Separation of Variables in Hamilton-Jacobi Theory

Hand the classical hydrogen atom to the Hamilton-Jacobi machinery in spherical coordinates and it falls open: three tidy one-dimensional integrals, energy and angular momentum appearing as integration constants. Now switch on a uniform electric field — the Stark problem — and the same spherical calculation jams completely. Rotate nothing, change no physics; simply re-describe the system in parabolic coordinates, and it falls open again. That strange sensitivity is the subject of this article. The Hamilton-Jacobi equation is a first-order nonlinear partial differential equation in $n$ coupled variables, and PDEs of that kind are generally hopeless. Separation of variables is the one great exception: an additive ansatz, $W=\sum_i W_i(q_i)$ , that — in the right coordinates — shatters the PDE into $n$ independent ordinary differential equations, each solvable by doing an integral. The engine behind the method is a single logical move used over and over. If the equation can be rearranged so that one coordinate and its derivative sit alone in a self-contained block, that block cannot communicate with the rest of the equation — so it must be frozen at a constant. Every such separation constant is a conserved quantity handed to you before you have solved anything. Mastering this technique buys you the practical half of integrability. The abstract Liouville-Arnold picture promises invariant tori when enough conserved quantities exist; separation of variables is the working machine that actually finds them, computes them, and — through Hamilton's characteristic function — converts them into explicit orbits, action variables, and frequencies.

Key takeaways

Separation of variables solves the time-independent Hamilton-Jacobi equation by positing that Hamilton's characteristic function splits additively, $W=\sum_i W_i(q_i)$ , each piece depending on one coordinate only.
Whenever one coordinate's dependence can be isolated into a self-contained block, that block must equal a constant — the separation constant — which is automatically a conserved quantity of the motion.
Each successful separation replaces one partial-derivative dependence by a constant and one ODE; full separation turns one PDE in $n$ variables into $n$ quadratures.
Separability belongs to the coordinate system, not to the system alone: the Kepler problem separates in spherical coordinates but not in Cartesian; the Stark problem separates in parabolic coordinates but not in spherical.
The $n$ separation constants are independent integrals in involution, so a fully separable system is Liouville integrable — separation is the practical route onto the invariant tori and to action-angle variables.
Key idea: if the characteristic function can be written $W=\sum_i W_i(q_i)$ and the Hamilton-Jacobi equation rearranges so that each coordinate appears only inside a self-contained block, those blocks must individually equal constants — the separation constants — turning one PDE into $n$ uncoupled quadratures.

The core insight

If the characteristic function can be written $W=\sum_i W_i(q_i)$ and the Hamilton-Jacobi equation rearranges so that each coordinate $q_i$ appears only inside a self-contained block, those blocks must individually equal constants — the separation constants — turning one PDE into $n$ uncoupled quadratures.

Where this lives today

The most spectacular modern separation is Carter's. Geodesic motion around a rotating (Kerr) black hole has only two obvious symmetries — time translation and axial rotation — one short of what integrability in four dimensions needs. In 1968 Brandon Carter tried the additive ansatz on the Hamilton-Jacobi equation in Boyer-Lindquist coordinates and it worked anyway, producing a fourth conserved quantity, the Carter constant, that corresponds to no spacetime symmetry at all (it descends from a hidden Killing tensor). Every photon trajectory in the ray-traced images of M87 and Sagittarius A produced by the Event Horizon Telescope collaboration — and every orbit model for the extreme mass-ratio inspirals that LISA is being built to hear — is computed by quadratures that exist because that separation succeeded. The same machinery runs quietly in less exotic places: satellite geodesy propagates orbits in the separable spheroidal approximations of Earth's field before adding perturbations; Rydberg-atom Stark spectroscopy still speaks the language of parabolic quantum numbers inherited from the 1916 separation; and in the theory of integrable systems, the search for "separation coordinates" for a given model remains an active research industry, with the Staeckel framework as its classical backbone.

The physical picture

Think of the Hamilton-Jacobi equation as a recipe in which all the coordinates are stirred together. Separation of variables asks a purely bookkeeping question: can the recipe be rearranged so that everything involving $q_1$ stands on one side of an equals sign and everything else on the other? If yes, the two sides cannot talk to each other — wiggle $q_1$ and only the left side could respond, yet the two sides are equal for every configuration — so each side is frozen at a constant. The equation has just handed you a conserved quantity without your solving anything. Repeat the argument coordinate by coordinate and the PDE collapses, one dimension at a time, into a cascade of one-dimensional problems.

Figure

The separation cascade — one coupled PDE shatters into n frozen blocks, each a quadrature

A cyclic coordinate is the trivial, warm-up case of this logic. If $q_1$ never appears in $H$ — only its momentum $p_1$ does — then the block is just $p_1$ itself: write $W_1=\alpha_1 q_1$ and the constant conjugate momentum $p_1=\alpha_1$ falls out immediately. Genuine separation generalizes the same move to coordinates that do appear, so long as coordinate and momentum appear bundled together in one isolable combination; that combination then plays the role of a generalized conserved momentum. The deepest — and most treacherous — feature of the method is that it is a property of the description, not just of the physics. The Kepler Hamiltonian written in Cartesian coordinates couples $x$ , $y$ , $z$ irreducibly through $V=-k/\sqrt{x^2+y^2+z^2}$ ; written in spherical coordinates, the same dynamics fall apart into three clean one-dimensional problems. The dynamics did not change. What changed is whether the coordinate surfaces are adapted to the symmetry of the problem.

Figure

Separability lives in the chart, not the system — the same Kepler orbit in two coordinate systems

Geometry explains what the constants are. Each separation constant pins the trajectory to one level set of a conserved function in phase space; with $n$ of them the orbit is confined to the intersection of $n$ level sets — an $n$ -dimensional torus, exactly the invariant torus promised by the Liouville-Arnold theorem. In the central-force problem you can see one of these constants in space: the constant $\alpha_\theta=L^2$ , together with $\alpha_\phi=L_z$ , confines the orbit to a band of latitudes. A satellite launched at inclination $i$ never strays poleward of latitude $\pm i$ — that visible band is a separation constant made flesh.

Figure

A separation constant you can see — the polar-angle block confines a satellite to a latitude band

In practice the cascade has a canonical order: peel off time first (giving the energy), then every cyclic coordinate (each giving its conserved momentum for free), and attack the genuinely coupled coordinates last. Each step strips one variable and leaves a smaller problem of exactly the same type.

The separation machinery

The starting point is the Time-Independent Hamilton-Jacobi Equation for Hamilton's characteristic function — both covered in their own articles, so here we state it and put it to work. It asks for a function $W(q_1,\dots,q_n)$ whose gradient, inserted in place of the momenta, holds the Hamiltonian at a constant value:

H\!\left(q_1,\dots,q_n,\frac{\partial W}{\partial q_1},\dots,\frac{\partial W}{\partial q_n}\right)=E \qquad (1)

where: $W$ — Hamilton's characteristic function, the generator of a canonical transformation to constant momenta; $\partial W/\partial q_i$ — identified with the momentum $p_i$ ; $E$ — the conserved energy, itself the first separation constant. Equation (1) is one equation in $n$ unknown partial derivatives — wildly underdetermined as it stands, which is exactly why an ansatz is needed. Time has already been peeled off to get here. For any Hamiltonian without explicit time dependence, the full principal function splits as $S(q,t)=W(q)-Et$ : substituting into $H+\partial S/\partial t=0$ , the coordinate side and the time side cannot communicate, so both equal the constant $E$ — the separation argument in its very first appearance. (The full story of that split belongs to the Time-Independent Hamilton-Jacobi Equation article; here it is move zero of the cascade.) Now the ansatz that gives the method its name. Posit that the characteristic function is an additive sum of single-variable functions:

W(q_1,\dots,q_n)=\sum_{i=1}^{n}W_i(q_i) \qquad (2)

where: each $W_i$ depends only on its own coordinate $q_i$ (and, parametrically, on the separation constants). Additive — not multiplicative as in the separation of linear PDEs — because what enters the Hamilton-Jacobi equation is the gradient of $W$ : an additive split makes $\partial W/\partial q_i=\mathrm{d}W_i/\mathrm{d}q_i$ depend on one variable only, which is the whole point. When the ansatz succeeds, the equation rearranges coordinate by coordinate into self-contained blocks, each frozen at its own constant:

H_k\!\left(q_k,\frac{\mathrm{d}W_k}{\mathrm{d}q_k}\right)=\alpha_k \qquad (3)

where: $H_k$ — the group of terms containing only $q_k$ and $\mathrm{d}W_k/\mathrm{d}q_k$ ; $\alpha_k$ — the $k$ -th separation constant. Each block is now an ordinary differential equation for one function of one variable. Solve it algebraically for the momentum and integrate:

\frac{\mathrm{d}W_i}{\mathrm{d}q_i}=p_i(q_i;\alpha_1,\dots,\alpha_n) \qquad\Longrightarrow\qquad W_i=\int^{q_i} p_i(q_i';\alpha)\,\mathrm{d}q_i' \qquad (4)

where: $p_i$ — the momentum as an explicit function of its own coordinate and the constants; $q_i'$ — a dummy integration variable. This is what "solved by quadrature" means: every remaining difficulty is a definite integral, an object a computer (or a 19th-century analyst with a table of elliptic integrals) evaluates without ever touching a differential equation again. One step remains: the constants $\alpha_i$ tell you which orbit, but not where the system is on it. The trajectory itself comes from differentiating $W$ with respect to the constants:

\beta_i=\frac{\partial W}{\partial \alpha_i}\ \ (i\neq 1),\qquad t+\beta_1=\frac{\partial W}{\partial E} \qquad (5)

where: $\beta_i$ — new constant coordinates conjugate to the constant momenta $\alpha_i$ . These are the standard type-2 transformation equations (the machinery of Generating Functions and the Four Types, applied with $W$ as $F_2$ ): the new Hamiltonian depends only on constants, so the new coordinates are frozen too — except the one conjugate to the energy, which ticks linearly as $t$ . Inverting the algebraic relations (5) yields $q_i(t)$ explicitly. Derivation C below turns this crank for the central-force orbit and watches Kepler's ellipse emerge.

Which equation answers which question

you want	use	notes
the PDE to be solved	eq. (1)	one equation, $n$ unknown derivatives
to attempt separation	ansatz (2), blocks (3)	success depends on the coordinate system
the momenta along the orbit	eq. (4), $p_i(q_i;\alpha)$	each momentum a function of its own coordinate
the explicit trajectory	eq. (5), $\beta=\partial W/\partial\alpha$	differentiate first, integrate after — often simpler
a separability criterion	Staeckel conditions	see Staeckel Conditions for Separability
actions and frequencies	$J_i=\frac{1}{2\pi}\oint p_i\,\mathrm{d}q_i$	see Action Integrals and Invariant Tori

Derivation AWhy every separated block is a constant of the motion

We prove the claim that powers the whole method: whenever the Hamiltonian depends on a coordinate-momentum pair $(q_1,p_1)$ only through a single combination $f_1(q_1,p_1)$ , that combination is an exact constant of the motion, and the separation ansatz automatically finds it. The strategy has two halves — first the freeze argument, which extracts the constant from the structure of the equation alone, then a one-line Poisson-bracket computation that confirms the frozen quantity is genuinely conserved along the dynamical flow. Watch the logic of the freeze step: it is not an approximation or a special solution, but a statement about functions of disjoint variable sets being equal identically.

Step 1. Assume the Hamiltonian has the isolable structure $H=h\big(f_1(q_1,p_1),\,q_2,\dots,q_n,p_2,\dots,p_n\big)$ — the pair $(q_1,p_1)$ enters only through the single function $f_1$ . Why this is the right hypothesis: it is precisely the structure the separation ansatz exploits; "the $q_1$ -dependence can be collected into one block" is this statement in words.

Step 2. Substitute the split $W=W_1(q_1)+W'(q_2,\dots,q_n)$ into the Hamilton-Jacobi equation:

h\!\left(f_1\!\big(q_1,\tfrac{\mathrm{d}W_1}{\mathrm{d}q_1}\big),\,q_2,\dots,\tfrac{\partial W'}{\partial q_2},\dots\right)=E

Why the block closes: because $W_1$ depends only on $q_1$ , the argument $f_1$ is a function of $q_1$ alone — all other dependence lives in the remaining slots of $h$ .

Step 3. Freeze the block. Vary $q_1$ while holding the other coordinates fixed: on the left side only the $f_1$ slot can respond, yet the right side is the fixed constant $E$ . If $h$ depends monotonically (in particular, nontrivially) on its first slot, the only way the equation can hold identically is

f_1\!\left(q_1,\frac{\mathrm{d}W_1}{\mathrm{d}q_1}\right)=\alpha_1=\text{const}

Why this is not circular: no solution of the dynamics has been used — the argument is purely about the functional structure of one identity in independent variables.

Step 4. Confirm conservation along the flow. Compute the Poisson bracket of $f_1$ with the Hamiltonian, using the chain rule on $H=h(f_1,\dots)$ : the only term involving derivatives with respect to $q_1$ or $p_1$ is proportional to $\{f_1,f_1\}$ ,

\frac{\mathrm{d}f_1}{\mathrm{d}t}=\{f_1,H\}=\frac{\partial h}{\partial f_1}\,\{f_1,f_1\}=0

Why the other terms vanish: $f_1$ depends only on $(q_1,p_1)$ , and the remaining arguments of $h$ contain neither, so their brackets with $f_1$ are identically zero.

Step 5. Iterate. With $f_1=\alpha_1$ frozen, the residual equation is a Hamilton-Jacobi problem in one fewer variable with $\alpha_1$ as a parameter; each further separation repeats Steps 1–4. Full separation therefore delivers $n$ functionally independent constants, and since each $f_i$ involves a disjoint coordinate-momentum pair (given the constants already fixed), they are in involution: $\{f_i,f_j\}=0$ . Verified: the $n$ constants in involution are exactly the hypothesis of the Liouville-Arnold theorem — separation implies Liouville integrability; matches Goldstein §10.4 and Arnold Ch. 10.

Go deeper · Derivation B · 2 minSeparating the central-force problem in spherical coordinates

Derivation BSeparating the central-force problem in spherical coordinates

This is the showcase of the entire method: a particle of mass $m$ in an arbitrary central potential $V(r)$ , fully separated in spherical coordinates $(r,\theta,\phi)$ , yielding the energy, the total angular momentum squared, and its polar projection as the three separation constants. The strategy follows the canonical order — cyclic coordinate first (that is $\phi$ ), then the polar block, then the radius — and each peel produces one constant and one quadrature. Watch how the constants nest: $\alpha_\phi$ appears inside the $\theta$ -block, and $\alpha_\theta$ inside the radial block — the cascade hands each level down to the next as frozen data.

Step 1. Write the Hamiltonian in spherical coordinates:

H=\frac{1}{2m}\left[p_r^2+\frac{p_\theta^2}{r^2}+\frac{p_\phi^2}{r^2\sin^2\theta}\right]+V(r)

Why these momenta: they are conjugate to $(r,\theta,\phi)$ from the spherical kinetic energy $\tfrac{m}{2}(\dot r^2+r^2\dot\theta^2+r^2\sin^2\theta\,\dot\phi^2)$ ; note $\phi$ appears nowhere — it is manifestly cyclic.

Step 2. Convert to the Hamilton-Jacobi equation by the replacement $p_i\to\partial W/\partial q_i$ :

\frac{1}{2m}\left[\left(\frac{\partial W}{\partial r}\right)^2+\frac{1}{r^2}\left(\frac{\partial W}{\partial\theta}\right)^2+\frac{1}{r^2\sin^2\theta}\left(\frac{\partial W}{\partial\phi}\right)^2\right]+V(r)=E

Step 3. Peel the cyclic coordinate. With the ansatz $W=W_r(r)+W_\theta(\theta)+W_\phi(\phi)$ , the coordinate $\phi$ appears only through $\mathrm{d}W_\phi/\mathrm{d}\phi$ , so that derivative must be constant:

\frac{\mathrm{d}W_\phi}{\mathrm{d}\phi}=\alpha_\phi \qquad\Longrightarrow\qquad W_\phi=\alpha_\phi\,\phi,\qquad \alpha_\phi=p_\phi=L_z

Why $L_z$ : the momentum conjugate to the azimuth is the polar component of angular momentum — the cyclic coordinate delivers its conserved momentum with zero work.

Step 4. Isolate the polar block. Multiply the equation through by $2mr^2$ and collect everything that depends on $\theta$ :

\left(\frac{\mathrm{d}W_\theta}{\mathrm{d}\theta}\right)^2+\frac{\alpha_\phi^2}{\sin^2\theta}=\alpha_\theta

Why the block freezes: the left side depends only on $\theta$ , while the rest of the equation depends only on $r$ ; equal functions of disjoint variables are constant. The constant is $\alpha_\theta=L^2$ , the squared total angular momentum: the left side is exactly $p_\theta^2+p_\phi^2/\sin^2\theta$ , which is $|\mathbf{L}|^2$ written in spherical components.

Step 5. What remains is purely radial:

\frac{1}{2m}\left(\frac{\mathrm{d}W_r}{\mathrm{d}r}\right)^2+\frac{\alpha_\theta}{2mr^2}+V(r)=E

where: the term $\alpha_\theta/2mr^2$ is the familiar centrifugal barrier — the angular cascade has deposited itself into an effective one-dimensional radial potential.

Step 6. Read off the three quadratures:

W_r=\int\!\sqrt{2m\big(E-V(r)\big)-\frac{\alpha_\theta}{r^2}}\;\mathrm{d}r,\qquad W_\theta=\int\!\sqrt{\alpha_\theta-\frac{\alpha_\phi^2}{\sin^2\theta}}\;\mathrm{d}\theta,\qquad W_\phi=\alpha_\phi\,\phi

Why this is complete: the PDE in three variables has become two definite integrals and one linear function; the three constants $(E,\,\alpha_\theta=L^2,\,\alpha_\phi=L_z)$ are the promised integrals in involution. *Verified: reproduces the standard result $\alpha_\phi=L_z$ , $\alpha_\theta=L^2$ , $\alpha_t=E$ ; matches Goldstein §10.7. The turning points of the $\theta$ -quadrature, $\sin\theta_{\min}=|L_z|/L$ , are the latitude band drawn in the third figure.*

Go deeper · Derivation C · 1 minFrom separation constants to the orbit — Kepler's ellipse by differentiation

Derivation CFrom separation constants to the orbit — Kepler's ellipse by differentiation

Separation delivers $W$ ; this derivation shows how the orbit falls out of it, using the planar Kepler problem ( $\theta=\pi/2$ , potential $V=-k/r$ , angular momentum $p_\phi=L$ ). The tool is equation (5): the coordinate $\beta$ conjugate to the constant $L$ is itself constant, and writing out $\beta=\partial W/\partial L$ produces a relation between $r$ and $\phi$ with no time in it — the geometric orbit equation. The practical trick to absorb: differentiate with respect to the constant before integrating, because $\partial W/\partial\alpha$ is usually a far simpler integral than $W$ itself. The strategy: write the planar $W$ , differentiate under the integral sign, substitute $u=1/r$ , and recognize an arccosine.

Step 1. In the orbital plane the characteristic function is $W=L\phi+W_r(r)$ with

W_r=\int^{r}\sqrt{2m\Big(E+\frac{k}{r'}\Big)-\frac{L^2}{r'^2}}\;\mathrm{d}r'

Why: this is Derivation B restricted to $\theta=\pi/2$ , where $\alpha_\theta=L^2$ and $p_\phi=L$ .

Step 2. Apply equation (5) with $\alpha=L$ , differentiating under the integral sign:

\beta=\frac{\partial W}{\partial L}=\phi-\int^{r}\frac{L\,\mathrm{d}r'}{r'^2\sqrt{2m\big(E+\tfrac{k}{r'}\big)-\tfrac{L^2}{r'^2}}}

Why differentiating helps: the integrand of $W_r$ contains a square root; its $L$ -derivative is algebraic in the same square root — one power tamer, and exactly integrable.

Step 3. Substitute $u=1/r'$ , $\mathrm{d}r'=-\mathrm{d}u/u^2$ . The $r'^{-2}$ measure cancels and the radicand becomes a downward parabola in $u$ :

\phi-\beta=\int^{u}\frac{L\,\mathrm{d}u'}{\sqrt{2mE+2mku'-L^2u'^2}}

Step 4. Complete the square in the radicand, $2mE+2mku-L^2u^2=\frac{m^2k^2}{L^2}\Big(1+\frac{2EL^2}{mk^2}\Big)-L^2\Big(u-\frac{mk}{L^2}\Big)^2$ , and the integral is a standard arccosine. Inverting:

u=\frac{1}{r}=\frac{mk}{L^2}\Big(1+e\cos(\phi-\phi_0)\Big),\qquad e=\sqrt{1+\frac{2EL^2}{mk^2}}

where: $\phi_0$ — the perihelion direction, absorbed from $\beta$ ; $e$ — the eccentricity. This is the conic-section orbit equation: an ellipse for $E<0$ , parabola for $E=0$ , hyperbola for $E>0$ . Why the result carries no time: $\beta$ is conjugate to $L$ , not to the energy — pairing with the energy constant, $t+\beta_1=\partial W/\partial E$ , is the separate relation that supplies the schedule $r(t)$ (Kepler's equation lives there). *Verified: matches the orbit equation derived by the effective-potential route in Goldstein §3.5, with the same eccentricity formula; circular orbit check: $e=0$ requires $E=-mk^2/2L^2$ , the minimum of the effective potential, as it must.*

Limiting cases

Case	Result	Meaning
cyclic coordinate	$W_k=\alpha_k q_k$ , $p_k=\alpha_k$	separation is automatic; the block is the momentum itself
all coordinates cyclic	$W=\sum_k\alpha_k q_k$	the trivial-Hamiltonian limit — dynamics already solved
one degree of freedom	$W=\int\!\sqrt{2m(E-V)}\,\mathrm{d}q$	energy conservation is the separation; every 1D system separates
Staeckel-separable	$p_i^2$ linear in the constants $\alpha_j$	the general orthogonal separable structure — see Staeckel Conditions for Separability
non-separable in the chosen chart	no additive $W$ exists in that chart	proves nothing by itself: try adapted coordinates — or the system may be genuinely non-integrable

Common trapIt did not separate in my coordinates, so the system is not integrable

This inference is wrong in both directions, and it costs real research time. Failing to separate in one chart proves nothing: the Kepler problem does not separate in Cartesian coordinates, the Stark problem does not separate in spherical coordinates — both are exactly separable in charts adapted to their symmetry (spherical and parabolic respectively). Conversely, integrability does not guarantee that any separating chart exists: the Toda lattice is Liouville integrable, yet its integrals do not come from coordinate separation. Separability is a sufficient — never a necessary — certificate of integrability, and it is a joint property of the dynamics and the coordinate system. When an additive ansatz jams, the honest conclusions are only: this chart is not adapted, and a systematic check (the Staeckel conditions, for orthogonal charts) is the next move.

Worked examples

1 · Anisotropic oscillator — separation in Cartesian coordinates (confidence builder)

A particle of mass $m=0.5\ \mathrm{kg}$ moves in the potential $V=\tfrac12 k_x x^2+\tfrac12 k_y y^2$ with $k_x=8\ \mathrm{N/m}$ and $k_y=18\ \mathrm{N/m}$ . It is released with amplitudes $A_x=0.10\ \mathrm{m}$ , $A_y=0.05\ \mathrm{m}$ . Separate the Hamilton-Jacobi equation, evaluate all separation constants numerically, and determine whether the orbit closes.

Step 1. The Hamilton-Jacobi equation is $\frac{1}{2m}\big[(\partial W/\partial x)^2+(\partial W/\partial y)^2\big]+\tfrac12 k_x x^2+\tfrac12 k_y y^2=E$ . Because the potential is already a sum of single-variable terms, Cartesian coordinates are the adapted chart: the ansatz $W=W_x(x)+W_y(y)$ splits the equation into $\frac{1}{2m}W_x'^2+\tfrac12 k_x x^2=\alpha_x$ and $\frac{1}{2m}W_y'^2+\tfrac12 k_y y^2=E-\alpha_x\equiv\alpha_y$ .

Step 2. Each block is a one-dimensional oscillator energy. Evaluate at the turning points: $\alpha_x=\tfrac12 k_x A_x^2=\tfrac12\times 8\times(0.10)^2=\mathbf{0.040\ J}$ and $\alpha_y=\tfrac12 k_y A_y^2=\tfrac12\times 18\times(0.05)^2=\mathbf{0.0225\ J}$ .

Step 3. The total energy is the sum of the frozen blocks: $E=\alpha_x+\alpha_y=0.040+0.0225=\mathbf{0.0625\ J}$ — the first separation constant is recovered as the sum, a consistency check on the cascade.

Step 4. The two quadratures give the familiar oscillations with $\omega_x=\sqrt{k_x/m}=\sqrt{8/0.5}=\mathbf{4.0\ rad/s}$ and $\omega_y=\sqrt{18/0.5}=\mathbf{6.0\ rad/s}$ .

Step 5. Closure test: $\omega_x/\omega_y=4/6=2/3$ is rational, so the Lissajous orbit closes, with common period $T$ set by the largest $T$ with $\omega_x T$ and $\omega_y T$ both multiples of $2\pi$ : $T=2\pi/2=\mathbf{3.14\ s}$ (two full $x$ -cycles and three full $y$ -cycles). An irrational ratio would instead carpet the rectangle $|x|\le A_x$ , $|y|\le A_y$ densely — the two-torus picture developed in Action Integrals and Invariant Tori.

2 · A satellite's latitude band from the polar block (exam level)

A satellite of mass $m=1000\ \mathrm{kg}$ is in a circular orbit of radius $r=7000\ \mathrm{km}$ around Earth ( $\mu=GM_\oplus=3.986\times10^{14}\ \mathrm{m^3/s^2}$ ) at inclination $i=60°$ . Evaluate the three spherical separation constants and find the band of latitudes the ground track can reach.

Step 1. Circular speed: $v=\sqrt{\mu/r}=\sqrt{3.986\times10^{14}/7.0\times10^6}=\sqrt{5.694\times10^7}=\mathbf{7546\ m/s}$ .

Step 2. Energy constant: $E=-\frac{\mu m}{2r}=-\frac{3.986\times10^{14}\times1000}{1.4\times10^{7}}=\mathbf{-2.85\times10^{10}\ J}$ — negative, as it must be for a bound orbit.

Step 3. Total angular momentum constant: $L=mrv=1000\times7.0\times10^6\times7546=\mathbf{5.28\times10^{13}\ kg\,m^2/s}$ , so $\alpha_\theta=L^2=2.79\times10^{27}\ \mathrm{kg^2m^4/s^2}$ .

Step 4. Polar projection: the orbit plane is tilted at $i=60°$ to the equator, so $\alpha_\phi=L_z=L\cos i=5.28\times10^{13}\times0.5=\mathbf{2.64\times10^{13}\ kg\,m^2/s}$ .

Step 5. The $\theta$ -block controls the reachable polar angles: $p_\theta^2=\alpha_\theta-\alpha_\phi^2/\sin^2\theta\ge 0$ requires $\sin\theta\ge|L_z|/L=\cos i=0.5$ , i.e. $\theta$ between $30°$ and $150°$ .

Step 6. Convert to latitude ( $\mathrm{lat}=90°-\theta$ ): the ground track oscillates between latitudes $\mathbf{-60°}$ and $\mathbf{+60°}$ — exactly the inclination, recovered here not from orbital geometry but from the positivity of a separated momentum block. At the turning latitudes $p_\theta=0$ : the track runs momentarily due east there, which is why ground tracks visibly "kiss" their extreme latitudes.

The mathematician behind the method

Origins · 1 minCarl Gustav Jacob Jacobi and how this was discovered

Carl Gustav Jacob Jacobi

1804–1851 · Königsberg and Berlin

Jacobi took Hamilton's optical-mechanical formalism — which Hamilton himself used mostly as an exact restatement of dynamics — and turned it into the most powerful solution engine classical mechanics has ever had. In his 1837 papers and the celebrated Königsberg lectures (published posthumously as the Vorlesungen über Dynamik), he recast the theory around a single partial differential equation and showed that solving mechanics means separating that equation. He then demonstrated the method's reach with a virtuoso piece: geodesic motion on a triaxial ellipsoid, separated in the elliptic coordinates he invented for the purpose — a problem with no obvious symmetry and no other known route. Characteristically sharp-tongued, he chided colleagues for treating mathematics as a servant of the senses; for him the separation method was mathematics leading physics somewhere neither could go alone.

History

1834

— Hamilton publishes the optical-mechanical theory of the principal and characteristic functions, the raw material of the method.

1837

— Jacobi reworks the theory into the Hamilton-Jacobi equation and introduces separation of variables as its systematic solution technique.

1839

— Jacobi separates geodesic flow on the triaxial ellipsoid in elliptic coordinates — the method's first conquest of a problem with no evident symmetry.

1891

— Paul Stäckel gives the complete algebraic criterion for separability of orthogonal coordinate systems, closing the question Jacobi opened.

1916

— Epstein and Schwarzschild quantize the Stark effect by separating hydrogen's Hamilton-Jacobi equation in parabolic coordinates — old quantum theory's finest hour.

1968

— Brandon Carter separates the Hamilton-Jacobi equation in the Kerr black-hole spacetime, revealing a hidden fourth constant with no symmetry to explain it.

Going further

References. Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., Ch. 10 (Sections 10.4-10.7) · Landau & Lifshitz, Mechanics, 3rd ed., Sec. 48 (Separation of variables) · Morin, Introduction to Classical Mechanics, Hamilton-Jacobi chapter · Taylor, Classical Mechanics, Ch. 13 · Arnold, Mathematical Methods of Classical Mechanics, Ch. 10

Provenance. Equations from the standard Hamilton-Jacobi formalism. A verifier should recompute: (1) the time separation $S=W-Et$ from $H+\partial S/\partial t=0$ (Goldstein Sec. 10.4, Landau Sec. 47); (2) the full spherical Kepler separation giving $\alpha_\phi=L_z$ , $\alpha_\theta=L^2$ , $\alpha_t=E$ and the three quadratures (Goldstein Sec. 10.7); (3) the Poisson-bracket proof that each separation block is conserved; (4) the Staeckel matrix form (Goldstein Appendix D / Sec. 10.6). Kepler energy-action relation $E=-2\pi^2 m k^2/(\sum J_i)^2$ from Goldstein Sec. 10.8. All cross-checked against Landau-Lifshitz Vol 1 Secs. 47-48 and Arnold Ch. 10. Worked-example numbers on this page ( $\alpha_x=0.040$ J, $\alpha_y=0.0225$ J, $\omega_x=4.0$ rad/s, $\omega_y=6.0$ rad/s, $v=7546$ m/s, $E=-2.85\times10^{10}$ J, $L=5.28\times10^{13}$ kg m²/s, latitude band $\pm60°$ ) are illustrative arithmetic recomputed from the displayed formulas.

Builds on: Hamilton-Jacobi Equation · Hamilton's Characteristic Function · Time-Independent Hamilton-Jacobi Equation · Hamilton's Principal Function Leads to: Staeckel Conditions for Separability · Action-Angle Variables · Liouville Integrable Systems · Frequencies from Action-Angle Variables

Why is the separation additive here, when PDE courses always separate multiplicatively?

Because the Hamilton-Jacobi equation constrains the gradient of $W$ , not $W$ linearly. An additive split $W=\sum_i W_i$ makes each derivative $\partial W/\partial q_i=W_i'(q_i)$ a function of one variable — which is what lets blocks isolate. Multiplicative separation is the natural move for linear PDEs like Laplace's or Schrödinger's equation, where products pass through the linear operator and divide out. The two are secretly the same trick: write $\psi=e^{iW/\hbar}$ and a multiplicative split of $\psi$ becomes an additive split of $W$ . That correspondence is not a coincidence — it is the seed of the connection to quantum mechanics explored later in the track (Connection to the Schrödinger Equation).

Is there an algorithm for finding the coordinates that separate a given system, or is it guesswork?

For orthogonal coordinate systems the answer is completely algorithmic in principle: the Staeckel conditions (next article) give a necessary and sufficient test for whether a chart separates a natural Hamiltonian, and in Euclidean 3-space the candidate orthogonal charts are classified — there are exactly eleven families (Cartesian, spherical, parabolic, elliptic, and their kin). So the honest workflow is: read the symmetry of the potential, shortlist adapted charts, test with Staeckel. What remains an art (and an active research area under the name "separation coordinates") is the general case — non-orthogonal charts, momentum-dependent potentials, and deciding whether any separating chart exists for a given integrable system. Failure of every chart you tried is evidence, never proof.

Do separation constants always mean something physical, like angular momentum?

No — and expecting them to is a common source of confusion. The constants are whatever the freeze argument isolates, and their physical transparency depends on the chart. In spherical coordinates the Kepler constants happen to be gloriously physical ( $E$ , $L^2$ , $L_z$ ). Separate the very same problem in parabolic coordinates and the constants are $E$ , $L_z$ , and a component of the Laplace-Runge-Lenz vector — still conserved, far less obvious. For Jacobi's geodesics on the ellipsoid, or the two-center problem, the separation constant is a quadratic invariant with no everyday name at all. The guarantee is only ever: functionally independent, in involution, conserved. Meaning is a bonus.

If a system separates in two different coordinate systems, do both give the same constants?

They give the same span of dynamics but generally different constants — different involutive sets whose members are functions of each other's. Hydrogen/Kepler is the canonical example: spherical separation yields $\{E,L^2,L_z\}$ , parabolic separation yields $\{E,L_z,A_z\}$ with $A_z$ a Laplace-Runge-Lenz component; $L^2$ is not a member of the second set. Multiple inequivalent separations are in fact a symptom of degeneracy: they can only coexist when the frequencies of the motion are commensurate, so that the invariant tori are non-uniquely fibered. That is why hydrogen — maximally degenerate — separates in four distinct chart families, while a generic central potential manages only spherical.

What happens to all this when I add a small perturbation, like Earth's equatorial bulge?

Strict separability is destroyed by almost any perturbation — the $J_2$ oblateness term couples $r$ and $\theta$ in a way no constant can freeze exactly. What survives is more interesting: the system becomes near-integrable. The unperturbed separation still provides the natural variables (actions and angles) in which perturbation theory is organized, secular drifts (like the famous nodal regression of satellite orbits) are computed, and the KAM theorem decides which invariant tori survive the perturbation. So a broken separation is not a dead end; it is the zeroth order of the modern theory of near-integrable dynamics. (A curiosity: the main $J_2$ problem admits an approximate Staeckel fit in oblate spheroidal coordinates, which is exactly how high-precision satellite theories start.)

Do I ever need W itself, or only its derivatives?

Almost always only the derivatives — and realizing this saves painful integrals. The orbit comes from $\beta_i=\partial W/\partial\alpha_i$ and the momenta from $p_i=\partial W/\partial q_i$ ; both can be formed by differentiating under the integral sign before integrating, which typically demotes a nasty radical to a manageable one (Derivation C lives on this trick). The actions $J_i$ likewise need only $p_i(q_i;\alpha)$ , never $W$ . The characteristic function is scaffolding: essential to the logic — it is the generating function that certifies everything is canonical — but rarely evaluated in closed form, and never needed numerically.

The freeze argument varies one coordinate while fixing the others — but along a real trajectory they all change together. Is the argument legitimate?

Yes, and this worry is worth settling precisely. The Hamilton-Jacobi equation is an identity in the coordinates: it holds at every point $q$ of the configuration region, not merely along a trajectory. When we "vary $q_1$ holding the rest fixed," we are comparing the identity at neighboring points of configuration space — a move about functions, not about motion. No claim is made that the system moves that way. The dynamical statement (the frozen block is conserved along the actual flow) is then verified separately, by the Poisson-bracket computation in Derivation A, Step 4. Structure first, dynamics second — the argument is airtight because the two are kept apart.

How does separability relate to quantum-mechanical separability of the Schrödinger equation?

Closely but not identically. Where classical Hamilton-Jacobi separation works, the corresponding Schrödinger equation usually separates in the same coordinates (hydrogen: spherical and parabolic — the quantum Stark effect is computed in parabolic quantum numbers to this day), and the classical separation constants become commuting operators. But the correspondence acquires corrections: the quantum equation cares about the ordering of operators and about curvature-like terms of order $\hbar^2$ that have no classical counterpart, so "quantum separability" can fail where classical separability holds (and, rarely, vice versa). The clean statement of the relationship goes through writing $\psi\sim e^{iS/\hbar}$ — the topic of Connection to the Schrödinger Equation, later in this track.

Separation of Variables in Hamilton-Jacobi Theory

Where this lives today

The physical picture

The separation machinery

Which equation answers which question

Limiting cases

Worked examples

1 · Anisotropic oscillator — separation in Cartesian coordinates (confidence builder)

2 · A satellite's latitude band from the polar block (exam level)

The mathematician behind the method

Going further

Before this

Where this leads