Killing Tensors and Carter Separability

Energy and axial angular momentum do not completely determine a Kerr geodesic: a particle can also move in latitude. The missing integral is quadratic in momentum and comes from a rank-two Killing tensor. Here we construct that integral, prove its independence from the familiar constants, and show how the same radial–angular split separates the scalar wave equation. The result applies locally to neutral geodesics and a minimally coupled scalar of constant mass on Kerr; it does not assert that arbitrary fields separate or that separated modes form a complete global basis.

Required background. Killing Fields and Isometry Groups supplies Killing equations and first integrals; Events, Intervals, and Minkowski Geometry supplies mass shell and four-momentum conventions.

Helpful background. Black-Hole Geometry and Mechanics supplies Kerr’s horizons; PDEs and Green Functions for Gravitation supplies separated boundary problems; Formulations of General Relativity distinguishes local separation from a global evolution problem; Frames, Coframes, and Anholonomy prepares the principal tensor’s geometric interpretation.

Killing tensors and momentum normalization

A symmetric tensor $K_{\alpha\beta}$ is a rank-two Killing tensor when

$$\nabla_{(\gamma}K_{\alpha\beta)}=0.$$

Use canonical momentum throughout. For a particle of rest mass $m_{\rm p}$, choose an affine parameter $\sigma$ so that

$$p^\alpha=\frac{dx^\alpha}{d\sigma}, \qquad H=\frac12g^{\alpha\beta}p_\alpha p_\beta =-\frac12m_{\rm p}^2.$$

For a massive particle, $p^\alpha=m_{\rm p}u^\alpha$ and $\tau_{\rm proper}=m_{\rm p}\sigma$, up to an additive constant. For a massless particle there is no proper-time normalization; an affine normalization fixes the scale of its momentum. In either case $p^\gamma\nabla_\gamma p^\alpha=0$, so

$$\begin{aligned} \mathcal K&=K_{\alpha\beta}p^\alpha p^\beta,\\ \frac{d\mathcal K}{d\sigma} &=p^\gamma p^\alpha p^\beta \nabla_{(\gamma}K_{\alpha\beta)} +2K_{\alpha\beta}p^\alpha p^\gamma\nabla_\gamma p^\beta =0. \end{aligned}$$

Thus the tensor equation proves conservation. It does not by itself prove that $\mathcal K$ is a new constant: the metric is a Killing tensor, but its contraction merely gives $2H$. Independence and separability need further checks.

A frequent normalization trap is to compare this $\mathcal K$ with a proper-velocity contraction. For $m_{\rm p}\ne0$,

$$K_{\alpha\beta}u^\alpha u^\beta =\frac{\mathcal K}{m_{\rm p}^2}.$$

The same distinction separates physical energy $E$ from specific energy $E/m_{\rm p}$. This momentum/affine normalization agrees with Frolov, Krtouš, and Kubizňák 2017, § 3.5, Eqs. (3.58)–(3.61), p. 54, PDF.

The Kerr split behind Carter separation

Assume Kerr with $\mu=GM>0$ and $a=J/M$, with $0<|a|\le\mu$. Work first in the Boyer–Lindquist exterior $r>r_+$, $0<\theta<\pi$, where

$$\begin{gathered} r_+=\mu+\sqrt{\mu^2-a^2},\\ \Delta=r^2-2\mu r+a^2, \qquad \Sigma=r^2+a^2\cos^2\theta. \end{gathered}$$

The limit $a\to0$ will be taken separately. The coordinates are singular at the horizon and axis, not the tensorial conservation law. The Kerr metric has the useful inverse-metric identity

$$\begin{aligned} \Sigma g^{\alpha\beta}p_\alpha p_\beta ={}&\Delta p_r^2+p_\theta^2\\ &-\frac{\big[(r^2+a^2)E-aL_z\big]^2}{\Delta} +\left(\frac{L_z}{\sin\theta}-aE\sin\theta\right)^2, \end{aligned}$$

where $E=-p_t$ and $L_z=p_\phi$ are constant because $t$ and $\phi$ are cyclic. Every term on the right depends on either $r,p_r$ or $\theta,p_\theta$, with $E,L_z$ as shared parameters. Crucially, $\Sigma=r^2+a^2\cos^2\theta$ is itself an additive radial–angular sum.

Separate the Hamilton–Jacobi equation

Let $W$ be the characteristic function, with $p_\alpha=\partial_\alpha W$. The mass-shell equation and separation ansatz are

$$\begin{gathered} g^{\alpha\beta}\partial_\alpha W\partial_\beta W=-m_{\rm p}^2,\\ W=-Et+L_z\phi+W_r(r)+W_\theta(\theta). \end{gathered}$$

There is no affine-time term in $W$. If the time-dependent principal function is wanted, it is $\overline S=W+\tfrac12m_{\rm p}^2\sigma$, satisfying $\partial_\sigma\overline S+H(x, \partial\overline S)=0$. The two Hamilton–Jacobi equations are distinguished in Frolov, Krtouš, and Kubizňák 2017, § 3.5, Eqs. (3.59)–(3.61), p. 54, PDF.

Set $P(r)=(r^2+a^2)E-aL_z$. Substitution into the inverse-metric identity gives

$$\begin{aligned} 0={}&\underbrace{\Delta(W_r')^2-\frac{P^2}{\Delta} +m_{\rm p}^2r^2}_{\text{radial part}}\\ &+\underbrace{(W_\theta')^2+ \left(\frac{L_z}{\sin\theta}-aE\sin\theta\right)^2 +m_{\rm p}^2a^2\cos^2\theta}_{\text{angular part}}. \end{aligned}$$

Since the two variables can vary independently, the angular part is a constant $\mathcal K$ and the radial part is $-\mathcal K$. Therefore

$$\begin{aligned} \Delta^2(W_r')^2 &=P^2-\Delta\big(m_{\rm p}^2r^2+\mathcal K\big),\\ (W_\theta')^2 &=\mathcal K- \left(\frac{L_z}{\sin\theta}-aE\sin\theta\right)^2 -m_{\rm p}^2a^2\cos^2\theta. \end{aligned}$$

This is an actual separation calculation, not an inference from stationarity and axisymmetry alone. The latter symmetries remove $t,\phi$ but generally leave a coupled two-variable problem. Compare Frolov, Krtouš, and Kubizňák 2017, § 3.5, Eqs. (3.61)–(3.66), pp. 54–55, PDF.

Two constants called “Carter”

A common alternative subtracts a combination of the manifest integrals:

$$\begin{aligned} Q&=\mathcal K-(L_z-aE)^2\\ &=p_\theta^2+\cos^2\theta \left[a^2(m_{\rm p}^2-E^2)+\frac{L_z^2}{\sin^2\theta}\right]. \end{aligned}$$

Both $Q$ and $\mathcal K$ are conserved. For an equatorial geodesic, $\theta=\pi/2$ and $p_\theta=0$, so $Q=0$ but $\mathcal K=(L_z-aE)^2$ generally does not vanish. Always identify which constant a radial equation uses.

Define the radial and polar potentials by

$$\begin{aligned} \mathcal V_r(r)&=P^2-\Delta\big(m_{\rm p}^2r^2+\mathcal K\big),\\ \Theta(\theta)&=Q-\cos^2\theta \left[a^2(m_{\rm p}^2-E^2)+\frac{L_z^2}{\sin^2\theta}\right]. \end{aligned}$$

Hamilton’s equations then give the first-order system

$$\begin{aligned} \Sigma\frac{dr}{d\sigma}&=\pm\sqrt{\mathcal V_r}, & \Sigma\frac{d\theta}{d\sigma}&=\pm\sqrt{\Theta},\\ \Sigma\frac{dt}{d\sigma} &=\frac{(r^2+a^2)P}{\Delta} +a(L_z-aE\sin^2\theta),\\ \Sigma\frac{d\phi}{d\sigma} &=\frac{aP}{\Delta}+\frac{L_z}{\sin^2\theta}-aE. \end{aligned}$$

Real motion requires $\mathcal V_r\ge0$ and $\Theta\ge0$. At an ordinary turning point the relevant square-root branch reverses; this is not a spacetime singularity. Horizon crossing instead requires a horizon-regular chart to remove the Boyer–Lindquist coordinate divergence.

An explicit tensor and a proof of integrability

The separated constant has a definition on the whole cotangent space, before any mass shell is fixed. Introduce

$$\begin{gathered} B=a^2\cos^2\theta,\\ \mathcal R=\Delta p_r^2-\frac{P^2}{\Delta}, \qquad \mathcal T=p_\theta^2+ \left(\frac{L_z}{\sin\theta}-aE\sin\theta\right)^2. \end{gathered}$$

Then $2\Sigma H=\mathcal R+\mathcal T$ and $\mathcal K=\mathcal T-2BH$. Equivalently, define the coordinate vector $v^\alpha=(\partial_\phi)^\alpha+a\sin^2\theta(\partial_t)^\alpha$. The symmetric tensor is

$$K^{\alpha\beta} =(\partial_\theta)^\alpha(\partial_\theta)^\beta +\frac{v^\alpha v^\beta}{\sin^2\theta} -Bg^{\alpha\beta}.$$

Its contraction is exactly $\mathcal K$, reducing on shell to $\mathcal T+B m_{\rm p}^2$. The vector $v$ in this expression is not claimed to be a Killing vector.

To prove the tensor equation without computing every connection component, use the canonical Poisson bracket

$$\{F,G\} =\sum_\alpha\left( \frac{\partial F}{\partial x^\alpha}\frac{\partial G}{\partial p_\alpha} -\frac{\partial F}{\partial p_\alpha}\frac{\partial G}{\partial x^\alpha} \right).$$

The separated functions satisfy $\{\mathcal R,\mathcal T\}=0$: their only shared variables are $p_t,p_\phi$, with no $t,\phi$ dependence. Since $\Sigma=r^2+B$, the product rule gives

$$\{H,\mathcal T\}=-\frac{H}{\Sigma}\{B,\mathcal T\}, \qquad \{H,B\}=-\frac{1}{2\Sigma}\{B,\mathcal T\}.$$

Hence $\{H,\mathcal K\}=\{H,\mathcal T\}-2H\{H,B\}=0$ for all momenta. For a quadratic momentum contraction, this identity is equivalent to the Killing-tensor equation: its cubic coefficient is the symmetrized covariant derivative of $K$. Thus the displayed tensor is genuinely Killing, rather than merely constant on one selected orbit.

All other pairwise brackets among $H,E,L_z,\mathcal K$ vanish by $t,\phi$ independence. Functional independence is also checkable:

$$\det\frac{\partial(H,E,L_z,\mathcal K)} {\partial(p_t,p_r,p_\theta,p_\phi)} =-\frac{2\Delta p_rp_\theta}{\Sigma}.$$

This minor is nonzero wherever $\Delta p_rp_\theta\ne0$, proving independence on a generic open set. Its vanishing at a turning point does not by itself prove that all other Jacobian minors vanish. We have four independent, commuting integrals for a four-coordinate Hamiltonian system: the conditions for local Liouville integrability. No compact invariant torus, global action–angle chart, or periodic orbit is being assumed.

What the principal tensor adds

Kerr’s principal tensor gives a geometric explanation of the split. It is a closed conformal Killing–Yano two-form $h$ satisfying

$$\nabla_\gamma h_{\alpha\beta} =g_{\gamma\alpha}\xi_\beta-g_{\gamma\beta}\xi_\alpha, \qquad \xi_\beta=\frac13\nabla^\alpha h_{\alpha\beta}, \qquad dh=0.$$

On a generic Kerr region its independent eigenvalue functions are $r$ and $a\cos\theta$. In four dimensions its Hodge dual on two-forms, $f=*h$, is Killing–Yano; the square $K_{\alpha\beta}=f_{\alpha\gamma}f_\beta{}^\gamma$, with the standard Kerr normalization, gives the tensor above. Reversing the orientation changes the sign of $f$, not its square. The explicit canonical forms and their square appear in Frolov, Krtouš, and Kubizňák 2017, § 3.2, Eqs. (3.20)–(3.26), p. 45, PDF.

Thus the latitude coordinate and fourth constant are not unrelated accidents. The more general construction belongs to Principal Killing-Yano Geometry; the explicit calculation above proves the implication needed here without assuming a classification theorem for other metrics.

For $a\ne0$, this tensor is irreducible: it cannot be a constant combination of the metric and products of Kerr’s two Killing vectors. Such products have no $rr$ or $\theta\theta$ components, whereas the ratios of our tensor to the inverse metric in those directions are respectively $-B$ and $r^2$.

In the Schwarzschild limit,

$$\mathcal K\longrightarrow p_\theta^2+\frac{L_z^2}{\sin^2\theta}=L^2, \qquad Q\longrightarrow L^2-L_z^2.$$

Now the principal tensor degenerates and the extra rotation generators make $K$ reducible, without destroying integrability (Frolov, Krtouš, and Kubizňák 2017, § 3.6, Eqs. (3.111)–(3.114), p. 63, PDF).

Scalar-wave separation and the Carter operator

Let $m_{\rm s}$ be the scalar-field mass, distinct from the test-particle mass $m_{\rm p}$, and consider $(\Box-m_{\rm s}^2)\Phi=0$. Using $\sqrt{-g}=\Sigma\sin\theta$, define the differential operators

$$\begin{aligned} \mathsf R &=\partial_r(\Delta\partial_r) -\frac{\big[(r^2+a^2)\partial_t+a\partial_\phi\big]^2}{\Delta},\\ \mathsf T &=\frac1{\sin\theta}\partial_\theta(\sin\theta\,\partial_\theta) +\frac{\big[\partial_\phi+a\sin^2\theta\,\partial_t\big]^2} {\sin^2\theta}. \end{aligned}$$

The coefficients inside the squared brackets are differentiated only in $t,\phi$. Direct substitution into the divergence formula for $\Box$ gives

$$\Sigma\Box=\mathsf R+\mathsf T.$$

Choose $\Phi=e^{-i\omega t}e^{im\phi}R(r)S(\theta)$, with integer $m$ for a single-valued scalar around the azimuth. Separation yields

$$\frac1{\sin\theta}\frac{d}{d\theta} \left(\sin\theta\frac{dS}{d\theta}\right) +\left[a^2(\omega^2-m_{\rm s}^2)\cos^2\theta -\frac{m^2}{\sin^2\theta}+A\right]S=0,$$

and

$$\frac{d}{dr}\left(\Delta\frac{dR}{dr}\right) +\left[ \frac{\big[(r^2+a^2)\omega-am\big]^2}{\Delta} -m_{\rm s}^2r^2-\lambda \right]R=0, \qquad \lambda=A+a^2\omega^2-2am\omega.$$

These are obtained by grouping the angular expression as $\mathsf T S-m_{\rm s}^2 B S=-\lambda S$, after its Fourier substitutions. The angular equation is the $s=0$ spheroidal equation, including for a massive scalar. Its effective parameter is

$$\gamma^2=a^2(\omega^2-m_{\rm s}^2), \qquad A=A_{0\ell m}(\gamma).$$

The scalar equation depends only on $\gamma^2$, so the choice of square-root sign does not change this angular problem. Only for $m_{\rm s}=0$ may we identify $\gamma$ with the site’s $c=a\omega$ and write $\lambda=A+c^2-2mc$. For nonzero mass, the radial shift still contains $a^2\omega^2$, not $\gamma^2$. Spin-Weighted Spheroidal Harmonics develops the general angular problem. This matches the convention in Angular Teukolsky Equation and Separation Constants. The massive scalar separation and commuting operators are given in Frolov, Krtouš, and Kubizňák 2017, § 3.5, Eqs. (3.71)–(3.76), pp. 56–57, PDF.

Why the operator commutes, and the eigenvalue sign

For the displayed tensor, its divergence-ordered scalar operator is

$$\mathcal C=\nabla_\alpha(K^{\alpha\beta}\nabla_\beta) =\mathsf T-B\Box.$$

The derivative of $\Sigma$ in the angular divergence cancels the first-derivative term from differentiating $B$ in $\nabla_\alpha(Bg^{\alpha\beta}\nabla_\beta)$. This explains why there is no omitted first-order correction in the second expression.

There is a short off-shell proof of commutation. The operators $\mathsf R,\mathsf T$ commute, as do $\mathsf R$ and multiplication by $B$. Using $\Sigma=r^2+B$ and the product rule for commutators gives

$$[\Box,\mathsf T]=\Sigma^{-1}[\mathsf T,B]\Box, \qquad [\Box,B]=\Sigma^{-1}[\mathsf T,B].$$

It follows that $[\Box,\mathcal C]=0$ on smooth scalars, before choosing a mode or imposing the wave equation. The same holds for a constant mass term. On the separated solutions,

$$\mathcal C\Phi=-\lambda\Phi.$$

The minus sign matters: the positive-divergence operator used here and on Symmetry Operators for Field Equations has eigenvalue $-\lambda$. Defining $\widehat K=-\mathcal C$ instead gives $+\lambda$, consistent with the principal-symbol replacement $p_\alpha\mapsto-i\nabla_\alpha$ in the classical quadratic constant.

For a generic Killing tensor, curvature can obstruct this scalar commutator; conservation of its classical contraction is insufficient. Carter’s general calculation identifies the extra conditions and their Ricci-flat specialization in Carter 1977, § VI, Eqs. (6.15)–(6.24), pp. 3409–3410. Here the explicit operator calculation proves the Kerr statement directly.

What separation does not guarantee

Regularity at both angular poles selects the eigenvalues $A_{\ell m}$, with $\ell\ge|m|$; at $a=0$, $A_{\ell m}=\ell(\ell+1)$. The angular problem is self-adjoint on the regular polar domain in the measure $\sin\theta\,d\theta$ whenever $\gamma^2$ is real. For real $a,m_{\rm s}$ this includes real and purely imaginary $\omega$; when $a=0$ the angular operator is independent of frequency. Generic complex $\gamma^2$ instead requires the non-self-adjoint continuation discussed on the spheroidal page. Horizon and infinity conditions must then be imposed on the radial equation. A local separation identity alone gives neither a quasinormal spectrum nor completeness of those modes.

The equation itself can also lose this separation while the metric remains exactly Kerr. For a smooth stationary-axisymmetric potential in $(\Box-V)\Phi=0$, the same product ansatz works when

$$\Sigma V(r,\theta)=F(r)+G(\theta).$$

On a connected coordinate rectangle this additive condition is equivalent to $\partial_r\partial_\theta(\Sigma V)=0$. A constant mass term satisfies it, since $\Sigma m_{\rm s}^2=m_{\rm s}^2r^2+m_{\rm s}^2B$. For a concrete counterexample, take dimensionless $\eta\ne0$ and

$$V=\frac{\eta r\cos\theta}{\mu\Sigma}, \qquad \partial_r\partial_\theta(\Sigma V) =-\frac{\eta}{\mu}\sin\theta\ne0.$$

This $V$ has dimensions of inverse length squared and is smooth in the stated exterior region. It prevents ordinary multiplicative radial–angular separation there, although the original geodesic Killing tensor is unchanged. This does not rule out every conceivable different coordinate system or field redefinition. Nor does a separable nonconstant potential necessarily preserve the unmodified $\mathcal C$: it can require a zeroth-order correction, as the next page’s commutator test explains.

The valid chain is therefore specific: Kerr geometry supplies an independent quadratic integral and a separable Hamiltonian; its scalar divergence operator also commutes with $\Box$; appropriate boundary conditions then define a spectral problem. Higher-spin lifts, deformed metrics, and global mode expansions require their own arguments.

Exercises

1. Specific constants and null rescaling. Rewrite the two separated geodesic potentials using specific quantities for $m_{\rm p}>0$. What happens instead under a constant rescaling of a null affine parameter?

Solution

Set $\mathcal E=E/m_{\rm p}$, $\mathcal L=L_z/m_{\rm p}$, $q=Q/m_{\rm p}^2$, and $k=\mathcal K/m_{\rm p}^2$. Dividing the momentum formulas by $m_{\rm p}^2$ replaces $(E,L_z,Q,\mathcal K,m_{\rm p}^2)$ by $(\mathcal E,\mathcal L,q,k,1)$, while replacing $d/d\sigma$ by $m_{\rm p}\,d/d\tau_{\rm proper}$ removes the remaining factor of $m_{\rm p}$ from the first-order equations. In particular $k=q+(\mathcal L-a\mathcal E)^2$.

For a null geodesic with $\sigma' = b\sigma$, $b>0$, momentum and $(E,L_z)$ scale by $b^{-1}$, while $(Q,\mathcal K)$ scale by $b^{-2}$. The trajectory is unchanged. When $E\ne0$, the ratios $L_z/E$ and $Q/E^2$ are independent of that affine normalization. There is no division by a nonzero rest mass in the null case.

2. Recover angular momentum from the wave operator. Set $a=0$. Determine $\mathcal C$, its eigenvalue on $Y_{\ell m}$, and the operator corresponding to the classical $Q=L^2-L_z^2$.

Solution

With $B=0$ and no $t$ term in $\mathsf T$,

$$\mathcal C=\frac1{\sin\theta}\partial_\theta (\sin\theta\,\partial_\theta) +\frac1{\sin^2\theta}\partial_\phi^2 =\Delta_{S^2}.$$

Therefore $\mathcal C Y_{\ell m}=-\ell(\ell+1)Y_{\ell m}$ and $-\mathcal C$ represents the positive angular-momentum square. The operator $-\mathcal C-(-i\partial_\phi)^2$ has eigenvalue $\ell(\ell+1)-m^2$, matching the classical distinction between $\mathcal K$ and $Q$. More generally on Kerr the wave analogue $-\mathcal C-(-i\partial_\phi-a\,i\partial_t)^2$ has eigenvalue $\lambda-(m-a\omega)^2=A-m^2$.

3. A compatible nonconstant potential. Let $F$ be a smooth radial function. Does $V=F(r)/\Sigma$ pass the radial–angular separation test? Does a generic potential $V=F(r)$ pass it when $a\ne0$?

Solution

The first choice gives $\Sigma V=F(r)$ and passes immediately: it alters only the radial equation after multiplication by $\Sigma$. For the second,

$$\partial_r\partial_\theta\big(\Sigma F(r)\big) =F'(r)\,\frac{dB}{d\theta}.$$

This is not identically zero for a generic nonconstant $F$ on rotating Kerr. Dependence of $V$ on only one coordinate is therefore not the right test: the weighted potential $\Sigma V$ must be additive. At $a=0$, $B=0$ and a radial potential is compatible with spherical separation, as expected.

References

• Carter, Brandon. “Killing Tensor Quantum Numbers and Conserved Currents in Curved Space.” Physical Review D 16 (1977): 3395–3414. DOI.
• Frolov, Valeri P., Pavel Krtouš, and David Kubizňák. “Black Holes, Hidden Symmetries, and Complete Integrability.” Living Reviews in Relativity 20 (2017): 6. DOI; Open PDF, arXiv version 3.

Further reading

• Carter, Brandon. “Hamilton–Jacobi and Schrödinger Separable Solutions of Einstein’s Equations.” Communications in Mathematical Physics 10 (1968): 280–310. DOI. The original investigation of a broader separable metric family.