WorksPhysics and Astronomy
The Mechanics of Curvature Inflation
Abstract
We examine slow roll directly in the Jordan-frame Hubble equation. For the quadratic Starobinsky action, this equation has an exact mechanical form. Comparing nearby solutions at the same Hubble value gives a local contraction law on the expanding branch with decreasing Hubble rate. A signed square root of the Hubble parameter also continues the vacuum background smoothly through its Hubble zeroes. For metric actions built from a smooth function of the Ricci scalar, coefficients depending only on the Hubble parameter require a derivative expansion. We derive the terms through cubic order and test both the equation residual and the error accumulated by an independently evolved approximation. A regularized inverse-logarithmic extension provides a nontrivial example. On the parameter slice studied, we prove a positive lower bound on the second curvature derivative of the action and compare separate Jordan- and Einstein-frame integrations. The third-order approximation reproduces the exact inflationary trajectory more accurately than the second-order approximation over the interval tested. We also compute the linear scalar and tensor spectra with numerical convergence checks. The benchmark gives a scalar spectral index of 0.97384 and a tensor-to-scalar ratio of 0.00527, evaluated 60 exact Einstein-frame e-folds before the first acceleration exit. These values depend on the chosen action and reference event. The mechanical form supplies approximation tests for the background; the scalaron and its perturbations follow from the original gravitational action.
Introduction
Slow roll is usually described through the motion of a scalar field on a nearly flat potential. In the Starobinsky model, the same approximation can also be examined directly in the equation for the Jordan-frame Hubble parameter. That equation contains terms proportional to the Hubble acceleration, velocity, and velocity squared. Comparing their sizes shows which balance supports slow evolution and how the approximation changes near the end of inflation. Here we derive these comparisons and test the resulting approximate trajectories against the exact background.
The Starobinsky action adds a positive curvature-squared term to the Einstein–Hilbert action [1, 2]. In metric gravity this introduces a scalaron, whose potential is nearly flat during inflation and quadratic near its minimum [3, 4]. The scalaron rolls during inflation and then undergoes coherent vacuum oscillations [5, 6]. The Einstein-frame potential makes this evolution and the calculation of fluctuations transparent. It provides a standard framework for studying inflationary dynamics and primordial perturbations in metric gravity [3, 4, 7, 8]. We restrict attention to metric theories; more general quadratic-curvature actions can contain additional degrees of freedom [9, 10].
We use the term mechanical for a particular organization of the homogeneous Hubble equation. Its acceleration, velocity-dependent, and algebraic terms resemble inertia, resistance, and a driving force. This interpretation is related to effective-fluid descriptions of higher-curvature terms [11]. The individual coefficients depend on the Hubble coordinate and on the normalization of the equation. They are useful for testing an approximation, but they do not define a physical particle mass or an independent channel of energy transfer. The field content remains that of the original action.
In the quadratic model, the mechanical equation is exact. We compare two decreasing-Hubble solutions at the same Hubble value, so an arbitrary shift of cosmic time does not enter the comparison. Their velocity difference obeys a contraction identity on their common expanding branch. This provides a precise local characterization of the approach toward the slowly evolving branch, complementing earlier analyses of inflationary attractor behavior in cosmology [6]. We also derive corrections to the nearly constant slow-roll velocity. During the later vacuum oscillations the Jordan-frame Hubble parameter repeatedly reaches zero. The vanishing of the highest-derivative coefficient at these points reflects a degeneracy of the Hubble-variable representation, rather than a physical singularity. A signed square-root variable gives a regular continuation. Matter production and thermal reheating require further equations [12, 13].
For a general smooth , the exact Hubble equation has curvature-dependent coefficients. Since the curvature itself contains the Hubble velocity, replacing those coefficients by functions of the Hubble parameter alone requires an expansion. We give that expansion through cubic derivative order, including the mixed acceleration–velocity term. We then evaluate the truncated equations on an exact background and integrate them separately from the same initial state. These are two different tests: a small instantaneous residual need not imply a comparably small trajectory error. This extends the exact quadratic formulation to general smooth theories through a controlled derivative expansion and makes it possible to assess not only whether a particular term is small, but also whether the resulting approximation reproduces the background over the interval of interest.
Our numerical example adds a regularized inverse-logarithmic correction to the quadratic action. Logarithmic curvature corrections and running quadratic-gravity actions have already been studied extensively [14, 15, 16]. The high-curvature action, exact Hubble phase-space equation, and leading inverse-field spectral scaling overlap with that earlier work. Here an Einstein–Hilbert term and a smooth low-curvature continuation allow us to follow the vacuum solution through zero curvature. The regularization is a phenomenological choice. On the parameter domain considered, we establish an analytic lower bound on and follow the background in both frames. § 4.2 gives the explicit relation to the running-gravity model. The purpose of the example is to test the derivative expansion on a regular branch.
We calculate scalar and tensor spectra from the standard quadratic perturbation actions [17, 18, 19]. The numerical modes also provide a check on potential slow-roll estimates [20, 21, 22]. The observational interpretation of a predicted tilt depends on the data combination and the cosmological parameter space [23, 24, 25, 26]. Our quoted values refer to specified e-fold events. Matching such an event to a dimensional CMB pivot requires a post-inflationary expansion history; testing statistical agreement requires a likelihood analysis. Neither is supplied by the mechanical rewriting.
The paper is organized as follows. § 2 treats the exact quadratic model, its contraction identity, and its regular Hubble zeroes. § 3 gives the canonical scalaron description and the general derivative expansion. § 4 introduces the regularized example and its branch properties. § 5 and § 6 present the background tests and linear spectra. Numerical mode details and convergence checks are collected in Appendix § 10. § 7 summarizes the results. We use signature .
Exact Hubble dynamics in the quadratic model
Constraint, trace and mechanical equation
The metric action is given by
on a spatially flat FLRW background. Here and . Unless an Einstein-frame derivative is written explicitly, a dot means . The vacuum equations include
The trace and Ricci equations propagate the constraint. In fact, for , they imply . Conversely, differentiating the constraint at recovers the spatial background equation. At a zero of , the trace–Ricci system should be retained instead of dividing the constraint by .
For with , we have , and . The Friedmann constraint can therefore be written as
where and have been used without approximation. The resulting exact equation is
In the normalization used here, this is
The force-form terms are , and . The signs here refer to the acceleration side of the equation, rather than to a dissipative stress tensor: the terms are produced entirely by rearranging the field equation. On a decreasing- branch, the first two oppose the motion. The quadratic force points toward increasing for either velocity sign, so it is not a conventional resistive force on both branches. The coefficient is an effective inertia, not a particle mass. Multiplication of the differential equation by a nonzero function changes these coefficients, while leaving its solutions unchanged. In particular, the terms called damping in this analogy do not describe particle production or energy transfer to a matter sector.
Define the Jordan-frame Hubble-flow parameters
On an interval with , , the exact term ratios are
Both are small on the relaxed slow-roll branch, rather than on every solution with an initially small . Dropping them relative to the linear term gives the familiar drift
Equation (7) is a consistency check of that approximation. A statement about attraction requires a comparison of neighboring solutions.
Local contraction and the slow-roll branch
On a monotone interval, let . Equation (4) becomes
If two solutions are compared at the same , their difference obeys the exact identity
The coefficient is positive, so decreases as decreases along their common interval. This comparison eliminates the arbitrary choice of where the two solutions sit in cosmic time: it asks whether two histories approach the same velocity when they pass through the same Hubble value. For an infinitesimal displacement about a reference solution, with increasing in time,
This is a conditional attraction statement in the phase plane: both trajectories must remain on the expanding, decreasing- branch. It is not a proof of a global basin of attraction, nor a comparison at equal cosmic times. In the leading drift regime the rate tends to , consistent with the known transient-attractor behavior [6].
To resolve the subleading drift, put and write . Equation (9) gives
The asymptotic branch with at large is
Consequently,
The inertial ratio on this branch is therefore smaller than a generic first-order slow-roll estimate.
Starting a numerical solution with only the leading velocity in Eq. (8) introduces a short transient and need not reproduce Eq. (14) immediately.
For the diagnostic plot we first allow the numerical trajectory to relax toward the decreasing-Hubble slow-roll branch, rather than treating a leading-order drift initial condition as exact. The figures use a solution started at , , evolved for more than 52 Jordan-frame e-folds before the displayed interval. The reference event is then selected 60 Jordan-frame e-folds before the first crossing. At that event,
The drift series reproduces this velocity to a relative difference of . At , and ; the quadratic term is still one sixth of the linear term (Figure 1). Thus the first loss of accelerated expansion does not coincide with quadratic dominance.
Figure 1. Term ratios along the relaxed Starobinsky trajectory. Left: . Right: the exact inertial-to-linear ratio and its slow-roll asymptotic estimate . The asymptotic estimate agrees closely with the exact inertial ratio on the relaxed slow-roll branch, but should not be extrapolated quantitatively to the end of inflation, where reaches unity. At the exit , so this plot does not show quadratic dominance. denotes the number of Jordan-frame e-folds remaining until the first crossing.
Signed square root and regular Hubble zeroes
On , the substitution removes the quadratic velocity term. Direct substitution into Eq. (4) gives
The positive root is only a local coordinate. Equation (16), however, is smooth at and can be used with a signed . Its two signs describe the same away from a zero. This is a regularizing parametrization, rather than a globally invertible field redefinition.
The change of variable is not merely algebraic convenience. It makes the late-time behavior easier to interpret because its equation has a smooth restoring term and an explicitly damped energy balance. In particular, a zero of is not by itself a singular event in spacetime.
The exact balance
bounds and in forward time for finite initial data. To check what its zero means physically, use the regular trace–Ricci system
At , the constraint requires , but allows . For an expanding tangency,
The cubic Taylor coefficient vanishes because Eq. (18) gives , and hence . The signed solution of Eq. (16) gives , in agreement with Eq. (18). It therefore continues through the tangency smoothly. The positive root alone behaves as and is not differentiable there. If , uniqueness of Eq. (18) gives the Minkowski solution.
At small amplitude, the signed variable has frequency . This is not the scalaron mass. Indeed,
so a leading harmonic produces a curvature oscillation with frequency . Averaging Eq. (17) over a period gives and, at late times,
The averaging applies after the coherent oscillations have entered their small-amplitude regime; the oscillatory term in and its averaged envelope should not be mistaken for separate fluids. This reproduces the known oscillatory vacuum behavior of the quadratic model [5, 8]. It describes coherent scalaron oscillations, not their conversion into a thermal matter sector [12, 13].
In the original Hubble variable one may instead define . Retaining the derivative of gives
The combined nonlinear term has magnitude times the linear term on the decreasing branch. Equations (17) and (22) are different mechanical balances for the same background. Neither defines an invariant fluid energy or an independent energy transfer into particles.
Canonical dynamics and a derivative expansion for general
Frame transformation and physical mass
On a branch with and locally invertible , the exact transformation is
The potential and homogeneous equations are [3, 4, 8]
For example, the transformed acceleration and friction combine to give
The trace equation then yields the scalaron equation in Eq. (25), while the Einstein-frame Friedmann equation follows from the transformed Jordan-frame constraint (2). The canonical friction is obtained from the full transformation, rather than from the Jordan-frame linear velocity term alone.
For the quadratic model,
Since on the expanding Einstein branch, . Its zero occurs at a passage through with positive scalaron velocity, where . This provides an independent check of the tangency in § 2.3. Figure 2 shows the canonical background. The monotone must not be identified with the oscillatory in Eq. (21).
Figure 2. Vacuum Starobinsky evolution in Einstein time. The initial field is ; its initial velocity is set by the leading slow-roll drift with the full Friedmann constraint imposed. Upper panels: the scalaron and the normalized Hubble parameter . Lower panels: the exit of inflation and the coherent late-time scalaron oscillations. In the small-amplitude limit, the frequency of tends to . No particle-production or thermalization equation is included. The oscillations are a vacuum property of the model and serve as a consistency check of the background evolution.
The disappearance of an effective mechanical coefficient at a special Hubble value has no bearing on the mass of the canonical scalaron. The latter is read from the potential, while ghost and tachyon criteria require perturbing the action around a specified background. Physical stability must be assessed from the action and the appropriate perturbation problem. The sign of a mechanical inertial coefficient is not a ghost criterion. We work with and ; these give the positive tensor kinetic coefficient and the conventional locally invertible curvature branch [27]. At a constant-curvature vacuum satisfying , the trace mass in Jordan time is
The canonical Einstein-frame mass there is . Positivity of alone does not prove positivity of this mass, and evaluating Eq. (28) along a rolling background does not give its complete perturbation mass.
Expansion and the first omitted terms
For a general smooth function, the exact Hubble equation is
with and . This formula is exact in vacuum and has no assumption of slow roll. Its dependence on the full is precisely why the simplified, position-only coefficients below require an additional approximation. For , solving this standard constraint for gives the exact Hubble phase-space equation used in the Supplemental Material of Ref. [16]. Its coefficients depend on the velocity through . Coefficients that depend on alone require an expansion.
On an interval with , write and , and adopt the derivative counting
Smoothness alone does not control the expansion. The curvature-derivative ratios and can offset the nominal powers of . Their behavior must be checked as varies from to . Near sharp curvature features, the Taylor remainder must be evaluated directly.
Using and , the second-order equation is
where all curvature derivatives below are evaluated at :
The linear coefficient can be checked without relying on the dictionary:
where and . The result differs from what one would obtain by expanding the potential-like term alone. The time derivative of and the curvature dependence of the coefficient of contribute at the same order. The coefficient is the sum : its pieces come from , the explicit velocity term, and , respectively. The potential expansion also supplies . Retaining only would therefore give the wrong quadratic coefficient.
At the next order,
In Eq. (30), both added terms have third derivative order. For example, the next inertial term is and has fourth order. The equations used for the two truncated trajectories are
where the relevant denominator must remain nonzero. These truncated equations are to be integrated only while the expansion ordering and the denominators remain under control. The residual of on an exact solution and the error of a solution evolved with are different quantities. We compute both in § 5.
Within Eq. (31), and a local coordinate removing its explicit quadratic velocity term satisfies
on , . This transformation removes a term of the chosen truncation, not the remainder or the scalaron’s physical nonlinear dynamics. For , Eq. (31) is exact and Eq. (36) reduces to the square-root parametrization.
A regularized inverse-logarithmic example
Action, derivatives and scalaron potential
Consider
with , , . The coefficient is dimensionless; has dimensions of inverse curvature. We study . The choice returns the quadratic Starobinsky action, so the numerical scan has a direct reference model. The shifted logarithm is positive and smooth for every real . The logarithm is regularized here specifically to permit the solution to cross during the vacuum oscillatory stage without encountering the logarithmic singularity of an unshifted expression. At , its inverse gives a correction that tends to zero, leaving a nonzero quadratic coefficient . At , the quadratic coefficient is instead . These choices define the interpolation used here; they do not establish a quantum origin or a unique ultraviolet completion. The motivation for considering an inverse logarithm is the running of curvature-squared couplings in quantum-gravity calculations and related logarithmically corrected inflationary actions [14, 16]. Our inverse-logarithmic factor is chosen to reproduce a comparable large-curvature functional dependence while keeping a smooth Minkowski branch. The shift inside the logarithm and its continuation through negative curvature are phenomenological choices; we do not derive them from a trace anomaly, a specified matter spectrum or a string effective action.
Introduce
The following functions keep the derivatives compact:
Direct differentiation gives
It is also useful to define
The exact potential and its mass at the Minkowski minimum are
Relation to running quadratic gravity
The large-number-of-matter-fields approximation of Ref. [16] has curvature dependence
where parametrizes the renormalization-group trajectory and is a mass scale. Suppressing the overall action normalization, the quadratic part of Eq. (37) has the same large-curvature form under the formal identification
The relation concerns the curvature functions. Here the Einstein–Hilbert term is present at all curvatures, and the regularized action has a Minkowski minimum. The running couplings, matter-loop content and Weyl-squared sector of Ref. [16] are outside this model. Their strong-coupling restriction on therefore cannot be transferred to the present parameter scan.
The common asymptotic behavior also appears in the potential. At fixed and ,
Indeed, and the leading potential correction is . Slow roll then gives
With Eq. (44), these are the leading spectral scalings of Ref. [16]. They are established asymptotics of this inverse-field plateau. The limit instead gives the exponential Starobinsky plateau; the fixed-positive- large-field limit is not uniform at .
The asymptotic formulas alone need not describe an event 60 e-folds before the exit. We therefore retain the full finite-curvature action in the numerical calculation. At the , , reference event, defined as 60 exact Einstein-frame e-folds before the first crossing (§ 5), the Einstein–Hilbert contribution is about 1.58% of the quadratic part of , while the logarithmic-to-Einstein–Hilbert ratio in is 0.983. Thus a small action fraction can still matter for the inflationary slope. The tests below concern the regular branch and the accuracy of the mechanical approximation in this specified extension.
Positive curvature derivative on the studied slice
Proposition 1. For , and , Eq. (37) satisfies at every real . The function is strictly increasing and has a unique negative zero . On , the scalaron potential is nonnegative and has a Minkowski minimum at .
Proof. Write , where
For , the discriminant of Eq. (47), regarded as a quadratic in , is . Since its leading coefficient is positive, . For , . On this interval, , and . Thus .
Since , is strictly increasing. Moreover, and as , giving one zero . Finally, and imply at every . Equation (42) proves the potential statement on .
The potential diverges as , because . An expanding Einstein-frame solution with finite initial scalaron energy cannot reach that boundary: its energy obeys . This is a branch-regularity statement, not a theorem of perturbative stability on every rolling solution. The bound just proved uses ; it should not be extrapolated to arbitrary positive without another analysis.
Mechanical coefficients and potential slow roll
At , Eq. (32) becomes
Neglecting and relative to in Eq. (31), the leading drift is therefore
The logarithmic deformation is thus visible already in the slow background drift, but its significance cannot be read off from the magnitude of the added term in the action alone. The correction changes both the drift and the potential. Their physical interpretation remains tied to the exact background and canonical scalaron.
We use because this slice admits the analytic lower bound on proved above. The choice supplies a smooth, resolved low-curvature scale for the post-inflationary continuation. Within the scan , the reference value is representative rather than an observational bound or a maximally allowed coupling: it gives an order-unity competition between the logarithmic and Einstein–Hilbert contributions to the inflationary potential slope, as quantified below. These values were chosen to stress-test the mechanical expansion, not by optimizing a cosmological likelihood.
For comparison with the mode calculation below, potential slow roll gives
and , . Its approximate e-fold integral is
The superscript distinguishes this endpoint prescription from the exact background criterion . For , , and , it gives and . At the potential pivot, the logarithmic contribution to is about 1.0069 times its Einstein–Hilbert contribution. It therefore cannot be dropped from the potential slope even when its fractional correction to the action is modest.
Background evolution and approximation errors
We integrate the exact vacuum equations and the truncated Hubble equations from the same reference state. Comparing the solutions at common Jordan times measures the accumulated error. Separate Jordan- and Einstein-frame integrations check the time and frame transformations.
Use dimensionless times , , Hubble parameters , and . The exact Jordan system is
The last equality is imposed on the initial state and monitored as a constraint. In the Einstein frame, put ; then
The matching data are
We use , and for the background comparison. The initial curvature is . With and , we choose the initial scalaron velocity from the leading slow-roll relation, , and determine from the full Friedmann constraint. Explicitly,
The reference event is selected 60 exact Einstein-frame e-folds before the first crossing. The direct Jordan integration starts from the transformed state at that event. Both time origins are reset there.
We use the DOP853 integrator in SciPy [28] with relative and absolute tolerances and and maximum dimensionless time step 0.0125. A comparison run uses , and step 0.025. The two frames are compared at 40 001 common events on . Maximum sampled absolute differences in , and are below ; the maximum absolute Jordan constraint residual is below . Relative residuals are poorly conditioned near simultaneous zeroes of and , so an absolute residual is reported. These checks measure a finite trajectory, rather than a solver-independent error bound.
The exact reference state, defined at the chosen 60-e-fold event, is
The first Einstein-frame acceleration exit occurs at and . The Jordan-frame event occurs 0.179064 Einstein-frame e-fold later. The e-fold relation between any two common events is
Thus the 60- reference in Figure 1 and the 60- reference here are distinct conventions. For the two Starobinsky events used in that figure, the same expansion is .
The vacuum logarithmic-model solution subsequently oscillates about (Figure 3). The sampled minima are and , consistent with § 4.3. The narrower features near the regularization scale are evolved with the exact equations. No slow-roll expansion is used for the post-inflationary trajectory.
Figure 3. Exact vacuum background for , , and . Upper panels: curvature evolution and the Jordan- and Einstein-frame acceleration parameters. The e-fold coordinate increases from the exact Einstein-frame reference event. Lower panels: the subsequent curvature oscillations and the positive functions and . Jordan time is measured from the reference event. This is a vacuum solution; no matter-production equation is included. The two end-of-inflation markers use and and therefore occur at different events.
Two types of comparisons are shown in the plots. The curvature and coefficient histories in Figure 3 come from the untruncated field equations, while the curves in Figure 4 test an approximation against that same exact trajectory. In particular, a small residual of an approximate equation does not ensure a proportionally small accumulated error in its solution. The points near the end of the plotted inflationary interval should not be used to extrapolate either truncated equation into the oscillatory era.
Figure 4. Approximation tests for the logarithmic benchmark. Upper left: the magnitudes of the retained inertial and quadratic terms relative to the linear term, evaluated on the exact background. Upper right: second- and third-order equation residuals divided by the force magnitude. Lower panels: the relative errors in the Hubble parameter and its velocity, respectively, obtained by independently integrating the truncated equations and comparing them with the exact solution at the same Jordan times. The e-fold coordinate is assigned using the exact background. The initial trajectory errors vanish and are clipped at for logarithmic plotting. Term ratios, equation residuals, and integrated trajectory errors are distinct diagnostics.
At the reference event, the second-order mechanical ratios and residual are
The omitted mixed term alone is of there; the cubic term also contributes.
Near the first Einstein-frame exit, the second-order residual reaches about of .
For a quantitative local check, define
on intervals where . This compares each truncated equation with the exact trajectory, rather than with its own numerical solution. A tolerance such as is a chosen one-per-mille accuracy criterion, not a model-independent slow-roll boundary. In the sampled inflationary interval of our benchmark, first exceeds at approximately and remains below through the first Einstein-frame acceleration exit. Where is small or zero, the same residual must instead be normalized by a nonsingular scale formed from the retained terms. These pointwise criteria say nothing on their own about the accumulated solution error, which is tested separately below.
To measure accumulated error, we evolve each equation in Eq. (35) from the same exact , at the reference event. We compare at common Jordan times, labeled by the exact trajectory’s elapsed Einstein e-folds . This avoids comparing different endpoints. The resulting errors are reported in Table 1. Table 1 uses the nearest points of the 40 001-event sample; their
| 30 | ||
| 50 | ||
| 55 | ||
| 59 | ||
| 60 |
Table 1. Accumulated errors of the second- and third-order equations, compared with the exact logarithmic-model trajectory at common Jordan-frame times.
e-fold labels differ from the displayed integers by at most 0.001. The third-order trajectory improves on the second-order one over this interval. The contrast is most informative toward the end of inflation: a small instantaneous residual does not imply that the integrated trajectory error will remain equally small. This example does not establish a uniform error bound for arbitrary models or license continuing the expansion through the sharp post-inflationary features (Figure 4).
Linear perturbations and conditional inflationary predictions
Tensor propagation and scalar fluctuations
We calculate scalar and tensor spectra from the standard gauge-invariant perturbation equations of metric gravity. The action and its constraints determine their kinetic terms and normalization. Appendix A gives the quadratic actions and numerical implementation used for the logarithmic benchmark.
For transverse–traceless tensor perturbations, and the linear scalar-curvature perturbation vanishes, . This does not make tensor propagation identical to general relativity: the evolving background changes the effective Planck mass. With no anisotropic-stress source, the tensor equation is [29, 30, 31]
The propagation speed remains unity; the additional damping is [32, 33]. In the quadratic model,
At high curvature during slow roll, . Along the exact Starobinsky trajectory used in § 2.2, the full damping coefficient divided by decreases from 2.98352 to 2.11494. The approximation cannot be extended to the final crossing (Figure 5).
Figure 5. Diagnostics along the relaxed Starobinsky background. Left: the magnitudes of the two mechanical power terms in the mechanical balance (22), normalized to the initial linear contribution. Right: exact tensor damping coefficient, its high-curvature slow-roll approximation and the constant- reference value. Mechanical power terms are diagnostics of the background equation, not physical reheating rates. The high-curvature approximation to tensor damping fails near the final crossing.
The scalar perturbation is described by the comoving curvature perturbation (distinct from the background Ricci scalar ). We restrict the mode calculation to the rolling inflationary branch, where comoving scalaron slicing is regular. In the Einstein frame the scalar and tensor canonical modes obey
where , primes denote conformal-time derivatives and . These are the established scalaron fluctuation equations, not additional consequences of the mechanical analogy [18, 19, 30].
Mode evolution, normalization and comparison with slow roll
We integrate the canonical scalar and tensor modes on the exact Einstein-frame background, taking as the evolution variable. The reference event is selected e-folds before the first exact crossing. The Bunch–Davies prescription [22, 34] is implemented at a finite, deep-subhorizon ratio ; the value of is varied to test the initialization. Appendix § 10 gives the background and mode equations, their initial conditions and the convergence tests. With , the dimensionless spectra are
Here and are the canonically normalized amplitudes defined in the Appendix § 10. The tensor power includes both helicities and approaches in the usual leading slow-roll limit.
The quoted spectra describe linear vacuum perturbations during inflation. Post-inflationary gravitational-wave production and scalaron decay require additional equations. The reference calculation uses and evaluates the frozen modes at , before the end of inflation. We determine the tilt with a centered five-point derivative in , and test the starting scale, integration tolerances and freeze-out time independently (Appendix § 10). The results in Table 2 are quoted to fewer digits than those numerical tests support.
| (modes) | (modes) | |||
| 0 | 0.96783 | 0.96775 | 0.002964 | 0.003007 |
| 1 | 0.97036 | 0.97027 | 0.003685 | 0.003730 |
| 2 | 0.97236 | 0.97226 | 0.004442 | 0.004489 |
| 3 | 0.97394 | 0.97384 | 0.005222 | 0.005271 |
| 4 | 0.97520 | 0.97511 | 0.006015 | 0.006066 |
| 5 | 0.97622 | 0.97613 | 0.006813 | 0.006865 |
Table 2. Potential slow-roll estimates and linear-mode spectra at the separately defined 60-e-fold reference events for the regularized inverse-logarithmic model.
The values of , and are used as specified test parameters for the regular branch, not as a best-fit point or a bound from CMB data. In particular, the benchmark is not claimed to be the largest observationally permissible departure from Starobinsky inflation. At , and , the results are The potential columns use Eq. (51) at
; the numerical columns use 60 e-folds to the exact acceleration exit. Both the slow-roll approximation and the endpoint prescription therefore enter their difference. This is a comparison of two stated prediction procedures, not an isolated measurement of a higher-order spectral correction. The scalar tilt differs by about and by roughly one percent over the values in Table 2 (Figure 6). For , the finite- calculation retains endpoint and logarithmic corrections that are absent from the large- estimates and . Those leading asymptotic expressions are not reference values at the precision of this table.
Figure 6. Potential shapes and conditional inflationary predictions for , . Upper left: the Einstein-frame potentials using the same overall . Upper right and lower left: the scalar spectral index and the tensor-to-scalar ratio , respectively, comparing potential slow-roll estimates with numerical linear-mode results at their separately defined 60-e-fold reference events. The numerical points are connected only as a visual guide. Lower right: the ratio of the logarithmic to Einstein–Hilbert contributions to the potential slope, evaluated at the potential slow-roll reference event. These are conditional theoretical predictions at specified reference events; they are not the results of a likelihood analysis. The slow-roll and numerical-mode calculations use different end-of-inflation prescriptions; their differences therefore cannot be attributed solely to higher-order slow-roll corrections.
The amplitude fixes the overall scale, conditional on this reference choice. If denotes Eq. (63) evaluated with , then
Taking [23] gives at and at . Since is held fixed, the corresponding dimensional changes when this normalization changes. The leading potential-amplitude expression is useful for comparison, but was not substituted for the numerical amplitude.
The parameters and e-fold choice are not fixed by the background analogy. For example, the same , model gives the following numerical sensitivity checks: Table 3 makes clear that the dependence on is larger than the numerical uncertainty. A data comparison must specify the post-inflationary expansion, match a dimensional pivot, and explore the model parameters. Cheslog et al. [26] show that allowing the effective relativistic energy density and neutrino masses to vary broadens the inferred tilt constraints and changes the apparent tension of Starobinsky inflation. Their likelihood analysis supplies observational context, rather than a test of the present action. The 60-e-fold choice used here is a numerical reference; its identification with a CMB pivot remains conditional on the post-inflationary history.
| 50 | 0.010 | 0.96791 | 0.006985 |
| 55 | 0.010 | 0.97115 | 0.006025 |
| 60 | 0.005 | 0.97338 | 0.004962 |
| 60 | 0.010 | 0.97384 | 0.005271 |
| 60 | 0.020 | 0.97433 | 0.005651 |
Table 3. Sensitivity of the numerical spectra to the e-fold reference and regularization parameter , with and .
What the mechanical ratio says about
The ratios introduced in the chart are useful for comparing terms in a background equation. They do not by themselves determine the scalar or tensor amplitudes. Nevertheless, in the high-curvature Starobinsky limit one may express the familiar leading tensor-to-scalar ratio in terms of a background mechanical ratio, once the frame conversion is performed correctly. On the expanding inflationary branch, where , , and , the exact vacuum background relation is
For high-curvature quadratic slow roll, and . Combining the leading spectrum result with gives
For this comparison, must be evaluated on the Jordan-frame background at the same pivot horizon-crossing event at which the Einstein-frame slow-roll spectrum is evaluated, with the two Hubble histories related by the conformal transformation. It is not a time-independent constant. In particular, is not a frame-invariant observable and the relation cannot replace the mode calculation. This is a rewriting of a standard approximation in a chosen background variable. On the 60- Starobinsky trajectory, it gives 0.0033148, while the same event has . These are not the numerical mode value at the different 60- reference in the table. Equation (66) is not valid for a general model or at the oscillatory stage. The homogeneous mechanical equation also supplies no independent stochastic tensor source.
Conclusion
The Jordan-frame Hubble equation gives a direct test of the slow-roll approximation in metric gravity. For the quadratic Starobinsky model the mechanical equation is exact. Its inertial-to-linear ratio on the relaxed branch is , while the quadratic-to-linear ratio is . The latter is only one sixth at the first Jordan-frame acceleration exit. The exit therefore cannot be identified with quadratic dominance in this normalization.
The comparison of neighboring trajectories is also precise. At a common decreasing Hubble value, the velocity difference contracts on the expanding branch. This removes the arbitrary time offset between histories. The result applies while both solutions remain on their common decreasing-Hubble interval; it does not establish a global attractor basin. At a Hubble zero the original equation becomes a poor coordinate description, although the trace–Ricci system remains regular. The signed variable passes smoothly through the tangency. Its small-amplitude frequency is half the curvature and scalaron frequency, as required by the relation between and .
For a general action, the curvature dependence of the coefficients matters. The second-order mechanical equation contains , and the cubic terms include both and an contribution. Small Hubble-flow parameters must therefore be accompanied by a controlled hierarchy of curvature derivatives. We tested the approximation in two ways: through its residual on the exact background, and through the error of a separately integrated trajectory at common Jordan times. In the regularized logarithmic benchmark the cubic truncation improves the inflationary trajectory over the interval studied. The sharp features during the later oscillations require the untruncated equations.
On the slice , , the example satisfies and has a regular positive- branch with a Minkowski minimum. Independent frame integrations agree within the reported numerical accuracy. The large-curvature action and leading inverse-field spectral scaling are shared with earlier running quadratic-gravity work; the tests here concern the smooth finite-curvature extension and its approximation errors.
The linear modes give and for , , , at the event 60 exact Einstein-frame e-folds before the first acceleration exit. These are conditional theoretical predictions. Their dependence on the e-fold reference is larger than the measured numerical uncertainty. A model for scalaron decay and the subsequent expansion would be needed to match a dimensional pivot, followed by a parameter likelihood analysis for an observational test. The vacuum oscillations studied here do not describe thermal reheating, and the mechanical velocity coefficient is not a scalaron decay rate.
Acknowledgments
This work is supported in part by the National Natural Science Foundation of China (NSFC) Grant Nos. W2533015, 12375052, and 11935009 and by the Tianjin University Self-Innovation Fund for Extreme Basic Research Grant No. 2025XJ21-0007.
Data and code availability
Source code for reproducing the numerical results is provided in the Supplemental Material [35].
Quadratic actions and numerical mode checks
Canonical actions and equations
For with , tracelessness gives at first order. The standard quadratic tensor action is [29, 30]
Variation gives Eq. (60). On an interval where comoving scalaron slicing is regular, the scalar action is
The conformal-time interval is common to the frames. The tensor pump field is proportional to , while . They give the canonical equations (62) used in Sec. 6.
Mode initial conditions and integration
For the numerical integration let , with selected e-folds before the first exact event. Define , , and . We set and . The exact inflationary background equations in this variable are
with . Initial data are the same constraint-consistent scalar drift at used in Sec. 5.
For stable evolution outside the horizon, we integrate the curvature and tensor amplitudes instead of differencing the pump fields numerically. Define and . They obey
The canonical vacuum at a finite subhorizon ratio gives, up to an irrelevant common phase,
This is a finite- approximation to the asymptotic short-wavelength vacuum. Its dependence on is tested, rather than assumed to vanish. The tensor normalization includes both polarizations. The default mode solver uses DOP853 with relative/absolute tolerances and maximum step 0.02 in . A centered five-point derivative at , , estimates . The initial finite- state approximates the asymptotic short-wavelength vacuum; its dependence on is tested below.
Convergence and independent checks
All background integrations for the mode calculation use and the constraint-consistent scalar drift of Eq. (55), with the model-specific value of . The e-fold background (69) is integrated with relative/absolute tolerances and maximum step 0.03. It reproduces the reference state obtained by the cosmic-time evolution. The following checks concern the numerical linear calculation; they do not estimate uncertainties from reheating or the phenomenological action.
At and 3, we repeat the calculation for , 200, 400, 800, halve the wavenumber spacing to 0.01, and tighten mode tolerances by ten while halving the maximum step to 0.01. Representative results are given in Table 4. The , rows use the tighter integration settings; the other rows use the defaults. Over the tested initializations, the starting-scale variations change by less than and by less than in relative terms in these tests. Halving the wavenumber spacing or tightening the integration settings changes by less than . The scalar and tensor canonical Wronskians are conserved to better than from initialization through , where cancellation has not yet made a numerical Wronskian check ill-conditioned.
| 0 | 100 | 0.02 | 0.96774735 | 0.003006557 |
| 0 | 200 | 0.02 | 0.96774733 | 0.003006551 |
| 0 | 400 | 0.01 | 0.96774729 | 0.003006547 |
| 0 | 400 | 0.02 | 0.96774729 | 0.003006547 |
| 0 | 800 | 0.02 | 0.96774732 | 0.003006547 |
| 3 | 100 | 0.02 | 0.97384475 | 0.005271005 |
| 3 | 200 | 0.02 | 0.97384472 | 0.005270998 |
| 3 | 400 | 0.01 | 0.97384465 | 0.005270992 |
| 3 | 400 | 0.02 | 0.97384465 | 0.005270992 |
| 3 | 800 | 0.02 | 0.97384469 | 0.005270993 |
Table 4. Mode-spectrum checks against the initial subhorizon ratio , the wavenumber spacing and the integration tolerances.
Evaluating the scalar amplitude at rather than 10 changes it by less than over all six scan points; the tensor change is of the same order. For and 3, extending the evaluation from to 12 changes both powers by less than . Figure 7 uses the later reference for values displayed through . These effects are below the precision used in the main table.
Figure 7. Scalar mode freeze-out for and at the exact - reference. Left: the power spectrum normalized to its value at . Right: the relative difference from that later reference value after horizon exit, displayed through . The nearly coincident curves show a similar approach to freeze-out in the two models. Tensor freeze-out is assessed separately using the same numerical procedure. This is a diagnostic of temporal convergence, not a new inflationary prediction.
A separate program evolves the canonical variables in Eq. (62), using analytic pump fields, rather than the amplitudes in Eq. (70). With the same initial prescription and evaluation at , the two formulations agree in scalar and tensor power to within in relative terms at and 3; their tilts differ by less than . This check tests the change of evolution variable and its normalization independently of the amplitude equations.
An independent normalization test uses a canonical power-law background with constant , for which is constant. The exact canonical solution, with an irrelevant constant phase suppressed, is
Here is a positive dimensionless mode argument, distinct from cosmic time. The numerical variables use the rescaled mode , which explains the two prefactors. Since at large , the canonical mode approaches up to that constant phase. Starting with this solution and its derivative at , the numerical scalar and tensor powers at agree with the analytic solution within in relative terms. Their ratio recovers to rounding accuracy. This check tests the evolution and normalization independently of the functions and of a slow-roll amplitude formula [22].
The dimensionless action and its first four derivatives were also checked by direct Taylor arithmetic at 75-digit precision for 55 curvature–parameter pairs. The sampled scaled differences are below .
References
- [1]A. A. Starobinsky, “A New Type of Isotropic Cosmological Models Without Singularity,” Phys. Lett. B 91, 99–102 (1980).
- [2]A. Vilenkin, “Classical and quantum cosmology of the Starobinsky inflationary model,” Phys. Rev. D 32, 2511–2521 (1985), doi:10.1103/PhysRevD.32.2511.DOI
- [3]B. Whitt, “Fourth Order Gravity as General Relativity Plus Matter,” Phys. Lett. B 145, 176–178 (1984).DOI
- [4]K. Maeda, “Towards the Einstein-Hilbert Action via Conformal Transformation,” Phys. Rev. D 39, 3159–3162 (1989), doi:10.1103/PhysRevD.39.3159.DOI
- [5]M. B. Mijić, M. S. Morris, and W.-M. Suen, “The R² cosmology: Inflation without a phase transition,” Phys. Rev. D 34, 2934–2946 (1986), doi:10.1103/PhysRevD.34.2934.DOI
- [6]K. Maeda, “Inflation as a transient attractor in R^2 cosmology,” Phys. Rev. D 37, 858 (1988), doi:10.1103/PhysRevD.37.858.DOI
- [7]T. P. Sotiriou and V. Faraoni, “f(R) Theories of Gravity,” Rev. Mod. Phys. 82, 451–497 (2010) [arXiv:0805.1726 [gr-qc]].DOI
- [8]A. De Felice and S. Tsujikawa, “f(R) theories,” Living Rev. Rel. 13, 3 (2010) [arXiv:1002.4928 [gr-qc]].DOI
- [9]K. S. Stelle, “Renormalization of higher derivative quantum gravity,” Phys. Rev. D 16, 953–969 (1977).DOI
- [10]K. S. Stelle, “Classical gravity with higher derivatives,” Gen. Rel. Grav. 9, 353–371 (1978).DOI
- [11]H. Khodabakhshi, M. Farhang, F. Shojai, and H. Lü, “Cosmology with Higher-Derivative Gravities,” Phys. Rev. D 110, L061504 (2024), doi:10.1103/PhysRevD.110.L061504 [arXiv:2405.02879 [gr-qc]].DOI
- [12]L. Kofman, A. Linde, and A. A. Starobinsky, “Reheating after inflation,” Phys. Rev. Lett. 73, 3195–3198 (1994), doi:10.1103/PhysRevLett.73.3195.DOI
- [13]L. Kofman, A. Linde, and A. A. Starobinsky, “Towards the theory of reheating after inflation,” Phys. Rev. D 56, 3258–3295 (1997), doi:10.1103/PhysRevD.56.3258 [arXiv:hep-ph/9704452].DOI
- [14]I. Ben-Dayan, S. Jing, M. Torabian, A. Westphal, and L. Zarate, “R^2 log R quantum corrections and the inflationary observables,” JCAP 09, 005 (2014), doi:10.1088/1475-7516/2014/09/005 [arXiv:1404.7349 [hep-th]].
- [15]M. Rinaldi, G. Cognola, L. Vanzo, and S. Zerbini, “Inflation in scale-invariant theories of gravity,” Phys. Rev. D 91, 123527 (2015) [arXiv:1410.0631 [gr-qc]], doi:10.1103/PhysRevD.91.123527.DOI
- [16]R. Liu, J. Quintin, and N. Afshordi, “Ultraviolet Completion of the Big Bang in Quadratic Gravity,” Phys. Rev. Lett. 136, 111501 (2026), doi:10.1103/6gtx-j455 [arXiv:2510.18733 [hep-th]].DOI
- [17]V. F. Mukhanov and G. V. Chibisov, “Quantum fluctuations and a nonsingular universe,” JETP Lett. 33, 532–535 (1981).
- [18]M. Sasaki, “Large Scale Quantum Fluctuations in the Inflationary Universe,” Prog. Theor. Phys. 76, 1036–1046 (1986), doi:10.1143/PTP.76.1036.DOI
- [19]V. F. Mukhanov, “Quantum theory of gauge-invariant cosmological perturbations,” Sov. Phys. JETP 67, 1297–1302 (1988).
- [20]A. R. Liddle, P. Parsons, and J. D. Barrow, “Formalizing the slow roll approximation in inflation,” Phys. Rev. D 50, 7222–7232 (1994) [arXiv:astro-ph/9408015].DOI
- [21]E. D. Stewart and D. H. Lyth, “A more accurate analytic calculation of the spectrum of cosmological perturbations produced during inflation,” Phys. Lett. B 302, 171–175 (1993) [arXiv:gr-qc/9302019], doi:10.1016/0370-2693(93)90379-V.DOI
- [22]I. J. Grivell and A. R. Liddle, “Accurate determination of inflationary perturbations,” Phys. Rev. D 54, 7191 (1996) [arXiv:astro-ph/9607096], doi:10.1103/PhysRevD.54.7191.DOI
- [23]Planck Collaboration, Y. Akrami et al., “Planck 2018 results. X. Constraints on inflation,” Astron. Astrophys. 641, A10 (2020) [arXiv:1807.06211 [astro-ph.CO]].
- [24]BICEP/Keck Collaboration, P. A. R. Ade et al., “Improved Constraints on Primordial Gravitational Waves using Planck, WMAP, and BICEP/Keck Observations through the 2018 Observing Season,” Phys. Rev. Lett. 127, 151301 (2021) [arXiv:2110.00483 [astro-ph.CO]].
- [25]ACT Collaboration, T. Louis et al., “The Atacama Cosmology Telescope: DR6 Power Spectra, Likelihoods and ΛCDM Parameters,” JCAP 11, 062 (2025), doi:10.1088/1475-7516/2025/11/062 [arXiv:2503.14452 [astro-ph.CO]].
- [26]Z. Cheslog, E. Finson, A. MacInnis, N. Sehgal, N. Afshordi, S. K. Nerval, and R. Hložek, “Current and Future Constraints on the Primordial Power Spectrum,” arXiv:2609.35964 [astro-ph.CO] (2026).arxiv.org/abs/2609.35964
- [27]A. D. Dolgov and M. Kawasaki, “Can modified gravity explain accelerated cosmic expansion?,” Phys. Lett. B 573, 1–4 (2003), doi:10.1016/j.physletb.2003.08.039 [arXiv:astro-ph/0307285].DOI
- [28]P. Virtanen et al., “SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python,” Nat. Methods 17, 261–272 (2020), doi:10.1038/s41592-019-0686-2.DOI
- [29]J. Hwang, “Cosmological perturbations in generalized gravity theories: Solutions,” Phys. Rev. D 42, 2601 (1990), doi:10.1103/PhysRevD.42.2601.DOI
- [30]J. Hwang and H. Noh, “Cosmological perturbations in generalized gravity theories,” Phys. Rev. D 54, 1460–1473 (1996), doi:10.1103/PhysRevD.54.1460.DOI
- [31]V. F. Mukhanov, H. A. Feldman, and R. H. Brandenberger, “Theory of cosmological perturbations,” Phys. Rep. 215, 203–333 (1992), doi:10.1016/0370-1573(92)90044-Z.DOI
- [32]I. D. Saltas, I. Sawicki, L. Amendola, and M. Kunz, “Anisotropic stress as a signature of nonstandard propagation of gravitational waves,” Phys. Rev. Lett. 113, 191101 (2014), doi:10.1103/PhysRevLett.113.191101 [arXiv:1406.7139].DOI
- [33]L. Lombriser and A. Taylor, “Breaking a dark degeneracy with gravitational waves,” JCAP 03, 031 (2016), doi:10.1088/1475-7516/2016/03/031 [arXiv:1509.08458].DOI
- [34]T. S. Bunch and P. C. W. Davies, “Quantum field theory in de Sitter space: Renormalization by point-splitting,” Proc. R. Soc. Lond. A 360, 117–134 (1978), doi:10.1098/rspa.1978.0060.DOI
- [35]See Supplemental Material at [URL will be inserted by publisher] for the numerical source code and instructions for reproducing the results.






