smooth local sections are closed under Lie brackets, and integrable if every point lies in an immersed submanifold with for all . By the Frobenius theorem [23], is integrable if and only if it is involutive, and then every point has a neighbourhood and a submersion with . The symbol denotes the internal direct sum of subspaces, and the convex hull of a set . For , and a chart at , the contingent cone of at , adapted from [24], is
The same argument gives for a diffeomorphism between open sets containing , and for in the interior of . For a control system on with input set , an input is a measurable , and denotes the corresponding Carathéodory solution with . A set is control invariant if for every there is an input such that is defined and lies in for all . We write . The following lemma is the only viability result used in the paper.
Lemma 1 (Viability in a chart): Let with and smooth, and let be nonempty, convex and compact. Let be a chart and nonempty and compact. If
then is control invariant.
Problem Formulation
Given the interconnected dynamics on a manifold
where , , with , and is nonempty, convex and compact. Subsystems are not blocks of the state vector since may not be a product and coordinates may not respect interconnection. We first formally define our notion of tangent bundle splitting.
Definition 1 (Complementary tangent-bundle splitting): For smooth, constant rank , distributions , define a complementary tangent bundle splitting if for every . Each has a unique decomposition with and . The splitting fixes subsystem directions without reference to coordinates, while the dynamics and the inputs in (2) may still couple the subsystems. On , let be the injection onto the -th block, the -th coordinate subspace, the projection onto along the other blocks, and , , the standard basis of . To impose constraints subsystem by subsystem, we need coordinates , , in which motion along changes only, that is, a diffeomorphism
where and are open and is identified with . In such a chart, subsystem constraint sets , each nonempty and compact, combine into the candidate set
which is compact and contained in the chart domain. The construction is local and does not assume that is globally a product. Our objective, shown in Figure 1 is to determine when the given splitting supports a local product representation of the constraints and when the resulting set is control invariant.

Figure 1. Adapted product chart for . Thin curves are integral curves of (blue) and (teal); maps them to coordinate lines, so . The product set pulls back to , whose faces are integral curves of and whose faces are integral curves of . On the face , the component is tangent to the face, so only , equivalently , determines whether the trajectory leaves .
Problem 1: Given the network (2), and the splitting in Definition 1,
determine when charts of the form (3) exist;
give conditions on the subsystem velocities, imposed through the sets and met by a single admissible input at each , under which (4) is control invariant;
describe the resulting invariant set and the invariance condition on , independently of the chart, so that the safety guarantee does not depend on the coordinates used.
Control Invariance from Tangent-Bundle Splitting
For the network (2) and the complementary splitting introduced in Definition 1, we first describe the associated subsystem projections and conditions under which the distributions admit compatible local product coordinates. These coordinates decompose the state only under an integrability condition characterised in Section III-B.
Compatible subsystem structure
The first step, smoothly assigning each tangent vector its subsystem components, requires no integrability.
Lemma 2 (Smooth subsystem projections). Let be a complementary splitting with inclusions . For , let restrict on each to the projection onto along . Then each is a smooth bundle homomorphism over , and
In particular, satisfies .
Proof. The identities hold fibrewise by uniqueness of the decomposition , , and . Smoothness is local. Around any point of , fix a chart on which each has a smooth local frame , and let have as columns the coordinate representations of . Then is smooth, and is invertible for every by the direct-sum hypothesis. In the induced chart of , reads , which is smooth because . Hence is smooth, and so is , since is an embedded submanifold of containing .
Applying these projections gives the subsystem velocities
This decomposition is chart-free and exists for every splitting.
Adapted product charts
Interpreting the distributions as local subsystem state spaces requires additional integrability and compatibility. A state of subsystem is a submersion from an open subset of onto an open subset of whose rate depends only on the -th component. For every tangent vector over its domain,
By a rank count, (6) holds iff . Collecting states of all subsystems gives a chart of the form (3).
Definition 2 (Locally product-compatible splitting). A diffeomorphism satisfying (3) is an adapted product chart. A complementary splitting is locally product compatible if every point of has an open neighbourhood that is the domain of an adapted product chart.
For a diffeomorphism , condition (3) is equivalent to for every , that is, to each being a state of subsystem in the sense of (6). The coordinate is constant along : motion in the directions of the other subsystems does not change it (see Figure 1). The next result characterises when such charts exist.
Proposition 1 (Product-compatible splittings). Let be a smooth complementary splitting, , and . The following are equivalent:
i) the splitting is locally product compatible,
ii) is involutive for every ,
iii) around every point of there are local frames of such that for all and all ,
iv) around every point of there are local frames of such that for all and all . Moreover, the condition in iii) holds for one choice of local frames iff it holds for every choice.
Proof. i) iv): take the coordinate fields of an adapted chart. iv) iii): immediate. iii) ii): for and , their bracket satisfies
proving frame independence. By Lemma 2, every local section of is a sum of local sections of , , and for \Rightarrow$ i): fix . By the Frobenius theorem, for each there is a submersion from a neighbourhood of into with . Let on the intersection of these neighbourhoods. For , gives ,
while makes injective, hence bijective onto . Thus for all , and is invertible. By the inverse function theorem, maps an open diffeomorphically onto an open set; choose open with and set . ▪
Remark 1: For , condition ii) states that and are involutive. For , involutivity of every is necessary but not sufficient. On , the line fields , and are complementary and involutive, but .
In adapted coordinates, the projections of Lemma 2 become constant coordinate projections, so the velocity decomposition (5) becomes a decomposition into coordinate blocks.
Lemma 3 (Coordinate form of projections): Let be an adapted product chart, with . Then, for , , or equivalently
Conversely, a diffeomorphism satisfying the second identity for all is an adapted product chart.
Proof: Fix and . Adaptedness gives , so is the decomposition of along the splitting. By its uniqueness, , and applying gives the first identity. Conversely, taking images in the second identity gives . ▪
Control invariance of product sets
Let be an adapted product chart and consider the candidate set (4). Since , the transformation rule for contingent tangent cones gives
with , so tangency conditions on become conditions on a product set in .
First, we study the properties of contingent cones under the adapted chart. We introduce the notion of geometric derivability for the elements in the cotingent cone.
Definition 3 (Geometric derivability (Definition 6.1 [24])): Let and . A vector is derivable if there exist and with and . The set is geometrically derivable at if every is derivable.
In the above definition, no continuity of on is required. Clarke regular sets, in particular convex sets, are geometrically derivable [24 Ch. 6]; so are finite unions of closed convex sets, because the contingent cone of a finite union of closed sets at is the union of the contingent cones of the members that contain . The following lemma states the regularity property used to decompose the tangent cone of a product set.
Lemma 4 (Tangent cones of product sets): Let with , and let . Then
If, moreover, is geometrically derivable at for every except at most one, then
Proof: If , there are and with ; since for , the points lie in , so . If , there are and with , and the -th block gives with ; this is the first inclusion in (9). For the second, is a finite sum of elements of , hence lies in , because the convex hull of a cone is a convex cone.
For (10), we show . Let be such that is geometrically derivable at for all , and let . Choose and with . For , let and for all large enough that . Then and , so . ▪
Main contribution
For , the subsystem velocities in adapted coordinates are
so that and along trajectories in . By Lemma 5, is the -th coordinate block of the projected velocity .
Theorem 5 (Control invariance of a local product set): Let be an adapted product chart and let be as in (4). Suppose that for every there exists such that
Then:
for the same ; if, in addition, every except at most one is geometrically derivable at , then ;
is control invariant.
Proof: i) Fix and a corresponding , and write . By (11) and (12), , which is contained in by Lemma 4 and equals under the geometric derivability assumption. Since is a linear isomorphism, (8) gives , and i) follows. ii) By i), Condition (1) holds on . Since is compact and contained in the domain of the chart , Lemma 1 applies with . ▪
The input in (12) is common to all subsystems: feasibility of each condition with a different input does not imply (12). Theorem 5 is stated in the chart. We now describe and condition (12) on . For , let $A(x) = {i \in \mathcal{J}: z_i(x) \in } be the set of active subsystems, and let be the -th face of .
Proposition 2 (The candidate set on ): Let be an adapted product chart and let be as in (4). Then:
i) , and every level set of is an integral manifold of . The set is , and each face is a union of pieces of integral manifolds of .
ii) for and , iff , and implies for every ;
iii) if , then ; if , then .
Proof: i) Since , is equivalent to . Each is a submersion with , so its level sets are tangent to . Since is a homeomorphism onto the open set and each is closed, , and . For ii), Let . By Lemma 3, . By Lemma 4, iff : sufficiency is , and necessity is the first inclusion in (9). With (8), this gives the equivalence. If , then , so and .
Lastly, for iii), , is an interior point of , so and is geometrically derivable at . If , (10) and (8) give , which equals by (b) and contains . If , then is an interior point of .
Blockwise control feasibility
To obtain a blockwise control structure compatible with the tangent bundle splitting, suppose that
where , each is nonempty, convex and compact, , and each satisfies
Equivalently, for . Then for , so depends on the input only through , and we write
Corollary 1 (Blockwise feasibility): Let and be as in Theorem 5, and assume (13) and (14). For , let
If for every and every , then is control invariant.
Proof: Fix . Choose for , and arbitrary otherwise. Then by (13), and by (15), so the -th condition in (12) holds for ; otherwise, it holds because . Theorem 5 applies. ▪
Remark 2 (Decentralised feedback): Corollary 1 reduces the viability condition (1) to separate input choices when state constraints have an adapted product form, input constraints are independent, and (14) holds. This separates control channels without requiring independent subsystem dynamics or ensuring decentralised feedback, since may depend on the full state. Writing , a decentralised selection must satisfy for all , . Theorem 5 guarantees admissible input trajectories. Additional regularity conditions are needed to ensure that a state-dependent selection yields well-posed closed-loop dynamics.
Illustrative example
Consider the following Kuramoto oscillator comprising of two interconnected subsystems
defined on subject to input constraints , with and . The parameters of the system are: are the drift and the coupling strength. The distributions , , satisfy , and . The adapted chart , mapping phase arcs to intervals, assigns each phase a unique , see Figure 2 top and middle. We now construct the adapted chart by defining the angular variables , and modulo and . This transformation is a smooth diffeomorphism on with inverse , modulo . The associated vector fields and distributions are: , and let and . These satisfy following linear independence of and ; additionally, we have and and . Therefore, is adapted to the splitting because its differential maps each distribution onto the corresponding coordinate direction in the coordinates. The new dynamics are:

Figure 2. Adapted coordinates for the Kuramoto oscillator in (16). (top) Invariant set on with two trajectories (teal) and (purple) starting from the initial conditions (teal) and (purple) respectively towards (black). (bottom) Adapted chart image . For both and , the vector field (red) shows that all flows remain within these sets.
where , , and and . Under this transformation, each input map, however, is aligned with its corresponding distribution, i.e., to and to respectively.
The constraints on for are and . The set is clearly nonconvex, but the adapted charts transform it into a square. The inputs satisfy the tangent conditions whenever . Since the input maps are aligned with the splitting, Corollary 1 provides control invariance for the entire set. We note that sets
with cannot be made invariant. For invariance to hold, must be contained in every interval , . Since is increasing on this interval, a nonempty intersection requires . Due to this, if , then no nonempty rectangle contained in can keep the full range of the first phase.
The example shows how adapted coordinates enable subsystem tangency conditions to certify the entire curved region without restricting it to a rectangle in original coordinates. The synchronisation conditions in [18] do not directly certify this region, while the invariant set construction presented in [12] only applies to coupled linear systems.
Conclusion and Future Work
This paper addressed the issue of design of safety sets from a distributed point of view. We generalised the Euclidean notion of coordinates, i.e., assignment of vector field coordinates to agents, using distributions of the tangent bundle. Using these distributions, we have constructed control invariant sets with a product structure through a tangent bundle splitting and adapted coordinates. The torus example shows how adapted coordinates represent an entire curved constraint region as a product, whereas no contained rectangle in the original coordinates spans the full first phase range under the stated conditions. The construction, therefore, preserves the full constraint region while allowing for subsystem boundary checks. Future work will address the selection of compatible distributions and constraint sets, and extension to larger regions with complex or learned geometry.
References
- [1]F. Blanchini, “Set invariance in control,” Automatica, vol. 35, no. 11, pp. 1747–1767, 1999.DOI
- [2]J.-P. Aubin, Viability Theory, ser. Modern Birkhäuser Classics. Boston: Birkhäuser, 2009.
- [3]B. Hernandez, P. Baldivieso-Monasterios, and P. Trodden, “Distributed MPC: Guaranteeing Global Stability from Locally Designed Tubes,” IFAC-PapersOnLine, vol. 50, no. 1, pp. 11 829–11 834, jul 2017.DOI
- [4]P. Trodden, J. M. Maestre, and H. Ishii, “On the limits of robust control under adversarial disturbances,” IEEE Transactions on Automatic Control, pp. 1–16, 2026.arxiv.org/abs/2507.03630
- [5]P. R. Baldivieso-Monasterios and P. A. Trodden, “Coalitional predictive control: Consensus-based coalition forming with robust regulation,” Automatica, vol. 125, p. 109380, 2021.arxiv.org/abs/2001.10240
- [6]E. Masero, P. R. Baldivieso-Monasterios, J. M. Maestre, and P. A. Trodden, “Robust coalitional model predictive control with plug-and-play capabilities,” Automatica, vol. 153, p. 111053, 2023.DOI
- [7]A. D. Ames, X. Xu, J. W. Grizzle, and P. Tabuada, “Control barrier function based quadratic programs for safety critical systems,” IEEE Transactions on Automatic Control, vol. 62, no. 8, pp. 3861–3876, 2017.DOI
- [8]P. Rabiee and J. B. Hoagg, “Soft-minimum and soft-maximum barrier functions for safety with actuation constraints,” Automatica, vol. 171, p. 111921, Jan. 2025.DOI
- [9]S. Prajna, A. Jadbabaie, and G. J. Pappas, “A framework for worst-case and stochastic safety verification using barrier certificates,” IEEE Transactions on Automatic Control, vol. 52, no. 8, pp. 1415–1428, 2007.DOI
- [10]H. Wang, A. Papachristodoulou, and K. Margellos, “Distributed safe control design and probabilistic safety verification for multi-agent systems,” Automatica, vol. 179, p. 112393, 2025.DOI
- [11]S. Bansal, M. Chen, S. Herbert, and C. J. Tomlin, “Hamilton-Jacobi reachability: A brief overview and recent advances,” 2017 IEEE 56th Annu. Conf. Decis. Control. CDC 2017, vol. 2018–January, pp. 2242–2253, 2018.arxiv.org/abs/1709.07523
- [12]P. Nilsson and N. Ozay, “Synthesis of separable controlled invariant sets for modular local control design,” in 2016 American Control Conference (ACC). IEEE, 2016, pp. 5656–5663.arxiv.org/abs/1510.02990
- [13]A. Pirastehzad, A. van der Schaft, and B. Besselink, “Modular verification and design of interconnected linear systems using (γ, δ)-similarity,” Automatica, vol. 190, p. 113046, Aug. 2026.DOI
- [14]V. Jurdjevic, Geometric Control Theory. Cambridge University Press, 1996.DOI
- [15]F. Bullo and A. D. Lewis, Geometric Control of Mechanical Systems: Modeling, Analysis, and Design for Simple Mechanical Control Systems. Springer, 2005, vol. 49.DOI
- [16]M. de Sa, P. Ong, and A. D. Ames, “From bundles to backstepping: Geometric control barrier functions for safety-critical control on manifolds,” arXiv preprint arXiv:2510.20202, 2025.
- [17]T. Ibuki, S. Wilson, A. D. Ames, and M. Egerstedt, “Distributed collision-free motion coordination on a sphere: A conic control barrier function approach,” IEEE Control Systems Letters, vol. 4, no. 4, pp. 976–981, 2020.DOI
- [18]F. Dörfler and F. Bullo, “Synchronization and transient stability in power networks and nonuniform Kuramoto oscillators,” SIAM Journal on Control and Optimization, vol. 50, no. 3, pp. 1616–1642, 2012.
- [19]T. Lee, “Geometric tracking control of the attitude dynamics of a rigid body on SO(3),” in Proceedings of the 2011 American Control Conference. IEEE, 2011, pp. 1200–1205.
- [20]M. Abdelgali, M. Krstic, and J. I. Poveda, “Non-holonomic gradient play: Leafwise nash equilibria, stability, and deception,” Sep. 2026.arxiv.org/abs/2609.38051
- [21]M. P. do Carmo, Riemannian Geometry, ser. Mathematics: Theory & Applications. Boston: Birkhäuser, 1992.
- [22]J. M. Lee, Introduction to Riemannian Manifolds, 2nd ed., ser. Graduate Texts in Mathematics. Cham: Springer, 2018, vol. 176.DOI
- [23]J. M. Lee, Introduction to Smooth Manifolds, 2nd ed., ser. Graduate Texts in Mathematics. New York: Springer, 2013, vol. 218.
- [24]R. T. Rockafellar and R. J.-B. Wets, Variational Analysis. Springer, 1998.DOI