Hyperbolicity and the Cauchy Problem

A covariant action does not by itself define a predictive initial-value problem. One must choose variables and gauge, reduce or propagate the constraints, and test the principal symbol of that formulation on the background and branch of interest. Real characteristic speeds are necessary but not sufficient: strong hyperbolicity requires a complete, uniformly controlled characteristic basis. Under explicit regularity and boundary hypotheses, that structure yields the energy estimate needed for local existence, uniqueness, and continuous dependence on initial data.

Required background. Variational and Symplectic Geometry supplies the constraint and gauge subspaces, while Ordinary Differential Equations supplies the Fourier-mode and energy-estimate logic used in the proof.

Helpful background. Einstein Equations provide the harmonic-gauge baseline, Gravitational Actions and Boundary Terms fix the underlying variational problem, and Formulations of General Relativity distinguish evolution equations, gauge conditions, and constraint propagation.

A strong-hyperbolicity theorem

Consider a quasilinear first-order reduction on a three-dimensional Cauchy slice without boundary,

$$\partial_tu=A^i(u,x)\partial_i u+F(u,x),$$

with initial data $u(0)=u_0$. Fix an open set of field values and a time direction for which the slice is noncharacteristic. For every real unit spatial covector $n_i$, define the principal symbol $A(n)=A^in_i$.

Theorem. Suppose the coefficients are sufficiently smooth, $u_0\in H^s$ with $s\gt5/2$, and there exists a Hermitian symmetrizer $H(n;u,x)$, smooth in $n$ and the fields, such that

$$cI\le H\le CI, \qquad H A(n)=A(n)^\dagger H$$

for constants $0\lt c\lt C$ uniform over the stated region. Suppose also that the gauge-fixed evolution preserves the constraint surface. Then the Cauchy problem is locally well posed in the corresponding Sobolev class: for nearby initial data satisfying the constraints there is a common short time on which solutions exist uniquely and depend continuously on those data.

The corresponding frozen-coefficient algebra requires real eigenvalues and a complete eigenvector matrix whose condition number is uniformly bounded for all $n$. The boundedness clause is essential near coincident characteristics. For the variable-coefficient theorem one also needs the smoothly varying symmetrizer stated above, not just a separate diagonalization at each sampled point. For systems second order in space or time, the theorem applies only after a specified first-order or pseudodifferential reduction whose symbol has the same uniform property. The symmetrizer definition and the local well-posedness implication are reviewed in Kovacs and Reall 2020, Appendix A.1, equations (119)–(124), PDF.

The statement concerns a local Cauchy problem on a slice without a physical boundary. With timelike or artificial boundaries, compatible maximally dissipative or constraint-preserving boundary conditions and an additional boundary-stability estimate are required. The theorem also does not cover coefficients below the regularity needed for the nonlinear estimates.

Proof architecture: the uniform energy estimate

Freeze the coefficients at one point and Fourier transform in space. With wavevector $k_i=|k|n_i$,

$$\partial_t\widehat u=i|k|A(n)\widehat u+\widehat F.$$

Define the frequency-space energy $E=\widehat u^\dagger H\widehat u$. Because $HA(n)$ is Hermitian, the two terms proportional to $|k|$ cancel in $dE/dt$. Derivatives of $H$, coefficient variation, and lower-order terms are bounded by a constant times $E$ plus the source norm. Uniform bounds on $H$ make this energy equivalent to $|\widehat u|^2$, independently of frequency.

Weighting by $(1+|k|^2)^s$, integrating over $k$, and controlling commutators gives an estimate of the form

$$\lVert u(t)\rVert_{H^s} \le C_T\left( \lVert u(0)\rVert_{H^s} +\int_0^t\lVert F(\tau)\rVert_{H^s}\,d\tau \right).$$

Gronwall’s inequality supplies continuous dependence and uniqueness. A standard iteration of the linearized problem gives local existence while the solution remains in the coefficient region where the symmetrizer is bounded. Finally, a homogeneous subsidiary evolution system for the constraints shows that initially vanishing constraints remain zero. This is the proof architecture, not a claim that every gauge choice supplies the required symmetrizer.

The Jordan-block counterexample

Weak hyperbolicity—real characteristic eigenvalues without uniform diagonalizability—is insufficient. Consider

$$\partial_tu=A\partial_xu, \qquad A=\begin{pmatrix}1&1\\0&1\end{pmatrix}.$$

The only characteristic speed is real, $1$, but $A=I+N$ has a nonzero nilpotent Jordan block $N$. A Fourier mode evolves as

$$\widehat u(t,k)=e^{iktA}\widehat u(0,k) =e^{ikt}\left(I+iktN\right)\widehat u(0,k).$$

For data with a component along the generalized eigenvector, the norm grows as $|k|t$. No estimate with a frequency-independent constant can bound the solution by initial data in the same Sobolev norm. Arbitrarily short wavelengths therefore destroy continuous dependence even though every characteristic speed is real. This explicit counterexample is the discriminating reason to demand strong rather than merely weak hyperbolicity.

What the quadratic action can and cannot show

The physical modes on the preceding pages used

$$S_{\mathrm{red}}^{(2)}=\frac12\int dt\,d^3x\,a^3 \left[ \dot q^{\mathsf T}K\dot q -\frac{1}{a^2}\partial_iq^{\mathsf T}G\partial_iq -q^{\mathsf T}M^2q +2\dot q^{\mathsf T}Bq \right].$$

On a frozen background, positive $K$ and a complete positive generalized spectrum of $Gv=c_s^2Kv$ are a valuable check on the physical high-frequency sector. They do not prove strong hyperbolicity of the full nonlinear equations. The lapse, shift, gauge drivers, constraint additions, and first-order reduction introduce principal-symbol blocks absent from the reduced action. Conversely, a poor gauge formulation can be weakly hyperbolic even when the physical quadratic modes are healthy.

Every conclusion must therefore name five ingredients: the covariant action, the background, the gauge-fixed formulation, the solution branch, and the weak-coupling or EFT regime. Changing any one can change the principal symbol without changing the others.

Formulation-dependent results in modified gravity

Papallo and Reall analyze principal symbols in generalized harmonic gauges. On generic weak-field backgrounds, they find weak but not strong hyperbolicity for Lovelock theory in harmonic gauge and an obstruction for general Horndeski theories within the generalized-harmonic class they study; special subclasses such as k-essence-like theories behave better. The hypotheses and linear-algebra mechanism are set out in Papallo and Reall 2017, § 2 and § 3.4. This is a result about specified gauges and generic weak-field backgrounds, not a theorem that the covariant theories admit no other well-posed formulation.

Kovacs and Reall subsequently construct modified harmonic gauges for which weakly coupled Lovelock and Horndeski equations are strongly hyperbolic. Their matrix criterion, weak-coupling continuity argument, and gauge construction are developed in Kovacs and Reall 2020, Appendix A and §§ 4.1–4.2. “Weak coupling” means that higher-derivative corrections to the principal symbol are controlled relative to the Einstein part on the chosen background; it is not a statement about small perturbation amplitude alone.

The two results are compatible. They teach that hyperbolicity is a property of a formulation in a regime. A successful weak-coupling formulation does not establish strong hyperbolicity on every strong-field branch, while failure of a conventional generalized harmonic gauge does not establish a formulation-independent impossibility theorem.

This page settles no empirical bound. Observational likelihoods belong to Tests of Gravity, stable low-energy quantum-gravity EFT belongs to Quantum Gravity, and generic PDE theory extends beyond this bounded gravitational application.

Common pitfalls

Equating real characteristics with well-posedness. The Jordan example has only a real characteristic speed but lacks a bounded eigenbasis and a same-norm energy estimate.

Suppressing the gauge in a theorem statement. Adding constraints or changing gauge drivers changes the principal symbol. Quote the formulation and gauge together with the theory.

Extending weak-coupling results to another branch. A symmetrizer controlled perturbatively near an Einstein background need not remain bounded where higher-derivative terms compete with the Einstein principal part.

Exercises

For the Jordan system above, take $\widehat u(0,k)=(0,1)^{\mathsf T}$. Compute the solution and explain why an $H^s\to H^s$ estimate cannot be uniform in $k$.

Solution

Since $N(0,1)^{\mathsf T}=(1,0)^{\mathsf T}$, the solution is $e^{ikt}(ikt,1)^{\mathsf T}$. Its norm grows at least as $|k|t$ times the initial norm. Choosing data concentrated at successively larger $k$ defeats any estimate with a constant independent of frequency, so the evolution loses one derivative and is not strongly hyperbolic in this formulation.

Why does a positive reduced kinetic matrix fail to prove well-posedness of the full gauge-fixed gravitational equations?

Solution

The reduced matrix tests physical perturbations after constraints and gauge components have been removed. A Cauchy formulation must also evolve gauge variables and propagate constraints; their principal-symbol blocks depend on gauge drivers, reductions, and constraint additions. Those blocks can contain a defective Jordan structure even when the physical action is positive, so a separate full-symbol and symmetrizer test is required.

References

• Kovacs, Aron D., and Harvey S. Reall. 2020. “Well-Posed Formulation of Lovelock and Horndeski Theories.” Physical Review D 101, 124003. DOI; arXiv:2003.08398.
• Papallo, Giuseppe, and Harvey S. Reall. 2017. “On the Local Well-Posedness of Lovelock and Horndeski Theories.” Physical Review D 96, 044019. DOI; arXiv:1705.04370.