Isometric Embedding & Generalized Synchronization

• Isometric embedding generalized synchronization is a technique that embeds a driving system’s dynamics into a reservoir state using a metric-preserving map.
• Continuous and discrete-time frameworks demonstrate how contraction conditions and Nash’s theorem guarantee a topologically faithful, smooth embedding of attractors.
• Applications in linear reservoirs and Takens’ delay-coordinate mappings illustrate its practical use in forecasting, data reconstruction, and robust dynamical modeling.

Isometric embedding generalized synchronization denotes a regime in which generalized synchronization $x=f(m)$ between a driving dynamical system and a driven reservoir or response system is not merely a functional dependence or a topological embedding, but an embedding that preserves the relevant metric structure. In the reservoir-computing literature, generalized synchronization is explicitly used as the mechanism that embeds the hidden source dynamics into the reservoir state space, and the synchronization map $f$ is the object that carries this representation. Continuous-time results establish conditions for existence, smoothness, and topological embedding properties of $f$, while discrete-time results show that a generic reservoir system admits a generalized synchronization that is a topological embedding of the input system’s attractor and that, for sufficiently high reservoir dimension given by Nash’s embedding theorem, there exists an isometric embedding generalized synchronization (Hart, 2022, Hart, 29 Aug 2025).

1. Formal notion of generalized synchronization and embedding

In continuous time, the source is a smooth vector field $\mathcal{V}$ on a smooth manifold $M$, with flow

$$\{\phi^t \in \mathrm{Diff}^1(M)\mid t\in\mathbb{R}\},\qquad \phi^{t_1+t_2}=\phi^{t_1}\phi^{t_2},$$

and observation function

$\omega \in C^0(M,\mathbb{R}^d).$

A continuous-time reservoir driven by this observation is written as

$\dot{x}(t)=F\big(x(t),\omega(\phi^t(m))\big),$

with $$F\in C^1(\mathbb{R}^N\times\mathbb{R}^d,\mathbb{R}^N)$$, Lipschitz in $x$. The reservoir system admits a $f$0-generalized synchronisation $f$1 if, for any initial $f$2 and any $f$3,

$f$4

A stronger uniform version requires a function $f$5 such that

$f$6

for all $f$7 (Hart, 2022).

In discrete time, the input system is $f$8 on a smooth manifold $f$9, with scalar observable $f$0, and the reservoir dynamics are

$f$1

where $f$2. On a $f$3-invariant compact set $f$4, generalized synchronization means that there exists $f$5 such that the graph

$f$6

is invariant,

$f$7

and nearby reservoir states converge to that graph along the input trajectory (Hart, 29 Aug 2025).

The embedding question concerns the regularity and injectivity of this synchronization map. A map $f$8 is an embedding of $f$9 if it is injective on $\mathcal{V}$0, a homeomorphism from $\mathcal{V}$1 onto $\mathcal{V}$2, and, in the $\mathcal{V}$3 context, typically immersive on $\mathcal{V}$4. The isometric refinement assumes that $\mathcal{V}$5 carries a Riemannian metric $\mathcal{V}$6, and requires

$\mathcal{V}$7

for all $\mathcal{V}$8 and all tangent vectors $\mathcal{V}$9. In that case, the synchronized reservoir state reproduces not only the topology and dynamics of $M$0, but also its Riemannian metric structure (Hart, 29 Aug 2025).

2. Continuous-time reservoir computers: synchronization as a Takens-like embedding mechanism

For continuous-time reservoir computers, a central structural result is that uniform $M$1-generalized synchronization is equivalent to uniform $M$2-asymptotic stability with respect to the family of inputs $M$3. A simple sufficient condition is a contraction condition in $M$4: if $M$5 is bounded, convex, $M$6-invariant, and there exists $M$7 such that

$M$8

for all $M$9 and all $$\{\phi^t \in \mathrm{Diff}^1(M)\mid t\in\mathbb{R}\},\qquad \phi^{t_1+t_2}=\phi^{t_1}\phi^{t_2},$$0, then the reservoir system is uniformly $$\{\phi^t \in \mathrm{Diff}^1(M)\mid t\in\mathbb{R}\},\qquad \phi^{t_1+t_2}=\phi^{t_1}\phi^{t_2},$$1-asymptotically stable and therefore admits a uniform $$\{\phi^t \in \mathrm{Diff}^1(M)\mid t\in\mathbb{R}\},\qquad \phi^{t_1+t_2}=\phi^{t_1}\phi^{t_2},$$2-GS $$\{\phi^t \in \mathrm{Diff}^1(M)\mid t\in\mathbb{R}\},\qquad \phi^{t_1+t_2}=\phi^{t_1}\phi^{t_2},$$3 (Hart, 2022).

When a uniform GS exists, it solves the quasilinear PDE

$$\{\phi^t \in \mathrm{Diff}^1(M)\mid t\in\mathbb{R}\},\qquad \phi^{t_1+t_2}=\phi^{t_1}\phi^{t_2},$$4

which expresses that the pushforward of the source flow under $$\{\phi^t \in \mathrm{Diff}^1(M)\mid t\in\mathbb{R}\},\qquad \phi^{t_1+t_2}=\phi^{t_1}\phi^{t_2},$$5 matches the reservoir vector field along the synchronization manifold $$\{\phi^t \in \mathrm{Diff}^1(M)\mid t\in\mathbb{R}\},\qquad \phi^{t_1+t_2}=\phi^{t_1}\phi^{t_2},$$6. The same framework also allows multiple generalized synchronisations simultaneously, each relative to a different invariant subset $$\{\phi^t \in \mathrm{Diff}^1(M)\mid t\in\mathbb{R}\},\qquad \phi^{t_1+t_2}=\phi^{t_1}\phi^{t_2},$$7, and this is connected to the multi-Echo-State-Property. In embedding language, different disjoint attracting sets can support multiple parallel embeddings of the same source manifold into different regions of reservoir space (Hart, 2022).

For a linear continuous-time reservoir

$$\{\phi^t \in \mathrm{Diff}^1(M)\mid t\in\mathbb{R}\},\qquad \phi^{t_1+t_2}=\phi^{t_1}\phi^{t_2},$$8

with $$\{\phi^t \in \mathrm{Diff}^1(M)\mid t\in\mathbb{R}\},\qquad \phi^{t_1+t_2}=\phi^{t_1}\phi^{t_2},$$9 symmetric positive definite and $\omega \in C^0(M,\mathbb{R}^d).$0, the GS map is given in closed form by

$\omega \in C^0(M,\mathbb{R}^d).$1

This map is a linear filter of the past observation signal and is described as a continuous-time analogue of a linear time-invariant reservoir mapping the delay coordinates of the source into reservoir space. The paper characterizes this as closely related to Takens’ embedding theorem: generalized synchronization in a linear reservoir is a Takens-like embedding mechanism that transforms time series from a dynamical system into a higher-dimensional representation that is locally diffeomorphic to the original system on a subset of interest (Hart, 2022).

Under additional assumptions—$\omega \in C^0(M,\mathbb{R}^d).$2, bounded $\omega \in C^0(M,\mathbb{R}^d).$3 and $\omega \in C^0(M,\mathbb{R}^d).$4, $\omega \in C^0(M,\mathbb{R}^d).$5, and controlled growth of the backward flow—the GS map is of class $\omega \in C^0(M,\mathbb{R}^d).$6. For a source vector field with finitely many fixed points, distinct eigenvalues at each fixed point, and linearly independent vectors

$\omega \in C^0(M,\mathbb{R}^d).$7

the GS map is a topological embedding on the set of fixed points for generic $\omega \in C^0(M,\mathbb{R}^d).$8. The same linear independence condition holds almost surely for randomly generated symmetric positive definite $\omega \in C^0(M,\mathbb{R}^d).$9 and random $\dot{x}(t)=F\big(x(t),\omega(\phi^t(m))\big),$0, so embedding of fixed points occurs almost surely for random linear reservoirs (Hart, 2022).

3. Generic embedding and isometric embedding in discrete-time reservoirs

The discrete-time embedding theory is formulated on a $\dot{x}(t)=F\big(x(t),\omega(\phi^t(m))\big),$1-invariant compact attractor $\dot{x}(t)=F\big(x(t),\omega(\phi^t(m))\big),$2. Whitney’s theorem is invoked in the form: if $\dot{x}(t)=F\big(x(t),\omega(\phi^t(m))\big),$3 has dimension $\dot{x}(t)=F\big(x(t),\omega(\phi^t(m))\big),$4 and the reservoir dimension satisfies

$\dot{x}(t)=F\big(x(t),\omega(\phi^t(m))\big),$5

then the set of embeddings $\dot{x}(t)=F\big(x(t),\omega(\phi^t(m))\big),$6 is open and dense. On that basis, and under Takens-type nondegeneracy conditions—$\dot{x}(t)=F\big(x(t),\omega(\phi^t(m))\big),$7 with finitely many periodic orbits in $\dot{x}(t)=F\big(x(t),\omega(\phi^t(m))\big),$8, and for each periodic orbit $\dot{x}(t)=F\big(x(t),\omega(\phi^t(m))\big),$9 of period $$F\in C^1(\mathbb{R}^N\times\mathbb{R}^d,\mathbb{R}^N)$$0, the derivative $$F\in C^1(\mathbb{R}^N\times\mathbb{R}^d,\mathbb{R}^N)$$1 has $$F\in C^1(\mathbb{R}^N\times\mathbb{R}^d,\mathbb{R}^N)$$2 distinct eigenvalues—the generalized synchronization of a generic reservoir system is an embedding of the attractor (Hart, 29 Aug 2025).

More precisely, for generic observation functions $$F\in C^1(\mathbb{R}^N\times\mathbb{R}^d,\mathbb{R}^N)$$3 and generic reservoir maps $$F\in C^1(\mathbb{R}^N\times\mathbb{R}^d,\mathbb{R}^N)$$4 among those maps that admit a GS, the generalized synchronization

$$F\in C^1(\mathbb{R}^N\times\mathbb{R}^d,\mathbb{R}^N)$$5

is an embedding of $$F\in C^1(\mathbb{R}^N\times\mathbb{R}^d,\mathbb{R}^N)$$6. The reservoir dynamics restricted to the GS attractor are then conjugate to the input dynamics restricted to $$F\in C^1(\mathbb{R}^N\times\mathbb{R}^d,\mathbb{R}^N)$$7: $$F\in C^1(\mathbb{R}^N\times\mathbb{R}^d,\mathbb{R}^N)$$8 In this setting, generalized synchronization gives a functional encoding $$F\in C^1(\mathbb{R}^N\times\mathbb{R}^d,\mathbb{R}^N)$$9, and the generic embedding result guarantees that this encoding is injective and topologically faithful (Hart, 29 Aug 2025).

The isometric extension assumes that $x$0 carries a Riemannian metric $x$1. A version of the Nash–Günther theorem is used in the form: if

$x$2

then there exists an isometric embedding $x$3. Combining this with the generic embedding GS theorem yields the existence statement: under the same periodic-orbit hypotheses on $x$4, and for generic observation functions $x$5, there exists a reservoir map $x$6 whose generalized synchronization

$x$7

is an isometric embedding of $x$8 (Hart, 29 Aug 2025).

This theorem is an existence theorem. It shows that, at sufficiently high reservoir dimension, one can construct a reservoir whose synchronized attractor is exactly the Nash isometric embedding of $x$9. The paper also states that these bounds are not tight. A plausible implication is that the result is primarily geometric and structural rather than a statement about typical low-dimensional engineering practice (Hart, 29 Aug 2025).

4. Linear reservoirs, explicit constructions, and the meaning of “isometric”

For linear reservoirs in discrete time,

$f$00

generalized synchronization typically has the form

$f$01

under suitable stability and convergence conditions such as $f$02. Two linear reservoir maps $f$03 and $f$04 are linearly isomorphic if there exists an invertible matrix $f$05 such that

$f$06

This coordinate freedom is used to obtain isometric generalized synchronization explicitly (Hart, 29 Aug 2025).

Under the assumptions that the eigenvalues of $f$07 are distinct, that $f$08 for all eigenvectors $f$09, that the GS series converges, and that the vectors

$f$10

are linearly independent, the paper constructs a new linear reservoir

$f$11

where $f$12 expresses an orthonormal basis of $f$13 with respect to $f$14, $f$15 is built from the columns

$f$16

and $f$17 is any rotation matrix. The associated GS

$f$18

is then an isometric embedding with respect to the metric $f$19 (Hart, 29 Aug 2025).

The continuous-time theory is more cautious. It proves smooth topological embedding on fixed points, and the details state that these results come close to bi-Lipschitz behavior at least locally, though the paper does not prove strict isometries. It also states explicitly that the paper does not claim that the GS $f$20 is an isometry, nor does it provide explicit bi-Lipschitz inequalities. Instead, it establishes topological embedding on fixed points and local diffeomorphic structure, preserving manifolds, dimensions, smooth structure, and linearization spectra. This suggests a near-isometric local picture in the sense of bounded local distortion, but the strict metric statement is reserved for the later discrete-time isometric theory (Hart, 2022).

The distinction is therefore precise. Topological embedding generalized synchronization preserves qualitative geometry and dynamical conjugacy. Isometric embedding generalized synchronization adds preservation of the Riemannian metric: $f$21 The former is generic at Whitney dimension; the latter is an existence result at Nash dimension, with an explicit coordinate construction in the linear case (Hart, 29 Aug 2025).

5. Relation to Takens, topological conjugacy, and broader notions of generalized synchronization

The reservoir-computing embedding results are explicitly connected to Takens’ theorem. Takens’ delay-coordinate map

$f$22

is the classical model of embedding by observation history. In the continuous-time linear reservoir, the synchronization map is a linear filter over all past times, and in the discrete-time theory a delay-line reservoir with

$f$23

has GS

$f$24

which is exactly the Takens delay map (Hart, 2022, Hart, 29 Aug 2025).

Outside reservoir computing, generalized synchronization has also been described as a geometric and topological relation between attractors. In a flat-coupled drive–response setting, the response attractor can, in strong cases, be regarded as a high-fidelity copy of the drive attractor: the response attractor can be embedded in a space of dimension equal to the drive’s, the first-return maps can be conjugated, and the map $f$25 between attractors behaves like a homeomorphism onto its image. The same paper is explicit that it does not use the term “isometric embedding” and that the mappings under discussion are topological or differentiable rather than distance-preserving. Its Type I and Type II generalized synchronization correspond to conjugated or $f$26-conjugated first-return maps, whereas Type III and Type IV involve multivalued or discontinuous mappings and loss of conjugacy (Letellier et al., 2024).

This broader comparison clarifies a recurring misconception. Generalized synchronization does not by itself imply metric preservation. In the reservoir setting, topological faithfulness arises generically under Whitney-type dimensional assumptions, and metric exactness requires the stronger Nash-type construction. In the flat-coupled systems taxonomy, strong generalized synchronization yields a topological embedding of the dynamics but not an isometric one. A plausible implication is that “embedding” and “isometric embedding” should be treated as distinct layers of structure rather than interchangeable descriptions (Hart, 29 Aug 2025, Letellier et al., 2024).

6. Dynamical consequences, robustness, and open problems

Once an embedding GS has been achieved, the continuous-time reservoir paper argues that the universal approximation theorem makes it possible to forecast the future dynamics of the source system and replicate its topological properties. In the linear case, one may define an autonomous reservoir system

$f$27

and approximate the unknown readout $f$28 with a random neural network. The paper proves a central limit theorem for such random neural networks: if

$f$29

then

$f$30

In that sense, generalized synchronization, embedding, and approximation together provide a route to reconstructing the source vector field on the synchronization manifold (Hart, 2022).

The Lorenz-63 example illustrates the local fidelity of this construction. A linear reservoir of dimension $f$31 with

$f$32

is driven by the scalar observation $f$33, and the fixed point

$f$34

is mapped to

$f$35

The reservoir states projected onto the first three principal components show a shape diffeomorphic to the Lorenz attractor, and the eigenvalues of the Jacobian of the autonomous reservoir dynamics at $f$36 match those of the source Jacobian at $f$37. The paper describes this as strong evidence of a diffeomorphic embedding there and notes that it is not an isometric embedding, but a spectrally faithful embedding (Hart, 2022).

The same continuous-time framework includes a noise-robustness result. For noisy observations modeled by

$f$38

the reservoir states converge in distribution to

$f$39

When $f$40 is constant, the stochastic error is a multivariate Ornstein–Uhlenbeck process. Thus the deterministic GS is preserved, and noisy reservoir states are random perturbations around the synchronization manifold (Hart, 2022).

A recent synthesis in theoretical neuroscience extends the embedding viewpoint further. It argues that a contractive recurrent circuit driven by structured sensory input can synchronize to the driving dynamics and that, under generic embedding conditions developed in the reservoir-computing literature, the resulting synchronization map can embed the low-dimensional sensory manifold into neural state space. It also states that for sufficiently high dimension satisfying Nash’s bound, an isometric embedding generalized synchronization exists, and proposes that Hebbian plasticity may crystallize the embedded manifold into recurrent connectivity. The central open problem is whether the Hebbian fixed point exists and preserves the embedding quality of the synchronization manifold (O'Reilly-Shah, 5 May 2026).

In this broader perspective, isometric embedding generalized synchronization is the strongest currently articulated form of the idea that synchronization can serve as an embedding mechanism. The established results separate three levels with some precision: functional dependence $f$41, generic topological embedding of the attractor, and isometric embedding at sufficiently high dimension. What remains open is the extent to which practical architectures, noisy finite-dimensional systems, and learned autonomous dynamics preserve not only topological faithfulness but also the metric structure encoded by an isometric synchronization map (Hart, 29 Aug 2025, O'Reilly-Shah, 5 May 2026).