Horndeski Theory¶

The four-dimensional single-scalar metric action family whose Euler-Lagrange field equations are constrained to remain second order.

Version: v3 · 2026-09-06 · History
Domain-specific #: 2020
Origin domain: gravitational physics
Subdomain: scalar-tensor gravity
Aliases: Horndeski's theory, Horndeski gravity, Horndeski scalar-tensor theory

Core Idea¶

Horndeski theory is the complete four-dimensional family of local, generally covariant theories built from one spacetime metric and one scalar field whose Euler–Lagrange equations for both fields contain derivatives of no higher than second order. Gregory Horndeski classified that family in 1974^[1]. A modern equivalent representation writes its gravitational action as a sum of four Lagrangian sectors governed by freely chosen functions \(G_2(\phi,X)\), \(G_3(\phi,X)\), \(G_4(\phi,X)\), and \(G_5(\phi,X)\). Here \(\phi\) is the scalar field and, in the common ((-+++)) signature convention,

\[ X=-\frac12 g^{\mu\nu}\nabla_\mu\phi\nabla_\nu\phi. \]

The abstraction is not merely “a theory with a scalar.” It is a constrained action architecture: derivative couplings that would ordinarily produce higher derivatives in the equations of motion occur in special combinations whose higher-derivative contributions cancel after variation. Choosing the four functions selects a particular model inside the family; specifying a background solution, matter coupling, parameters, and initial or boundary data makes that model predictive. The family supplies a common language in which researchers can derive backgrounds and perturbations, locate familiar subfamilies, test stability, compare observational restrictions, and identify when a proposed scalar–tensor interaction lies outside the second-order class.

The historical superlative must be bounded carefully. Horndeski is the most general family satisfying the stated assumptions and the second-order-equation requirement. It is not the most general healthy scalar–tensor theory now known. Beyond-Horndeski and degenerate higher-order scalar–tensor theories can have higher-order field equations while a degeneracy constraint prevents the generic extra Ostrogradsky mode^[2]. “Second order” is therefore Horndeski's defining classification rule and a powerful sufficient safeguard against a generic nondegenerate higher-derivative instability, not a necessary or complete certificate of physical viability.

Structural Signature¶

Using \(\phi_{\mu\nu}=\nabla_\mu\nabla_\nu\phi\), \(\Box\phi=g^{\mu\nu}\phi_{\mu\nu}\), and \(G_{iX}=\partial G_i/\partial X\), a common modern convention is

\[ S=\int d^4x\sqrt{-g}\left(\sum_{i=2}^{5}{\cal L}_i+{\cal L}_{m}\right), \]

with

\[ {\cal L}_2=G_2(\phi,X), \qquad {\cal L}_3=-G_3(\phi,X)\Box\phi, \]

\[ {\cal L}_4=G_4(\phi,X)R+G_{4X}\left[(\Box\phi)^2-\phi_{\mu\nu}\phi^{\mu\nu}\right], \]

and

\[ \begin{aligned} {\cal L}_5={}&G_5(\phi,X)G_{\mu\nu}\phi^{\mu\nu}\\ &-\frac{G_{5X}}{6}\left[(\Box\phi)^3-3\Box\phi\,\phi_{\mu\nu}\phi^{\mu\nu} +2\phi_{\mu}^{\ \nu}\phi_{\nu}^{\ \rho}\phi_{\rho}^{\ \mu}\right]. \end{aligned} \]

Signs, integrations by parts, scalar normalization, and names for the free functions vary among authors. The particular printed (G_i) are therefore representation-dependent bookkeeping objects, not direct observables. The recognition test is structural:

Field-content role: one metric \(g_{\mu\nu}\) and one scalar \(\phi\) furnish the gravitational-sector degrees of freedom.
Action role: local generally covariant scalar densities built from those fields and their derivatives are varied to obtain dynamics.
Free-function role: functions of \(\phi\) and (X) select members of the family and organize nested subclasses.
Cancellation role: curvature and second-derivative scalar terms carry tuned relative coefficients.
Invariant: variation yields at most second-order equations for both metric and scalar under the defining assumptions.
Model-completion role: a matter-coupling prescription, background, parameters, and data must be declared before observation claims follow.
Health role: kinetic matrices, propagation speeds, strong-coupling scales, and other stability conditions are checked after membership; membership alone does not make every model viable.

In a regular region, the theory normally describes the two tensor polarizations of the metric plus one scalar mode^[3]. That degree-of-freedom count can fail at singular parameter choices, strongly coupled backgrounds, or poorly posed limits, so it is a generic structural expectation rather than an unconditional theorem about every written function choice.

What It Is Not¶

Horndeski theory is not generic modified gravity. Curvature corrections, extra vector fields, multiple scalars, nonlocal models, higher-dimensional theories, and Lorentz-violating constructions may all modify general relativity without belonging to this family. Nor is every scalar–tensor action Horndeski: an arbitrary derivative coupling generally produces higher-order equations unless it takes the classified form or is equivalent to it by allowed identities and boundary terms.

It is not a single fixed cosmological model. The four functions contain broad freedom. Statements such as “Horndeski predicts cosmic acceleration,” “Horndeski screens locally,” or “Horndeski has a particular gravitational-wave speed” omit the function choice, background, matter frame, and scale on which the claim depends.

It is not conformal gravity. Four-dimensional conformal gravity is organized around local Weyl invariance and a curvature-squared action, ordinarily leading to fourth-order metric equations. Horndeski is organized around one scalar, one metric, general covariance, and the second-order cancellation condition. It is not Lovelock gravity either: Lovelock classifies metric curvature invariants giving second-order metric equations in general dimension; ordinary four-dimensional Lovelock dynamics reduce to Einstein gravity with a cosmological constant, up to topological terms^[4]. Horndeski obtains additional scalar–tensor dynamics through a different field-content and cancellation architecture.

It is not equivalent to the effective field theory of dark energy. A background-dependent unitary-gauge EFT parameterizes operators and perturbation functions useful for cosmology, while Horndeski specifies a covariant action family. Mappings exist under stated assumptions, but neither label can be substituted for the other without preserving the level of description.

Finally, it is not synonymous with covariant Galileon, generalized Galileon, generalized G-inflation, beyond Horndeski, or DHOST. The first three are historically and mathematically close formulations or subfamilies whose exact relation depends on convention. Beyond-Horndeski and DHOST deliberately enlarge the healthy scalar–tensor landscape beyond the second-order-equation boundary.

Scope of Application¶

The home domain is classical gravitation and its cosmological applications. Horndeski reasoning is used when constructing or comparing scalar–tensor models of inflation, late-time acceleration, modified gravitational clustering, compact objects, and screened departures from general relativity. It also organizes perturbation theory: once a background is chosen, one derives tensor and scalar kinetic coefficients, gradient terms, mixing, effective gravitational couplings, and observables. It is consequently both a model-generation framework and a diagnostic classification.

The scope is restricted by the assumptions behind the classification. The usual modern identity is four-dimensional, local, generally covariant, and contains one scalar plus one metric. Multi-scalar theories, vector–tensor theories, nonlocal interactions, explicitly broken covariance, and higher-dimensional constructions require other categories even when they share some formulas. Matter is often assumed minimally coupled to a Jordan-frame metric, but this is not forced by naming the gravitational action. A conformal or disformal change of variables can make one sector look simpler while moving complexity into matter couplings; physical interpretation must track which metric matter follows.

The theory is also an effective description with a domain of validity. Derivative interactions that are harmless at classical field-equation level may become strongly coupled above a cutoff, and radiative stability depends on the model and symmetries. Cosmological viability does not imply Solar-System or binary-system viability. Likewise, an observational restriction at low redshift and gravitational-wave frequencies does not mechanically settle the ultraviolet completion.

Clarity¶

The clearest membership question is counterfactual: if the displayed scalar and metric action is varied without imposing a special background, do all third and higher derivatives cancel from both Euler–Lagrange equations? A positive result, together with the field-content, dimensionality, locality, and covariance assumptions, places the action in Horndeski or an equivalent representation. Cancellation only on a homogeneous background is insufficient; an interaction may hide higher derivatives when symmetry has removed them.

A second question separates family membership from model health: after the member is selected, do its relevant backgrounds have the intended number of degrees of freedom with acceptable kinetic and gradient behavior? A member can pass the identity test yet have ghosts, gradient instabilities, superluminal regimes that demand careful interpretation, a vanishing kinetic coefficient, an unacceptably low strong-coupling scale, or phenomenology already excluded. These are bad or restricted Horndeski models, not evidence that the family identity has disappeared.

A third question separates representation from physics. Two sets of (G_i) may differ by integrations by parts or field transformations while describing equivalent dynamics under appropriate conditions. Conversely, a disformal transformation accompanied by a different matter-coupling declaration may change what observers call the physical metric. Comparison therefore proceeds through actions, transformations, degrees of freedom, and observables rather than through function names alone.

Manages Complexity¶

Without the Horndeski classification, a model builder confronting scalar derivative couplings must vary each candidate action, locate all higher derivatives, discover cancellations term by term, and determine whether the resulting equations propagate unwanted modes. The family compresses that search into four function slots with known compensating combinations. This turns an unstructured space of derivative interactions into a navigable function space.

The compression supports several recurring operations. Researchers can recover a subclass by setting functions to zero or choosing a restricted functional dependence; translate a covariant model into background and perturbation coefficients; identify which function derivatives control kinetic mixing or tensor propagation; and apply theory, stability, local-gravity, and cosmological filters in sequence. The family also provides a common comparison frame for papers that otherwise use different names, integrations by parts, or parameterizations.

This complexity management does not eliminate calculation. The background equations, constraint reduction, quadratic perturbation action, matter response, screening regime, and observables still depend on the selected functions. Rather, Horndeski organizes where those calculations begin and which consistency gates must remain distinct.

Abstract Reasoning¶

Horndeski reasoning can be expressed as an ordered pipeline.

Membership gate. Declare the fields, dimension, covariance and locality assumptions. Rewrite the action in a recognized basis or directly vary it. Verify the general second-order cancellation, allowing legitimate boundary terms and identities but not relying on a special solution.

Subclass gate. Inspect the (G_i). If only (G_2) is nontrivial and (G_4) is constant, the model lies in a minimally coupled k-essence-like sector. Scalar dependence in (G_4) introduces nonminimal coupling. (X)-dependent (G_4) and nontrivial (G_5) encode richer derivative couplings. This immediately predicts which calculations and restrictions are likely to matter, though it does not replace them.

Degree-of-freedom and health gate. Choose a background and derive the reduced kinetic and gradient matrices after constraints are handled. Positive kinetic coefficients and acceptable squared propagation speeds are background-dependent requirements. A second-order equation prevents the generic Ostrogradsky construction but does not prove these inequalities, a well-posed Cauchy problem, radiative naturalness, or weak coupling throughout the desired regime.

Observation gate. Map the covariant functions into observable or effective parameters for the regime under study. The multimessenger event GW170817 and its gamma-ray counterpart constrained the low-redshift tensor speed to be extremely close to the speed of light^[5]. Under the additional requirement that luminality hold robustly on general cosmological backgrounds, this sharply restricts many quartic and quintic derivative couplings—often summarized in common conventions as requiring (G_{4X}) to be negligible and (G_5) to be effectively constant in the relevant regime. It does not remove (G_2), (G_3), or a scalar-dependent (G_4), and it does not by itself prove that every surviving model is viable.

Boundary gate. If an action has higher-order equations but a degenerate kinetic structure removes the extra mode, it belongs to beyond-Horndeski or DHOST territory rather than passing the Horndeski membership rule. If it adds fields or breaks other constitutive assumptions, route it to the corresponding broader theory family.

These gates license useful negative inferences. A claimed “Horndeski prediction” with no selected functions is underdetermined. A theory with second-order equations but two independent scalars does not qualify under the standard identity. An observationally excluded member remains an instance of the family even though it is no longer a plausible model of nature.

Knowledge Transfer¶

Transfer is strongest inside gravitational physics. The same action-classification apparatus can be carried from inflation to late-time dark energy, from homogeneous backgrounds to perturbations, and from cosmology to compact-object or screening analyses, provided the field and coupling assumptions remain explicit. The family also teaches a general field-theory lesson: apparently dangerous higher-derivative terms can be rendered constrained by correlated coefficients, and degree-of-freedom counting must follow the full constraint structure rather than the highest visible derivative alone.

That lesson can inform the study of DHOST, effective field theory, and constrained mechanical systems, but the name “Horndeski” should not travel with the generic lesson. Outside single-scalar metric gravity, the portable residue is already expressed by broader abstractions such as variational formulation, constraint, cancellation, degeneracy, invariance, and degrees of freedom. Calling any tuned higher-derivative system “Horndeski-like” is analogy, not exact instantiation.

Examples¶

Canonical scalar or quintessence. Take \(G_2=X-V(\phi)\), \(G_4=M_{\mathrm{Pl}}^2/2\), and (G_3=G_5=0). The result is Einstein gravity plus a canonical scalar potential. This simple corner demonstrates that Horndeski contains ordinary scalar cosmology; the family is not defined by exotic derivative interactions.

k-essence. Allow an arbitrary \(G_2(\phi,X)\) while retaining constant (G_4) and vanishing (G_3,G_5). Nonlinear kinetic structure changes scalar pressure and propagation while remaining inside the family.

Brans–Dicke-like nonminimal coupling. A scalar-dependent \(G_4(\phi)\), together with a corresponding scalar kinetic term in (G_2), produces the familiar variable gravitational-coupling architecture after conventions and field normalizations are fixed. “Brans–Dicke” names a restricted model family, not all of Horndeski.

Covariant Galileon. Special polynomial choices of the free functions reproduce covariant Galileon interactions. Curvature counterterms compensate derivative terms so that the general field equations remain second order^[6]. This is a canonical demonstration of the cancellation role, but covariant Galileon is a subclass rather than an unrestricted alias.

Luminal post-GW170817 sector. A model with nontrivial (G_2), (G_3), and \(G_4(\phi)\), but without the derivative couplings that robustly shift tensor speed, illustrates how the multimessenger constraint prunes the function space without deleting the class. Scalar stability, local tests, background fit, and strong coupling still require separate checks.

Nonexample: arbitrary curvature-squared gravity. Adding (R^2) or Weyl-tensor-squared terms to a metric action does not become Horndeski simply because the purpose is modified gravity. The field content, derivative structure, and resulting equations differ. Metric (f(R)) models can sometimes be rewritten as a particular scalar–tensor representation through a nonsingular auxiliary-field/Legendre transformation; that correspondence is a bounded relationship, not identity with the full family.

Nonexample: a degenerate higher-order theory. A DHOST action can display third or higher derivatives in its equations yet propagate only the intended scalar and tensor modes because its kinetic matrix is degenerate. It may be healthy in an appropriate regime, but it fails Horndeski's second-order recognition rule.

Structural Tensions¶

Generality versus predictivity. Four free functions provide wide model coverage, but the unselected family makes few unique predictions. A productive analysis narrows the functions by symmetry, effective-field-theory power counting, stability, and data rather than treating flexibility as explanatory success.

Derivative richness versus control. The classified combinations admit sophisticated interactions without generic higher-order equations. The same interactions can produce strong coupling, kinetic mixing, or background-dependent instabilities. Passing the cancellation test opens the analysis; it does not finish it.

Screening versus testability. Nonlinear derivative interactions can suppress deviations from general relativity near dense sources. Successful screening may reconcile cosmology with local tests, yet it can also hide effects or introduce a low cutoff. One must specify the regime and transition rather than claim screening from family membership.

Covariant action versus phenomenological coordinates. The (G_i) make field content and covariance visible, while effective cosmological functions can make observation fits more direct. Translation is powerful but background- and convention-dependent; neither representation is a universally sufficient description.

Frame simplicity versus matter complexity. Conformal or disformal transformations can simplify the gravitational action, but they may complicate matter coupling and the interpretation of clocks, rulers, or light cones. Formal equivalence requires transforming the whole system, not just one sector.

Second-order clarity versus broader health. Horndeski offers an exceptionally clean syntactic rule. DHOST demonstrates that degeneracy, rather than equation order alone, is the deeper condition governing an extra Ostrogradsky mode. The older boundary remains useful precisely because it is narrower and easier to recognize.

Cosmological freedom versus multimessenger restriction. Function freedom allows acceleration and modified growth, while gravitational-wave speed measurements remove broad regions under reasonable robustness assumptions. The tension is resolved through conditional model selection, not through the claim that the entire class is ruled out.

Structural–Framed Character¶

Horndeski theory is strongly structural, with an estimated structural–framed aggregate of 0.0. Its membership is determined by fields, covariance, an action, derivative order, and cancellation identities. It is discovered through mathematical variation rather than assigned by an institution, negotiated by convention, or evaluated by a normative standard. Names and sign conventions vary, but those presentational choices do not create membership.

The abstraction is nevertheless domain-specific. “Structural” does not mean “prime.” Removing scalar fields, Lorentzian spacetime, curvature, covariant derivatives, action variation, and the four-dimensional classification problem leaves a generic pattern of constrained variation and cancellation that is already represented by broader abstractions. Those gravitational commitments are constitutive, not decorative framing.

Structural Core vs. Domain Accent¶

The structural core is a constrained generative family: freely variable components occupy named slots, compensating combinations preserve an invariant under an operation, special settings recover nested subclasses, and downstream tests prune the remaining space. That skeleton is portable to many taxonomies and model architectures.

The domain accent is indispensable to Horndeski identity. The objects are a spacetime metric and scalar field; the generating object is a generally covariant action; the operation is Euler–Lagrange variation; the invariant is second-order field equations; and the compensations combine curvature with second covariant derivatives of the scalar in four dimensions. The function slots (G_2) through (G_5), tensor/scalar degrees of freedom, matter frames, gravitational-wave propagation, and cosmological stability are not optional examples. They define what the classification is for.

Accordingly, the core may inspire cross-domain reasoning, but only gravitational instances satisfying the complete recognition boundary instantiate Horndeski theory. A software architecture with four plug-ins or a mechanical model with tuned constraints is not a Horndeski instance merely because it shares the generative-family shape.

Horndeski theory instantiates Principle of Least Action: its defining data are organized in an action, and varying that action generates the metric and scalar equations. The relation is compositional rather than taxonomic. A theory family is not itself a subtype of the variational principle; it uses that principle as its generative mechanism. This is the sole proposed DAG edge because it is both literal and minimal.

It also relates in prose to Constraint, Cancellation, Degrees of Freedom, Invariance, Symmetry, and Classification. Constraint and Cancellation describe why the derivative combinations work. Degrees of Freedom separates apparent equation order from physical propagation. Invariance and Symmetry guide admissible action terms. Classification describes the family-building role. None alone, and no simple conjunction of them, supplies the scalar–tensor field content, four-dimensional covariance, (G_i) decomposition, or general second-order-equation boundary.

Relationships to Other Abstractions¶

Current abstraction Horndeski Theory Domain-specific

Parents (1) — more general patterns this builds on

Horndeski Theory is a kind of Principle of Least Action Prime

Horndeski theory instantiates Principle of Least Action: its defining data are organized in an action, and varying that action generates the metric and scalar equations.

Hierarchy path (1) — routes to 1 parentless root

Horndeski Theory → Principle of Least Action

Neighborhood in Abstraction Space¶

Horndeski Theory sits in a sparse region of the domain-specific corpus (94^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Algebraic Structure & Reciprocity Theorems (5 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

General scalar–tensor theory: broader label; may include higher-order, degenerate, multiple-scalar, or differently coupled constructions.
Generalized G-inflation/generalized Galileon: modern formulations crucial to the rediscovery and practical representation of Horndeski, but their naming and scope are not automatically exact aliases in every source.
Covariant Galileon: a particular derivative-interaction subclass.
Beyond Horndeski/GLPV: extension whose equations may be higher order while constraints preserve the intended degrees of freedom.
DHOST: still broader degenerate higher-order scalar–tensor families, with Horndeski occupying a subset of the healthy landscape.
Brans–Dicke theory, k-essence, quintessence, and metric (f(R)): special models or representable subfamilies under declared transformations and assumptions.
Effective field theory of dark energy: a background-oriented operator/parameter framework, not the same covariant family.
Conformal gravity: a Weyl-invariant curvature-squared theory with a different defining symmetry and derivative structure.
Lovelock gravity: a metric curvature classification in arbitrary dimension, not the single-scalar family.
Ostrogradsky instability: a pathology of generic nondegenerate higher-derivative dynamics; avoiding generic Ostrogradsky reasoning does not uniquely identify Horndeski or guarantee every other stability property.

References¶

[1] Horndeski. “Second-order scalar-tensor field equations in a four-dimensional space”. International Journal of Theoretical Physics, 1974. Original 1974 paper classifying the most general four-dimensional scalar-tensor theory with second-order field equations. registry ↩

[2] Langlois. “Dark energy and modified gravity in degenerate higher-order scalar–tensor (DHOST) theories: A review”. International Journal of Modern Physics D, 2019. Review stating the degeneracy condition under which beyond-Horndeski and DHOST theories avoid the generic extra Ostrogradsky mode despite higher-order field equations. registry ↩

[3] Kobayashi. “Horndeski theory and beyond: a review”. Reports on Progress in Physics, 2019. Review stating that Horndeski theory generically propagates two tensor polarizations plus one scalar degree of freedom in regular regions of parameter space. registry ↩

[4] Lovelock. “The Four-Dimensionality of Space and the Einstein Tensor”. Journal of Mathematical Physics, 1972. Proves the four-dimensional reduction of Lovelock gravity to the Einstein tensor plus cosmological constant; the preceding general-dimension classification of Lovelock invariants is established in the earlier Lovelock (1971) paper, which this source does not supply. registry ↩

[5] Abbott, et al. “Gravitational Waves and Gamma-Rays from a Binary Neutron Star Merger: GW170817 and GRB 170817A”. The Astrophysical Journal Letters, 2017. Joint GW/gamma-ray multimessenger observation (GW170817/GRB170817A) constraining the low-redshift gravitational-wave speed to be extremely close to the speed of light. registry ↩

[6] Deffayet, Esposito-Farèse, and Vikman. “Covariant Galileon”. Physical Review D, 2009. Constructs the covariant Galileon and shows that added curvature counterterms cancel higher-derivative terms so the field equations remain second order. registry ↩