p-Rough Path Space Ωₚ(V)

• p-Rough path space Ωₚ(V) is the quotient of weakly geometric p-variation paths modulo tree-like equivalence, capturing essential non-parametric path information.
• It supports multiple natural topologies—induced, quotient, and metric—with distinct properties in separability, compactness, and metrizability.
• Its rich algebraic and differential structures enable applications in signature-based machine learning, rough integration, and solving rough differential equations.

A $p$-rough path space, denoted $Ω_p(V)$, is the canonical quotient space of weakly geometric $p$-variation rough paths over a Banach or inner-product space $V$, modulo tree-like equivalence. It is fundamental to the modern, parameterization-invariant theory of signatures for paths, and serves as the domain for signature-based machine learning, rough integration, and differential equations driven by highly irregular signals. The construction, topological structure, and metric properties of $Ω_p(V)$ encode the essential non-parametric, pathwise information arising from multidimensional iterated integration.

1. Construction of $Ω_p(V)$ and Tree-like Equivalence

Let $V$ be a finite-dimensional inner-product space. The space of $p$-weakly geometric rough paths, $\mathrm{WG}\Omega_p(V)$, consists of continuous paths $X : [0, T] \rightarrow G^{(m)}$ with $X_0 = 1$, where $G^{(m)}$ is the step-$m$ free nilpotent group over $V$ (with $m = \lfloor p \rfloor$). These paths have finite $p$-variation: $$\|X\|_{p\text{-var};[0,T]} := \left(\sup_{\mathcal P}\sum_{i} d\big(X_{t_i}, X_{t_{i+1}}\big)^p\right)^{1/p} < \infty$$ with distance induced by the Carnot–Carathéodory metric. Lyons’s extension theorem ensures a unique continuous lift of each $X \in \mathrm{WG}\Omega_p(V)$ to the full group-like elements $G^{(*)}$ in the tensor algebra, with the signature map $\text{SS}(X) := X_T^{<\infty}$ (the total signature).

Tree-like equivalence $X \sim_\tau Y$ holds if and only if the concatenation $X * (\text{reversal of } Y)$ factors through a real tree, equivalently if and only if their signatures coincide, $SS(X) = SS(Y)$. Each equivalence class $[X]$ contains a unique tree-reduced representative, up to reparameterization.

Define the quotient space

$$Ω_p(V) := \mathrm{WG}\Omega_p(V) \,/\, \sim_\tau = \{ [X] : X \in \mathrm{WG}\Omega_p(V) \}$$

This identifies parameterized paths under the robust signature invariant, removing redundant parameterization and “tree-like” excursions, and thus encodes only essential pathwise information (Cass et al., 2024).

2. Topologies on $Ω_p(V)$

Three natural classes of topologies can be placed on $Ω_p(V)$ (Cass et al., 2024):

• Induced (metrizable) topologies $\chi_m$: Any metric $m$ on $Ω_p(V)$ for which the signature map $$SS: (\mathrm{WG}\Omega_p(V), d_{p\text{-var}}) \rightarrow (Ω_p(V), m)$$ is continuous defines a metrizable topology. Typically, continuity is reinforced by requiring $SS$ to be continuous when $\mathrm{WG}\Omega_p(V)$ is equipped with the $q$-variation metric for some $q>p$.
• The quotient topology $\chi_\tau$: This is the final topology making the projection $\pi: \mathrm{WG}\Omega_p(V) \rightarrow Ω_p(V)$, $X \mapsto [X]$, continuous when $\mathrm{WG}\Omega_p(V)$ carries its $p$-variation topology. This topology is Hausdorff but not first-countable, not regular, and therefore not metrizable.
• The metric topology $\chi_{\mathscr D}$: For two equivalence classes $[X], [Y]$ with tree-reduced, Hölder-control parameterizations $X^\star, Y^\star$, set

$$\mathscr d\left([X], [Y]\right) := d_{p\text{-var}}(X^\star, Y^\star)$$

This is a well-defined metric on $Ω_p(V)$, giving rise to the topology $\chi_{\mathscr D}$.

Each inclusion

$$\chi_m \subsetneq \chi_\tau \subsetneq \chi_{\mathscr D}$$

is strict, and all these topologies are Hausdorff (Cass et al., 2024).

3. Topological and Metric Properties

The main topological properties of $Ω_p(V)$ under these topologies include (Cass et al., 2024):

• Separation and compactness: Under any metric $m$ for which $SS$ is continuous from the $q$-variation topology ($q>p$), $(Ω_p(V), \chi_m)$ is separable, $\sigma$-compact, and Lusin, but not a Baire space, not locally compact, nor completely metrizable.
• Quotient topology ($\chi_\tau$): Hausdorff, but not metrizable due to failure of first-countability and regularity.
• Metric topology ($\chi_{\mathscr D}$): Hausdorff and separable, but not complete for $p > 1$. However, for $p=1$, $(Ω_1(V), \chi_{\mathscr D})$ is Polish (separable and completely metrizable). Furthermore, in any of these topologies, compact sets of $Ω_p(V)$ cannot contain open balls. Balls of bounded $p$-variation are compact in $\chi_m$, yet every nonempty $\chi_m$-open set contains classes with arbitrarily large $p$-variation. None of these topologies is locally compact for $p > 1$.

A summary of topological features:

Topology	Metrizable	Separable	Complete	Locally Compact	Polish
$\chi_m$	Yes	Yes	No	No	No
$\chi_\tau$	No	—	—	—	No
$\chi_{\mathscr D}$	Yes	Yes	No ($p>1$)	No	Yes ($p=1$)

Proofs use the uniqueness of tree-reduced representatives, Hölder-control reparameterizations, and compactness criteria via Arzelà-Ascoli-type arguments, while non-local compactness and incompleteness reflect the possibility of concatenations with long excursions (Cass et al., 2024).

4. Functional-Analytic Structure and Perturbations

There exists a distinguished vector subspace $\mathscr H^p(V) \subset Ω_p(V)$ with a well-defined addition $\boxplus$ and scalar multiplication $\odot$, making $\mathscr H^p(V)$ into a vector space isomorphic to a class of additive indexed functionals $\mathfrak I^p(V)$ (Geller et al., 21 Jan 2026). The operations are constructed as follows:

• Addition $\boxplus$: For $H, \widetilde H \in \mathscr H^p(V)$, form the pointwise sum at each level, then apply the rough path sewing lemma to restore multiplicativity.
• Scalar multiplication $\odot$: Induced via a mapping between $\mathscr H^p(V)$ and the additive index class.
• Extension to $Ω^p(V) \times \mathscr H^p(V)$: For any base rough path $X$ and perturbation $H$, $X\boxplus H$ is defined via rough sewing.

Key structural results:

• Associativity: $$(X\boxplus H) \boxplus \widetilde H = X\boxplus (H\boxplus\widetilde H)$$.
• Trivial kernel: $X\boxplus H = X$ if and only if $H$ is the neutral element.
• Displacement invariance: Enlarging $\mathscr H^p(V)$ to the space of almost rough paths $\mathscr H^{\mathrm{am},p}(V)$ does not enlarge the set $\{X\boxplus H\}$.

This algebraic structure encodes the perturbation theory and “displacement orbits” within the rough path space (Geller et al., 21 Jan 2026).

5. Differential and Metric Geometry of $Ω_p(V)$

The differential structure of $Ω_p(V)$ is formulated by associating to each path a tangent space of equivalence classes of curves, capturing the nontrivial differentiable directions of functionals on rough path space (Qian et al., 2011). The construction utilizes:

• Variational curves: For a base path $X$ and direction $Y$, form variations $X(\varepsilon)$ corresponding to infinitesimal deformations.
• Tangent vectors: Identified as pairs $[Z, \phi]$, where $Z$ is a higher-level rough path (encoding perturbation) and $\phi$ an independent second-level increment.
• Differentiability: For any $F: WGΩ_p(V) \rightarrow W$, and $[Z, \phi]$ at $X$, the directional derivative $D_v F(X)$ is defined as the limit along the variational curve.
• Flow equations: One can solve the rough flow equation $U'(t) = F(U(t))$ with $U(0) = X_0$ whenever $F$ is locally (or globally) Lipschitz in the appropriate rough-variation metric, with existence and uniqueness results established via compactness and a Grönwall-type argument.

This structure enables the analysis of smooth functionals, variational calculus, and the flow of rough differential equations within the rough path manifold (Qian et al., 2011).

6. Applications, Functional Analysis, and Further Directions

The quotient rough path space $Ω_p(V)$ is essential for the non-parametric analysis of signatures and the formulation of integration, control, and differential equations driven by highly non-smooth (e.g., stochastic or arbitrary) signals. The topologies ensure invariance under parameterization and the robustness of path signature-based analysis.

The results on topological and metric properties have direct implications for learning algorithms based on signatures, the approximation and perturbation theory of rough paths, and the well-posedness of rough differential equations in both Banach and manifold contexts. The structural understanding of vector spaces and tangent directions underlies advanced functional-analytic and algebraic approaches in the theory of rough paths (Cass et al., 2024, Geller et al., 21 Jan 2026, Qian et al., 2011).

7. Connections to the Signature Map and Compactness

The signature map $SS: \mathrm{WG}\Omega_p(V) \rightarrow G^{(*)}$ plays a central role: $Ω_p(V)$ is precisely the space of paths identified up to identical signatures. Under appropriate topologies, $SS$ is continuous with respect to the $q$-variation metric ($q > p$), and its level sets coincide exactly with tree equivalence classes. Compactness of $p$-variation balls in $Ω_p(V)$ follows, but these have empty interior, underscoring the richness of the space. When $p = 1$, the metric topology is completely metrizable (Polish), providing a firm analytic foundation (Cass et al., 2024).

The analytical and algebraic properties of $Ω_p(V)$ therefore form the bedrock for much of modern rough path theory and its applications in stochastic analysis, geometric data analysis, and dynamical systems.