Attractor Sculpting in Dynamical Systems

• Attractor sculpting is a design-oriented approach that intentionally modifies invariant sets’ stability, geometry, and topology across diverse dynamical systems.
• Various methods—such as periodic skeletons in reservoir computing, A-portrait diagnostics, and linking matrix validation—enable precise control over attractors’ properties.
• Sculpting leverages local stability analysis and bifurcation thresholds to engineer desired chaotic, periodic, or topological behaviors in complex systems.

Attractor sculpting denotes a family of interventions on dynamical invariant sets in which the target of design is not only stability, but also geometry, topology, or the spatial distribution of attraction and repulsion along the set. In the works grouped under this label, sculpting is realized in several distinct ways: by using a periodic skeleton to induce a chaotic attractor with prescribed shape in a reservoir computer, by classifying and selecting scaling critical points in two-field inflation, by revealing hidden organizing structures through generalized Floquet exponents and attractiveness portraits, and by constructing genus–1 chaotic templates from linking matrices (Kabayama et al., 2024, Christodoulidis et al., 2019, Guan, 2014, Olszewski et al., 2018).

1. Conceptual scope

The cited literature does not use a single universal formalism for attractor sculpting. Instead, it develops several operational meanings of the term, each tied to a different class of dynamical systems and a different control lever. The common feature is intentional intervention in the structure of an attractor or invariant set rather than passive observation.

Context	Object being sculpted	Principal mechanism
Reservoir computing	Geometry of a chaotic attractor	Periodic skeleton, teacher forcing, edge-of-chaos bifurcation
Two-field inflation	Background scaling attractor	Field-space curvature, potential gradients, bifurcation of critical points
Dynamical-systems diagnostics	Internal organization of invariant sets	GFE and A-portrait reveal local attraction/repulsion
Genus–1 templates	Topological template of a chaotic attractor	Linking matrix validation, crossing-order extraction, minimal-height drawing

In this broad sense, attractor sculpting may mean designing a phase portrait so that it follows a specified contour, selecting a non-geodesic inflationary trajectory with prescribed turning properties, exposing hidden periodic skeletons embedded in complicated attractors, or enforcing a target topological template through torsions and permutations. This suggests that the notion is best understood as a design-oriented viewpoint on attractors rather than a single algorithmic doctrine.

2. Semi-supervised geometry control in reservoir computing

In "Designing Chaotic Attractors: A Semi-supervised Approach" (Kabayama et al., 2024), attractor sculpting is defined as intentional shaping of a chaotic attractor’s geometry by using a simple periodic time series as a template, called a skeleton, and letting a reservoir computer fail "just right" at learning that template. The dynamical substrate is an $N$-dimensional leaky integrator echo state network (LESN) with $D$-dimensional input $u_k$, state $x_k \in \mathbb{R}^N$, element-wise $\tanh$ nonlinearity, leak rate $a \in (0,1]$, internal weights $W$, and input weights $W_{\mathrm{in}}$. In driven mode, the state update is

$$x_{k+1} = (1-a)x_k + a \cdot \tanh(\rho W x_k + \sigma W_{\mathrm{in}} u_k).$$

The linear readout produces $z_k = W_{\mathrm{out}}^\top x_k$, and $D$0 is fitted by ridge regression,

$D$1

After training, the closed-loop system replaces the teacher signal by the model output and evolves as

$D$2

with post-training effective recurrent matrix $D$3.

The skeleton is a periodic time series chosen to encode the desired contour. A canonical example is the $D$4D Lissajous curve

$D$5

which yields a figure-8-like template. Open-loop learning quality is quantified by

$D$6

and successful open-loop learning is defined empirically by $D$7. Closed-loop chaos is diagnosed by the maximum Lyapunov exponent, $D$8, while geometry is assessed by a deviation index $D$9. For the Lissajous skeleton, eliminating time yields the curve constraint $u_k$0, and the deviation index is

$u_k$1

A trajectory is considered "along" the skeleton when $u_k$2.

The core mechanism is bifurcation-induced training failure. The paper distinguishes pre-training and post-training effective spectral radii,

$u_k$3

but emphasizes that in driven reservoirs stability is governed by the conditional Lyapunov exponent (CLE). External input suppresses reservoir chaos; even with $u_k$4, driven reservoirs can converge if the CLE is negative. As $u_k$5 increases, the driven CLE rises and crosses zero at the edge of chaos, denoted $u_k$6. Near this crossing, convergence slows, the closed-loop periodic orbit embedded by training becomes unstable, and small prediction errors accumulate rather than dissipate. The skeleton then survives as an unstable periodic orbit, visible as a long transient aligned with the template, before destabilizing via period-doubling and crisis events into an untrained chaotic attractor whose output remains close to the skeleton’s global geometry.

The algorithmic workflow is explicit: choose a skeleton; initialize $u_k$7 and $u_k$8; train $u_k$9 by teacher forcing; retain settings with $x_k \in \mathbb{R}^N$0; estimate the driven CLE as a function of $x_k \in \mathbb{R}^N$1 to locate $x_k \in \mathbb{R}^N$2; find a supervised point $x_k \in \mathbb{R}^N$3 with $x_k \in \mathbb{R}^N$4 and strong shape alignment; then sweep $x_k \in \mathbb{R}^N$5 and identify semi-supervised points $x_k \in \mathbb{R}^N$6 where $x_k \in \mathbb{R}^N$7 and $x_k \in \mathbb{R}^N$8 remains small. Validation uses Lyapunov exponents, Kaplan–Yorke dimension,

$x_k \in \mathbb{R}^N$9

and phase-portrait inspection.

The Lissajous example is the canonical quantitative demonstration. With $\tanh$0, the edge of chaos was found around $\tanh$1, where the driven CLE changes sign. The convergent region $\tanh$2 has $\tanh$3 and closed-loop periodic attractors; the chaotic region $\tanh$4 has $\tanh$5 and $\tanh$6. Near $\tanh$7, specifically $\tanh$8, the system exhibits alternating windows of periodic solutions $\tanh$9 and band chaos $a \in (0,1]$0 with $a \in (0,1]$1, consistent with period-doubling followed by a crisis. Other skeletons yield analogous outcomes: a Van der Pol template gives a chaotic attractor with $a \in (0,1]$2; a hand-drawn "@" symbol gives $a \in (0,1]$3; a Rössler limit cycle in $a \in (0,1]$4D gives $a \in (0,1]$5; and a piano triad waveform yields a chaotic output with $a \in (0,1]$6. Typical settings are $a \in (0,1]$7, $a \in (0,1]$8, $a \in (0,1]$9, and $W$0.

Several limitations are intrinsic to the method. Periodic or quasi-periodic skeletons with clear contours are most suitable; overly complex or noisy templates may degrade alignment or yield trivial dynamics. The semi-supervised interval may contain multiple disjoint periodic windows, slow convergence near $W$1 makes results sensitive to $W$2, and greater chaos intensity may reduce geometric fidelity. Success is therefore non-guaranteed, although semi-supervised points $W$3 are reported consistently between $W$4 and $W$5 across seeds.

3. Local attraction, repulsion, and hidden organizing structure

A different sense of attractor sculpting appears in "Generalized Floquet Exponent, Attractiveness Portrait and Structure Hidden in an Attractor" (Guan, 2014). Here the emphasis is not on constructing an attractor from a prescribed signal, but on revealing, analyzing, and thereby reshaping the internal organization of invariant sets by identifying where attraction and repulsion are distributed. Two tools are central: the generalized Floquet exponent (GFE), which is a finite-window analogue of Floquet and Lyapunov exponents, and the attractiveness portrait, or A-portrait, which is a directional field drawn along trajectories.

For a trajectory $W$6 of an autonomous system $W$7, let $W$8. Over a finite window $W$9, define the principal fundamental matrix by

$W_{\mathrm{in}}$0

The generalized Floquet exponents are then the eigenvalues of the logarithmic monodromy rate,

$W_{\mathrm{in}}$1

equivalently $W_{\mathrm{in}}$2 where $W_{\mathrm{in}}$3 are eigenvalues of $W_{\mathrm{in}}$4. In practice, the real parts $W_{\mathrm{in}}$5 are used as local contraction and expansion rates. The paper contrasts GFE with several frozen-coefficient proxies—LEJ, LEO, and LEY—but proves that for any fixed window their sums agree: $W_{\mathrm{in}}$6 As $W_{\mathrm{in}}$7, LEJ, LEY, and GFE converge to the same local limit. The authors stress that GFE is fundamentally local: overly long windows may misrepresent local behavior because a numerical trajectory approximates a nearby exact trajectory only over finite times.

The A-portrait is defined from the frozen Jacobian $W_{\mathrm{in}}$8 along sampled points $W_{\mathrm{in}}$9. For each real eigenvalue, a line segment is drawn at the sample point in the corresponding eigendirection, with length proportional to $$x_{k+1} = (1-a)x_k + a \cdot \tanh(\rho W x_k + \sigma W_{\mathrm{in}} u_k).$$0; blue denotes $$x_{k+1} = (1-a)x_k + a \cdot \tanh(\rho W x_k + \sigma W_{\mathrm{in}} u_k).$$1 and red denotes $$x_{k+1} = (1-a)x_k + a \cdot \tanh(\rho W x_k + \sigma W_{\mathrm{in}} u_k).$$2. For a complex-conjugate pair $$x_{k+1} = (1-a)x_k + a \cdot \tanh(\rho W x_k + \sigma W_{\mathrm{in}} u_k).$$3, two crossing segments are drawn in the eigenplane and colored by the sign of $$x_{k+1} = (1-a)x_k + a \cdot \tanh(\rho W x_k + \sigma W_{\mathrm{in}} u_k).$$4. The algorithm is direct: integrate a representative trajectory, sample it at fixed times, compute $$x_{k+1} = (1-a)x_k + a \cdot \tanh(\rho W x_k + \sigma W_{\mathrm{in}} u_k).$$5, eigendecompose it, and place the corresponding glyphs along the trajectory. The paper uses $$x_{k+1} = (1-a)x_k + a \cdot \tanh(\rho W x_k + \sigma W_{\mathrm{in}} u_k).$$6, $$x_{k+1} = (1-a)x_k + a \cdot \tanh(\rho W x_k + \sigma W_{\mathrm{in}} u_k).$$7 for the Lorenz attractor, $$x_{k+1} = (1-a)x_k + a \cdot \tanh(\rho W x_k + \sigma W_{\mathrm{in}} u_k).$$8, $$x_{k+1} = (1-a)x_k + a \cdot \tanh(\rho W x_k + \sigma W_{\mathrm{in}} u_k).$$9 for the Silnikov attractor, and long runs over $z_k = W_{\mathrm{out}}^\top x_k$0 for invariant tori and a limit cycle in the improved Nosé–Hoover oscillator.

These diagnostics expose hidden structures that ordinary trajectory plots may obscure. In the Van der Pol limit cycle, the A-portrait shows repulsive red segments near the $z_k = W_{\mathrm{out}}^\top x_k$1-axis regions of the cycle, indicating pronounced anisotropy in local transverse dynamics. In the Lorenz attractor, blue segments are predominantly normal to the local sheets of the attractor and red segments are tangent to them, with stronger repulsion in some wing regions, consistent with reinjection and folding. In the Silnikov system

$z_k = W_{\mathrm{out}}^\top x_k$2

the GFE and A-portrait become especially diagnostic near bifurcations. For a limit cycle of rotation number $z_k = W_{\mathrm{out}}^\top x_k$3 at $z_k = W_{\mathrm{out}}^\top x_k$4, $z_k = W_{\mathrm{out}}^\top x_k$5, the reported GFE is $z_k = W_{\mathrm{out}}^\top x_k$6, which matches the expected pattern of one near-zero and two negative exponents. Near splitting, at $z_k = W_{\mathrm{out}}^\top x_k$7, $z_k = W_{\mathrm{out}}^\top x_k$8, the GFE becomes $z_k = W_{\mathrm{out}}^\top x_k$9, displaying two near-zero values and strong red segments in the would-be separation direction. At $D$00, two symmetric limit cycles appear, with GFE $D$01. At $D$02, period-doubled cycles of rotation number $D$03 have GFE $D$04. For a high rotation number $D$05 orbit at $D$06, $D$07, the leading GFE is positive, $D$08, and the A-portrait shows pronounced red tangential segments.

The paper further introduces the smallest invariant closed set (SICS), defined as a nonempty invariant closed set with no nonempty proper invariant closed subset. For a bounded positive semi-orbit, the attractive set $D$09 is connected and is a SICS. Limit cycles, tori, and more complicated invariant closed sets may therefore be SICSs. In the improved Nosé–Hoover oscillator

$D$10

the paper reports three interlocked structures: two clusters of invariant tori and one dissipative limit cycle. Their A-portraits reveal complementary attractiveness and repulsiveness, clarifying how the sets coexist and constrain one another.

In this framework, sculpting consists in using A-portraits and windowed GFEs to identify leverage points for bifurcation control. Persistent red segments, near-zero leading GFEs, and localized weakening of attraction mark regions where parameter changes or feedback can split cycles, double rotation numbers, or regularize a chaotic set into a periodic orbit. The method is diagnostically rich but not free of ambiguity: window choice remains delicate, visual density can obscure interpretation, and extension beyond low-dimensional continuous-time systems is nontrivial.

4. Topological sculpting through templates and linking matrices

A third formulation is topological rather than metric. In "Visualizing the Template of a Chaotic Attractor" (Olszewski et al., 2018), the object of sculpting is the template of a chaotic attractor bounded by a genus–1 torus. The paper adopts the standard definition of a template as a compact branched two-manifold with boundary and smooth expansive semiflow built locally from joining and splitting charts. It explicitly restricts scope to attractors bounded by a genus–1 torus, such as Rössler or Malasoma attractors, and excludes more complex cases such as Lorenz attractors bounded by a genus–3 torus.

The primary encoding device is the linking matrix $D$11, an $D$12 symmetric integer matrix. Diagonal entries $D$13 specify torsions of branch $D$14, and off-diagonal entries $D$15 specify the number and sign of permutations between branches $D$16 and $D$17. A torsion is a twist of a branch with itself; a permutation is an exchange of position of two branches. The sign convention assigns positive or negative crossings according to the orientation used in the Melvin–Tufillaro representation, with strips ordered at the bottom from the back-most on the left to the front-most on the right.

The paper’s principal contribution is an end-to-end validator and renderer. Validation proceeds in three stages. First, the matrix must have correct form: square, symmetric, integer-valued. Second, it must satisfy continuity constraints. The diagonal elements must differ by exactly one from their diagonal neighbors; any matrix entry must differ from its neighbors by at most one; and an order array must be derivable by the Melvin–Tufillaro algorithm, containing each branch index exactly once. Repetition in this array implies tearing and invalidates a genus–1 template. Third, it must satisfy determinism constraints: certain $D$18 diagonal minors are forbidden up to addition of a global torsion, and the final bottom order must pass a planarity test in which alternating over/under arcs connecting consecutive labels can be drawn without intersections. Failure of this planarity test implies a choice point and hence non-determinism.

Once a matrix is validated, the next problem is compactness. The same linking matrix can generate multiple drawings, and the paper seeks the most concise template by maximizing the number of permutations executed per level of the drawing, equivalently minimizing height. Only currently adjacent branches may permute at a given level. The required off-diagonal permutations are collected into a multiset, and a breadth-first search tree is constructed whose nodes contain the remaining permutations and the current branch order. From any node, feasible next moves are the permutations that are both still required and currently adjacent. Breadth-first search ensures that the first valid leaf reached corresponds to the shortest path and hence to a minimal-height template.

The drawing stage converts the validated, optimally scheduled template into SVG. Torsions are drawn first from the diagonal of $D$19, then permutations are drawn level by level according to the shortest path through the search tree. The geometric primitive is a cubic Bézier curve. Given start point $D$20 and end point $D$21, the control points are placed halfway in height between the two points and straight above or below them. For a torsion, one Bézier curve is drawn, a small white circle erases the crossing center, and the second curve is drawn to produce the proper over/under configuration. For a permutation, the sign determines rendering order so that the upper strip covers the lower one.

The paper supplies explicit failure and success cases. A $D$22 matrix $D$23 is rejected because its final order is $D$24, which duplicates bottom positions. A matrix $D$25 passes form and continuity but fails the planarity test. A $D$26 matrix $D$27 passes all checks and yields valid final order $D$28, after which the tool draws the corresponding template. The method is practically effective but bounded by its assumptions. It does not apply to higher-genus templates, enforces determinism by design, and the breadth-first search can suffer combinatorial explosion; one $D$29 case with $D$30 permutations ran out of memory.

Within this literature, topological attractor sculpting therefore means selecting torsions and pairwise permutations through $D$31, verifying that the resulting object is a valid deterministic genus–1 template, extracting the minimal-height crossing order, and producing a compact topological rendering.

5. Scaling attractors in two-field inflation

In "Scaling attractors in multi-field inflation" (Christodoulidis et al., 2019), attractor sculpting is recast in a cosmological setting. The underlying system is a two-field model with curved scalar geometry $D$32 and potential $D$33,

$D$34

On a spatially flat FRW background,

$D$35

A scaling solution is defined by constant $D$36 and constant Hubble-normalized velocities $D$37. These are described as a one-parameter generalization of de Sitter solutions and as a natural starting point for the study of non-slow-roll slow-turn behavior.

The paper formulates scaling solutions as critical points of an autonomous system. With $D$38 and $D$39,

$D$40

For scaling solutions with $D$41,

$D$42

A broad class of two-field geometries with one transitive isometry is written as

$D$43

with $D$44. In projected variables $D$45 and $D$46, one has $D$47 and the autonomous system becomes

$D$48

The central classification comprises several families of critical points. In the shift-symmetric case $D$49, there are three. The geodesic or gradient solution is

$D$50

which exists for $D$51. Its stability eigenvalues in the invariant $D$52 subspace are

$D$53

and it is stable for

$D$54

For hyperbolic geometry $D$55, the hyperbolic spiral solution is

$D$56

with

$D$57

and it exists when $D$58. This solution is stable whenever it exists. The kinetic solution is $D$59 with $D$60.

For product potentials $D$61, the gradient valley attractor occurs at $D$62, with $D$63 and stability requiring $D$64 together with positive valley curvature $D$65. For potentials of the form $D$66, the paper identifies a frozen or orbital attractor with $D$67 and

$D$68

at a constant $D$69 determined by

$D$70

Its Hubble slow-roll parameter is

$D$71

For exponential $D$72 and $D$73, this frozen solution is always stable when it exists. The paper also studies an extremum attractor, geodesic along $D$74, and shows that negative curvature can destabilize it through geometrical destabilization.

Bifurcation theory is integral to the sculpting perspective. In hyperbolic geometry, the geodesic branch $D$75 is stable for $D$76 and becomes unstable at $D$77, where two stable non-geodesic branches $D$78 are born in a supercritical pitchfork. The selected branch has smaller $D$79. In the frozen-versus-extremum setting with $D$80 and $D$81, frozen solutions appear only for $D$82; at that threshold the extremum branch loses stability and two frozen branches at $D$83 emerge.

The paper emphasizes that background attractor stability is determined by the Jacobian eigenvalues of the autonomous system, not generically by the isocurvature effective mass $D$84. For frozen, isometry-aligned trajectories one has $D$85, so $D$86 is necessary and sufficient for background stability in that special case. In general curved manifolds, however, $D$87 is neither necessary nor sufficient. This corrected criterion is central to the engineering program.

Practical sculpting rules follow directly from the classification. Negatively curved manifolds with an isometry, such as $D$88 or $D$89, support robust sustained turning; $D$90 sets the curvature scale and hence the turn rate. Geodesic attractors require weak or positive curvature, $D$91, and a valley in the orthogonal direction. Spiral or frozen attractors require $D$92 or satisfaction of the frozen condition, respectively. Because

$D$93

for spiral and frozen classes, one can tune $D$94 through the combination $D$95, while the turn rate scales as $D$96. The paper further notes that in large-turn regimes one may have $D$97, so steep potentials can coexist with slow expansion, a point discussed in connection with swampland constraints.

6. Shared principles, diagnostics, and limitations

Taken together, these works suggest that attractor sculpting operates through three recurrent levers: modification of local stability, imposition of geometric or topological constraints, and controlled approach to bifurcation thresholds. In the reservoir-computing framework, the relevant threshold is the edge of chaos where the driven CLE crosses zero and closed-loop instability converts a supervised orbit into an untrained chaotic attractor (Kabayama et al., 2024). In the inflationary framework, sculpting proceeds by arranging critical-point existence and stability through curvature scale $D$98, potential gradients $D$99, and pitchfork bifurcations between geodesic and turning branches (Christodoulidis et al., 2019). In the GFE/A-portrait framework, leverage points are windows with near-zero or positive local exponents and regions of concentrated red segments (Guan, 2014). In the template framework, admissible sculpting is bounded by continuity, determinism, and genus–1 topology, with minimal-height realization obtained by breadth-first scheduling of adjacent permutations (Olszewski et al., 2018).

A second shared feature is the centrality of diagnostics that are more refined than global asymptotic stability. The semi-supervised LESN method uses open-loop $u_k$00, driven CLE, closed-loop MLE, the shape-alignment index $u_k$01, and the Kaplan–Yorke dimension. The inflationary classification relies on autonomous critical points, Jacobian eigenvalues, turn rate $u_k$02, and in special cases $u_k$03. The GFE/A-portrait program resolves local attraction and repulsion along the attractor rather than only a global Lyapunov average. The template approach uses structural validators—neighbor continuity, forbidden $u_k$04 blocks, planarity, and final-order feasibility—in place of metric stability criteria.

The limitations are correspondingly domain-specific. In reservoir computing, the sculpted attractor may depend sensitively on seed, $u_k$05, and fine scanning near $u_k$06; complex or noisy skeletons can degrade alignment. In scaling inflation, exact analytic control relies on exponential dependencies or slowly varying parameters, and background stability does not guarantee fluctuation stability. In GFE/A-portrait analysis, window size is not fully objective, dense visualizations can be difficult to interpret, and high-dimensional extensions are challenging. In template design, the genus–1 assumption excludes Lorenz-type attractors, determinism is enforced rather than inferred, and minimal-height search can become memory-intensive.

The combined literature therefore presents attractor sculpting not as a single method, but as a coherent research program: specify an intended invariant-set property, identify the structural variables that control it, and use local stability analysis or topological constraints to drive the system toward the desired attractor while remaining attentive to bifurcation structure and failure modes.