---
title: Keplerian Image Transformations
url: https://www.emergentmind.com/topics/keplerian-image-transformations
type: topic
---

# Keplerian Image Transformations

“Keplerian image transformations” is best treated as an *Editor’s term* for transformation families that recast objects governed by central projection or Keplerian dynamics into alternative geometric representations while preserving a physically significant structure. In the supplied literature, the phrase spans two technical domains. In calibrated perspective imaging, the closest rigorous counterpart is the class of rigidity preserving image transformations, for which the only depth-independent image-plane warps induced by rigid motion are rotational homographies [2201.13065]. In celestial mechanics, related constructions map Keplerian trajectories, orbital states, or relative-motion solutions into circles, oscillator variables, geodesics, or alternative coordinate charts through polar reciprocation, regularizing lifts, and state/element reparameterizations [1107.0852], [1901.06876], [1612.07395], [2605.12982], [2404.18141], [2108.09608].

## 1. Scope and unifying idea

The literature does not present a single formal theory under the name “Keplerian image transformations.” Instead, it presents several precise transformation classes, each attached to a different invariant. In the perspective-imaging setting, the invariant is compatibility with rigid camera motion under unknown depth. In the Kepler problem, the invariant may be conic structure, canonicity, regularity at collision, preservation of the fixed-energy trajectories as geodesics, or the correct transformation of probability densities under reparameterization. This suggests a unifying characterization: a Keplerian image transformation is not an arbitrary deformation, but a map selected because it preserves the physically relevant structure of a Keplerian or perspective system.

Under that reading, the main families in the cited material are sharply differentiated. Polar reciprocation is an involutive geometric map on central-force trajectories that sends Kepler conics to circles or circular arcs [1107.0852]. The Jacobi–Maupertuis and Jacobi–Eisenhart frameworks absorb the Kepler potential into a metric and reparametrized flow, so that fixed-energy dynamics become geodesic motion [1612.07395]. The Kustaanheimo–Stiefel and Lissajous–Kustaanheimo–Stiefel constructions lift the 3D Kepler problem to a constrained 4D oscillator and then re-express that oscillator in action-angle variables adapted to perturbation theory [1901.06876]. State-vector/orbital-element maps treat the Kepler problem as a reparameterization problem with explicit Jacobian determinants, which is essential for Bayesian inference [2605.12982]. In close-proximity dynamics, Lyapunov–Floquet transport between QNS, Cartesian, and spherical coordinates provides analytically linked geometric “images” of the same linearized Keplerian relative motion [2108.09608].

## 2. Perspective-consistent image transformations

In the calibrated pinhole-camera model used in “Rigidity Preserving Image Transformations and Equivariance in Perspective,” a visible 3D point
\[
x=\begin{bmatrix}x_0\\x_1\\x_2\end{bmatrix}, \qquad x_2>0
\]
projects to
\[
\pi(x)= \begin{bmatrix} x_0/x_2\\ x_1/x_2 \end{bmatrix}\in \mathbb{R}^2.
\]
Given a set \(B\subset \mathbb{P}^2\) containing at least two points, an image-plane map \(\eta\) is rigidity preserving on \(B\) if there exists a rigid motion \(\rho\) of \(\mathbb{R}^3\) such that
\[
\eta\circ \pi(x)=\pi\circ \rho(x), \qquad \text{for all } x\in \pi^{-1}(B).
\]
The quantifier over \(\pi^{-1}(B)\) is the decisive condition: the transformation must hold for all 3D points lying on the viewing rays through \(B\), so unknown scene depth is built into the definition. The main theorem is exact:
\[
\eta \text{ is rigidity preserving} \iff \eta \in RH,
\]
where \(RH\) is the set of rotational homographies, namely homographies represented in normalized image coordinates by matrices \(R\in SO(3)\). Thus the only depth-independent image-plane transformations induced by rigid motion for all possible 3D content are the warps generated by pure camera rotations [2201.13065].

This result excludes several transformations that are often treated as generic image symmetries. A 2D translation by \(\tau=(\tau_0,\tau_1)^T\) is represented in homogeneous coordinates by
\[
H= \begin{bmatrix} I_{2\times2} & \tau\\ 0 & 1 \end{bmatrix},
\]
which is generally not orthogonal and hence not in \(SO(3)\). More strongly, if a rigid motion has nonzero translation,
\[
\rho(x)=Rx+t, \qquad t\neq 0,
\]
then its image effect is depth dependent: points at different depths move by different amounts in the image, so no single global image warp exists under unknown depth. The paper therefore treats ordinary CNN-style translation equivariance as geometrically mismatched for 3D inference from perspective images. In pixel coordinates, a physically valid augmentation by pure camera rotation has homography
\[
H = K R_{\mathrm{aug}} K^{-1},
\]
and the associated camera pose update is \(R_{\mathrm{aug}}R\) with unchanged camera center \(c\). On the representation side, the paper shows that \(RH\cong SO(3)\), and develops a pitch-yaw approximation \(PY\) in which a radial remapping
\[
\tan(r)\mapsto r
\]
—essentially the azimuthal equidistant projection—allows ordinary convolutions on the resampled grid to approximate equivariance to pitch-yaw camera rotations. Experimentally, both the \(PY\)-equivariant representation and rotational-homography augmentation improve competitive baselines on LINEMOD and Occlusion LINEMOD, while localization results are more mixed [2201.13065].

## 3. Polar reciprocation of Kepler orbits

In Davis’s treatment of central-force motion, the geometric image transformation is polar reciprocation. For a trajectory \(\tau\) in a central force field centered at \(O\), with velocity \(\vec \nu=\dot{\vec r}\) and angular momentum per unit mass
\[
\vec C = \vec r \times \dot{\vec r}, \qquad C=|\vec C|,
\]
the reciprocal point \(P_*\) of \(P\in\tau\) lies on the line through \(O\) perpendicular to \(\vec \nu\), with magnitude \(OP_*=r_\perp^{-1}\), where \(r_\perp\) is the perpendicular distance from \(O\) to the tangent line at \(P\). Analytically,
\[
\overrightarrow{OP_*} = \frac{\vec \nu \times \vec C}{C^2},
\qquad
\vec r_*(t)=\dot{\vec r}(t)\times \frac{\vec C}{C^2}.
\]
The transformed curve is therefore the hodograph, rescaled by \(C^{-1}\) and rotated clockwise by \(90^\circ\), but unlike the hodograph the construction is involutive:
\[
(P_*)_* = P,
\qquad
\vec r_{**}=\vec r.
\]
That self-inverse property is the construction’s main structural distinction [1107.0852].

For the Kepler problem,
\[
\ddot{\vec r} = -\gamma \frac{\vec r}{r^3},
\]
the reciprocal curve has especially simple form. Defining
\[
\rho \equiv \frac{\gamma}{C^2},
\]
the paper derives
\[
\overrightarrow{OP_*} = \rho(\vec\varepsilon_*+\hat e_r).
\]
Since \(\rho \hat e_r\) traces a circle of radius \(\rho\) centered at \(O\), the polar reciprocal \(k_*\) of a Kepler orbit \(k\) is a circle if the entire angular range is realized, or a circular arc for open trajectories. Applying the same transformation again recovers the original orbit and yields
\[
\frac{1}{r} = \overrightarrow{OP_*}\cdot \hat e_r
          = \rho(\vec\varepsilon_*\cdot \hat e_r +1).
\]
Choosing the polar axis along \(\vec\varepsilon_*\) gives
\[
\frac{1}{r}=\rho(1+e\cos\phi),
\qquad
e=|\vec\varepsilon_*|,
\]
which is the standard polar equation of a conic with focus at \(O\). The reciprocal-circle description therefore turns the conic nature of Kepler orbits into a direct consequence of a circle plus an involution. The same framework also yields a uniform ruler-and-compass construction for elliptical, parabolic, and hyperbolic trajectories, and the vector \(\vec\varepsilon_*\) is identified, up to scale, with the Laplace–Runge–Lenz vector [1107.0852].

## 4. Geodesic, canonical, and regularizing transforms

A second major family of Keplerian image transformations replaces the original dynamics by geodesic or oscillator dynamics. In the Jacobi–Maupertuis framework, a natural Lagrangian
\[
L(x,\dot x)=\frac12 g_{ij}(x)\dot x^i\dot x^j-U(x)
\]
at fixed energy \(E\) is recast using the Jacobi metric
\[
ds_J^2 = 2m(E-U(x))\,g_{ij}(x)\,dx^i dx^j,
\]
with Maupertuis time change
\[
d\sigma = (E-U(x))\,dt.
\]
For the Kepler potential \(U(r)=-\alpha/r\), the planar metric becomes
\[
ds^2=\left(E+\frac{\alpha}{r}\right)(dr^2+r^2d\theta^2),
\]
and its Gaussian curvature is
\[
K_G=-\frac{\alpha E}{2(rE+\alpha)^3}.
\]
The sign of \(E\) then matches the sign of curvature and the conic class:
\[
E<0 \Rightarrow K_G>0 \quad \text{(ellipse)}, \qquad
E=0 \Rightarrow K_G=0 \quad \text{(parabola)}, \qquad
E>0 \Rightarrow K_G<0 \quad \text{(hyperbola)}.
\]
Within the same survey, Bohlin’s transformation
\[
x+iy=\frac{(q_1+i q_2)^2}{2}
\]
maps the planar Hooke oscillator into the Kepler Hamiltonian, while Houri’s canonical swap
\[
x^i=p_i,\qquad p_i=x^i
\]
and Milnor’s momentum inversion
\[
w=\frac{p}{|p|^2}
\]
produce constant-curvature metric forms and anomaly-adapted regularized representations. The contact-geometric formulation further interprets Kepler flow on a fixed energy hypersurface as a Reeb flow after time reparametrization [1612.07395].

The Kustaanheimo–Stiefel regularization and its Lissajous–KS refinement push this strategy further. With defining vector \(\boldsymbol c\), KS lifts the 3D physical position \(\mathbf{x}\) to a quaternion \(\mathbf{v}\) by
\[
\alpha \mathbf{x} = \mathbf{v}\,\mathbf{c}\,\overline{\mathbf{v}},
\qquad
r=\|\mathbf{x}\|=\frac{|\mathbf{v}|^2}{\alpha},
\]
and introduces Sundman time
\[
\frac{d\tau}{dt}=\frac{\alpha}{4r}.
\]
The transformed Hamiltonian is that of a perturbed isotropic oscillator in four variables, supplemented by the bilinear invariant
\[
\mathcal{J}(\mathbf{v},\mathbf{V})= \left(\overline{\mathbf{v}}\wedge\overline{\mathbf{V}}\right)\cdot\mathbf{c}=0,
\]
which encodes the extra KS fibre degree of freedom. The Lissajous–Kustaanheimo–Stiefel transformation applies Lissajous variables separately on the \((v_0,v_3)\) and \((v_1,v_2)\) oscillator planes, then recombines them so that one action becomes
\[
\Gamma=\mathcal{J}(\mathbf{v},\mathbf{V}),
\]
with conjugate angle \(\gamma\) cyclic. The physical reduction is therefore the simple constraint \(\Gamma=0\), taking the regularized \(4.5\) degrees-of-freedom system back to \(3.5\) degrees of freedom. Because LKS avoids any reference to the orbital plane, it remains regular for most circular inclined orbits, equatorial elliptic orbits, and all radial orbits except polar radial ones, including degenerate rectilinear ellipses that are problematic for Delaunay-like or plane-based action-angle variables [1901.06876].

## 5. Orbital-state reparameterizations and computational transforms

In astrodynamics and Bayesian orbit inference, the relevant transformation is often the map between Cartesian state vectors and Keplerian elements. The default bound-orbit parameterization in “On the Reparameterization Between Cartesian Position-Velocity Vectors and Orbital Elements in the Kepler Problem” is
\[
(\bm{x},\bm{v})=(x,y,z,v_x,v_y,v_z)
\quad \leftrightarrow \quad
(a,e,i,\Omega,\omega,M),
\]
with standard textbook forward relations through Kepler’s equation, the eccentric anomaly \(E\), and the perifocal-frame vectors
\[
\bm{r}_{\rm pf}= \begin{pmatrix} a(\cos E-e)\\ a\sqrt{1-e^2}\sin E\\ 0 \end{pmatrix},
\qquad
\bm{v}_{\rm pf} = \frac{\sqrt{\mu a}}{r}
\begin{pmatrix} -\sin E\\ \sqrt{1-e^2}\cos E\\ 0 \end{pmatrix},
\]
followed by
\[
\bm{x} = R_z(\Omega)\,R_x(i)\,R_z(\omega)\,\bm{r}_{\rm pf},
\qquad
\bm{v} = R_z(\Omega)\,R_x(i)\,R_z(\omega)\,\bm{v}_{\rm pf}.
\]
Its principal analytic result is the closed-form Jacobian determinant
\[
\det\left(\frac{\partial(\bm{x},\bm{v})}{\partial(a,e,i,\Omega,\omega,M)}\right)
= \frac{1}{2}\,\mu^{3/2}\, a^{1/2}\, e\, \sin i.
\]
This makes the singularity structure explicit: the map is singular at \(e=0\), where periapsis and \(\omega\) are undefined, and at \(i=0,\pi\), where the ascending node and \(\Omega\) are undefined. The same paper gives reduced Jacobian factorizations, alternative parameter choices such as \(P\) or \(\tau\), and the density-conservation rule
\[
p(\bm{x}, \bm{v})\,d\bm{x}\,d\bm{v} = p(\boldsymbol{\theta})\,d\boldsymbol{\theta},
\]
so that priors transform by multiplication or division by the absolute Jacobian. In astrometric orbit fitting, matching priors in this way yields identical posteriors across parameterizations, while sampling in \((\bm{x},\bm{v})\) improves effective sample size per core-minute by a factor of \(2\)–\(7\) and reduces divergent transitions relative to orbital-element sampling [2605.12982].

A separate numerical literature addresses the implementation of the inverse map from a Cartesian spacecraft state vector
\[
\mathbf{x}=\{\mathbf{r},\mathbf{v}\}
\]
to classical orbital elements
\[
\mathbf{o}=\{a,e,i,\Omega,\omega,\nu\}.
\]
“Computing Classical Orbital Elements with Improved Efficiency and Accuracy” recasts this conversion as a floating-point software problem. The paper shows that standard branch-heavy, \(\arccos\)-based implementations suffer from domain failures, tolerance-dependent special cases, loss of precision at small inclination from
\[
i=\arccos(h_z/h),
\]
and cancellation error at small eccentricity in
\[
\mathbf{e}=\frac{1}{\mu}(\mathbf{v}\times\mathbf{h})-\hat{\mathbf{r}}.
\]
Its branchless ATAN2-based algorithm AL3 instead uses
\[
i=\operatorname{ATAN2}\!\left(\sqrt{h_x^2+h_y^2},\,h_z\right), \qquad
\Omega=\operatorname{ATAN2}(h_x,-h_y),
\]
constructs an in-plane basis \((\hat{\mathbf{n}},\mathbf{b},\hat{\mathbf{h}})\), and evaluates
\[
\omega=\operatorname{ATAN2}(\mathbf{e}\cdot\mathbf{b},\mathbf{e}\cdot\hat{\mathbf{n}}),
\qquad
\ell=\operatorname{ATAN2}(\mathbf{r}\cdot\mathbf{b},\mathbf{r}\cdot\hat{\mathbf{n}}),
\qquad
\nu=\ell-\omega.
\]
The method handles arbitrary eccentricity and inclination except the degenerate zero-angular-momentum case. On large random benchmarks it improves RMS state-vector reconstruction error by more than two orders of magnitude relative to a corrected conventional implementation and reduces runtime by about \(43\%\) on the reported hardware; on some low-inclination tests the conventional method is up to \(5\) orders of magnitude worse [2404.18141].

## 6. Representation-to-representation transforms in Keplerian relative motion

In the linearized chief–deputy problem about a closed Keplerian chief orbit, the relevant image transformation is an analytic transport between coordinate representations of the same periodic linear dynamics. The paper on fundamental solutions in relative motion uses three states: local Cartesian coordinates
\[
\bm{x}_c = (x,y,z,\dot x,\dot y,\dot z)^\top,
\]
differential quasi-nonsingular elements
\[
\doe = (\delta a,\delta\theta,\delta i,\delta q_1,\delta q_2,\delta\Omega)^\top,
\]
and local spherical coordinates
\[
\bm{x}_s = (\delta r,\theta_r,\phi_r,\delta\dot r,\dot\theta_r,\dot\phi_r)^\top.
\]
If two representations are related by an orbit-periodic linear map
\[
\bm x = [G(t)]\doe,
\]
then their state-transition matrices satisfy
\[
[\Phi_{\bm x}(t,t_0)] = [G(t)] [\Phi_{\doe}(t,t_0)] [G(t_0)]^{-1},
\]
and their Lyapunov–Floquet reductions are related by
\[
[P_{\bm x}(t)] = [G(t)][P_{\doe}(t)][G(t_0)]^{-1},
\qquad
[\Lambda_{\bm x}] = [G(t_0)][\Lambda_{\doe}][G(t_0)]^{-1}.
\]
This makes the coordinate-image relation exact: once an LF transformation is known in one representation, its counterpart in any linearly related periodic representation follows analytically [2108.09608].

For the Keplerian case, the QNS reduction yields an almost trivial LTI system with only one Jordan-chain coupling, representing semimajor-axis-driven along-track phase drift. Transporting that normal form into Cartesian and spherical coordinates produces modal decompositions with four in-plane modes, two purely out-of-plane modes, and one drift mode. The reported geometric shapes include an offset circle mode, a teardrop-shaped mode, and, in spherical coordinates, a one-dimensional along-track oscillatory mode. The spherical transformed dynamics further admit a stationary plane in the active in-plane subspace, characterized by
\[
\bm\rho\cdot\bm n=0,
\]
with the vector field parallel to that plane and speed proportional to the distance from it. A plausible implication is that, in this literature, “image transformation” often means not a raster warp but an analytically exact change of geometric picture: the same Keplerian relative trajectory becomes easier or harder to interpret depending on whether it is rendered in QNS, Cartesian, or spherical coordinates [2108.09608].

Across these domains, the common principle is restrictive rather than permissive. Ordinary image translations are not generally compatible with rigid perspective motion [2201.13065]. Polar reciprocals are not merely hodographs, because they are self-inverse [1107.0852]. Classical orbital elements are not globally regular coordinates, because their singularities at \(e=0\) and \(i=0,\pi\) are geometric degeneracies rather than numerical accidents [2605.12982]. LKS does not eliminate all singularities, but relocates them away from the highly eccentric and rectilinear regimes for which regularization is most useful [1901.06876]. In that sense, Keplerian image transformations are best understood as structure-preserving recodings: each transformation narrows the class of admissible maps in order to expose a conserved geometry, a regularized dynamics, or a measure-correct statistical representation.

Source: https://www.emergentmind.com/topics/keplerian-image-transformations