The Nonlinear Stability of Minkowski Space

The nonlinear stability of Minkowski space says that sufficiently small, suitably decaying vacuum perturbations of flat initial data have a maximal globally hyperbolic development that exists for all time, is causally geodesically complete, and disperses toward flat geometry in quantified asymptotic senses. The theorem is global and geometric: it is not merely long-time existence of coordinate components or boundedness of a linearized field.

Required background. Analysis for Mathematical Relativity supplies weighted Sobolev estimates; The Einstein Equation supplies the reduced vacuum system; PDEs and Green Functions for Gravitation supplies wave energy and decay.

Helpful background. Causality and Global Structure clarifies completeness and conformal asymptotics.

What “small asymptotically flat data” means

Let $(\Sigma,h,K)$ be a smooth vacuum initial data set on a manifold diffeomorphic to $\mathbb R^3$, satisfying

$$R(h)+K^2-K_{ij}K^{ij}=0, \qquad D^j(K_{ij}-Kh_{ij})=0,$$

with

$$K_{ij}=-h_i{}^\mu h_j{}^\nu\nabla_\mu n_\nu.$$

Smallness is measured relative to $(\delta,0)$ in a high-order weighted Sobolev or geometric curvature norm. The weights encode decay toward Euclidean data, allow the asymptotic charges appropriate to the theorem, and distinguish an asymptotically flat end from a merely locally small perturbation. The constraints are part of the hypothesis; arbitrary small pairs $(h-\delta,K)$ are not admissible data.

Different proofs use different gauges and decay assumptions. Christodoulou and Klainerman use a geometric maximal foliation and curvature energies. Lindblad and Rodnianski use wave coordinates and the weak-null structure of the reduced equations. Their hypotheses and asymptotic statements are not identical, though both establish nonlinear stability of the same flat spacetime geometry. Shen compares the proof systems and borderline decay issues in Shen 2026, survey overview.

A geometric theorem statement

Here is a concrete sufficient data class from the wave-coordinate proof. Choose a reference length $L$ and rescale all geometric lengths by $L$. Use dimensionless coordinates $y=x/L$, $\rho=\lvert y\rvert$, and dimensionless tensor components $h$ and $k=LK$ in the rescaled geometry. In a global asymptotically Cartesian chart, write

$$h_{ij}=\delta_{ij}+\chi(\rho)\frac{\mu}{\rho}\delta_{ij}+q_{ij}, \qquad \mu=\frac{2GM_{\mathrm{ADM}}}{L},$$

where the smooth cutoff $\chi$ is zero for $\rho\le1/2$ and one for $\rho\ge3/4$. The factor of two translates the coefficient of the source’s $1/r$ tail into the site’s physical ADM-mass convention. Assume smooth, constraint-satisfying data with $q=o(\rho^{-1-\alpha})$ and $k=o(\rho^{-2-\alpha})$ for some $\alpha\gt0$, and define

$$\begin{aligned} \mathcal E_N={}&\sum_{\lvert I\rvert\le N} \left\|(1+\rho)^{1/2+\gamma+\lvert I\rvert} \partial_y\partial_y^I q\right\|_{L^2(d^3y)}\\ &+\sum_{\lvert I\rvert\le N} \left\|(1+\rho)^{1/2+\gamma+\lvert I\rvert} \partial_y^I k\right\|_{L^2(d^3y)}. \end{aligned}$$

For $N\ge6$, fixed sufficiently small $\gamma\gt0$, and $\mathcal E_N+\mu\le\varepsilon$ sufficiently small depending on $\gamma$, the vacuum development is future causally geodesically complete and has quantitative decay in global wave coordinates. Time reversal, $K\mapsto-K$, gives past completeness. This is the vacuum specialization of Lindblad–Rodnianski 2010, Theorem 1.1, PDF pp. 5–6. Christodoulou and Klainerman’s geometric norms and sharper curvature hierarchy are developed in Christodoulou–Klainerman 1993, Chapters 8–17.

The conclusion is uniqueness up to a diffeomorphism fixing the initial data. It does not say that $g_{\mu\nu}-m_{\mu\nu}$ decays in every coordinate system, that the ADM mass vanishes, or that every small unweighted norm suffices.

The global proof architecture

The argument combines six mechanisms.

1. Local development. A hyperbolic reduction gives a solution while a continuation norm remains finite.
2. Bootstrap region. Assume weighted energies are bounded by $C\varepsilon$ on a maximal time interval.
3. Decay from commutation. Vector fields or null foliations convert energies into pointwise decay, with stronger decay for derivatives tangent to outgoing cones.
4. Null structure. The worst parallel null interactions cancel or enter a hierarchy that can be integrated.
5. Strict improvement. Error integrals are bounded by a quantity smaller than the bootstrap allowance, such as $C_0\varepsilon+C_1(C\varepsilon)^2$.
6. Continuation and completeness. The improved estimates prevent the breakdown quantities from diverging, extend the solution indefinitely, and make the affine-length integrals along causal geodesics diverge.

Each step is necessary. Energy boundedness without decay leaves quadratic errors nonintegrable in time. Coordinate decay without constraint propagation need not solve the vacuum equations. Infinite coordinate time without lapse and connection control does not prove geodesic completeness.

Why the nonlinear terms are borderline

A radiating linear wave in $3+1$ dimensions decays roughly like $t^{-1}$ in the wave zone $r\sim t$; decay inside the cone is stronger under suitable data assumptions. For example, a compactly supported free wave in flat $3+1$ dimensions eventually vanishes at any fixed spatial point by Huygens’ principle. A generic quadratic source is therefore borderline near the expanding light cone: in an energy estimate, a coefficient of size $t^{-1}$ has a logarithmically divergent time integral. The classical null condition removes products in which both derivatives point in the same bad null direction. Einstein’s wave-coordinate equations obey a weaker hierarchical condition rather than the classical null condition in every component.

This is why the theorem is not a routine perturbation of linear stability. The gauge, constraint equations, asymptotic mass tail, and hierarchy among components all enter the global closure. The next pages isolate the vector-field estimates and null structure used here.

Limitations

The smallness assumption excludes generic strong-field data and says nothing by itself about thresholds between dispersion and black-hole formation. The vacuum theorem does not automatically include fluids, whose shocks may form, or massive fields with different decay. It also does not establish a smooth Penrose compactification for every admitted decay class; peeling and completeness of null infinity are stronger asymptotic questions.

Finally, stability is orbital and asymptotic modulo gauge and asymptotic parameters. A compactly supported coordinate transformation can make metric components look large while leaving the geometry flat, so component closeness is not a gauge-invariant conclusion.

Solved diagnostic: why a strict improvement matters

Use the proposed bootstrap bound below to improve its constant by a factor of two. What changes if the integrable time weight is replaced by $(1+s)^{-1}$? Here time is measured in a fixed reference unit.

Solution

Assume a bootstrap energy $E(t)\le4\varepsilon^2$ and derive

$$E(t)\le \varepsilon^2+C\int_0^t(1+s)^{-3/2}E(s)^{3/2}\,ds.$$

Using the bootstrap bound,

$$E(t) \le\varepsilon^2 +8C\varepsilon^3\int_0^\infty(1+s)^{-3/2}ds =\varepsilon^2+16C\varepsilon^3.$$

If $\varepsilon\le(16C)^{-1}$, then $E(t)\le2\varepsilon^2$, strictly improving $4\varepsilon^2$. Openness and local continuation then extend the bootstrap interval. If the time weight were $(1+s)^{-1}$, the integral would grow logarithmically and this uniform closure would fail without additional structure.

References

• Christodoulou, Demetrios, and Sergiu Klainerman. The Global Nonlinear Stability of the Minkowski Space. Princeton Mathematical Series 41. Princeton University Press, 1993. DOI.
• Lindblad, Hans, and Igor Rodnianski. “The Global Stability of Minkowski Space-Time in Harmonic Gauge.” Annals of Mathematics 171 (2010): 1401–1477. arXiv:math/0411109.
• Shen, Dawei. “The Stability of Minkowski Spacetime.” arXiv:2605.26506, 2026.