Dynamics of Relaxed Inflation

The cosmological relaxation of the electroweak scale has been proposed as a mechanism to address the hierarchy problem of the Standard Model. A field, the relaxion, rolls down its potential and, in doing so, scans the squared mass parameter of the Higgs, relaxing it to a parametrically small value. In this work, we promote the relaxion to an inflaton. We couple it to Abelian gauge bosons, thereby introducing the necessary dissipation mechanism which slows down the field in the last stages. We describe a novel reheating mechanism, which relies on the gauge-boson production leading to strong electromagnetic fields, and proceeds via the vacuum production of electron-positron pairs through the Schwinger effect. We refer to this mechanism as Schwinger reheating. We discuss the cosmological dynamics of the model and the phenomenological constraints from CMB and other experiments. We find that a cutoff close to the Planck scale may be achieved. In its minimal form, the model does not generate sufficient curvature perturbations and additional ingredients, such as a curvaton field, are needed.


Introduction
During the past few decades, numerous ideas to solve the hierarchy problem have been put forth. The majority of these ideas require the introduction of new symmetries to protect the Higgs mass. Such symmetries lead to the prediction of new degrees of freedom at the electroweak (EW) scale, but none of them has yet been found in experiments. In Ref. [1], a new mechanism was proposed as an alternative solution to the hierarchy problem. The proposal relies on an axion-like field, dubbed relaxion, coupled to the Higgs field in an effective Lagrangian. The relaxion, during its cosmological evolution, scans the Higgs mass, and finally settles on a local minimum where the Higgs has the observed mass, which is parametrically smaller than the cutoff of the effective Lagrangian. The larger the cutoff, the more successful the mechanism in addressing the hierarchy problem. Several works  have elaborated on various aspects of this framework.
In the original proposal [1], the entire scanning takes place in the background of inflation, which provides constant Hubble friction necessary for the relaxion to maintain slow roll and eventually stop at the local minimum. This setup is rather constraining, rendering the cutoff scale significantly below the Planck scale. A natural question to ask is: can some of the restrictions be ameliorated by promoting the relaxion to play the role of the inflaton itself? In this article, we answer this question positively, extending the discussion presented in [30]. One pleasant consequence of this promotion is that we can indeed achieve a higher cutoff compared to the previous proposal. The price to pay is that the dynamics of the relaxion require a friction mechanism that can remain efficient after the reheating process ends, preventing the overshooting of the EW minimum. Here we consider friction from the tachyonic production of gauge bosons, due to the time-dependent background of the rolling relaxion. Not only this slows down the field efficiently, but also provides an interesting and novel mechanism for reheating. A possible issue related to such a friction mechanism is that one risks overproducing the cosmological perturbations. We show there is a region of parameter space where this is avoided.
The relaxion dynamics proceeds in three regimes. The first consists of a long period of inflation, with standard slow roll due to the flatness of the potential. The second corresponds to the last O(20) e-folds of inflation, where slow roll is due to dissipation via gauge-boson production. At this stage, reheating takes place. The third is after reheating, where the relaxion keeps rolling, the friction from gauge-boson production is still present and sufficient to allow for the field to settle on the correct local minimum. Thus, the final stage of relaxation of the EW scale occurs after the end of inflation. This is described with detail in Sections 3 and 4. The last two regimes share similarities with models of axion inflation that have been studied previously in Refs. [31][32][33][34]. The relevance of this type of dissipation, in the context of relaxion models, has been also discussed in Refs. [19,24].
An important aspect of our work is the actual reheating mechanism, which, to the best of our knowledge, has not been explored before. 1 The photons that are produced in the last stages of inflation have a large occupation number and very low momentum. This is a

Axion inflation and photon production
In this section we review some aspects of axion inflation that are relevant to our framework. The inflaton will also play the role of the relaxion in the next section, but for now we are only interested in the dissipation mechanism due to particle production. We couple Abelian gauge fields to the inflaton, whose time evolution leads to the non-perturbative production of gauge field quanta. This production has two important effects: (1) it backreacts on the inflaton and slows it down, (2) it provides a mechanism to reheat the universe at the end of inflation. Once a large number of coherent photons are produced, the reheating process follows through the production of e + e − pairs via the Schwinger mechanism, and the subsequent thermalization of the system. After that happens, it is important to take into account thermal effects in the gauge-boson production. We discuss Schwinger and thermal effects in Section 5 and in Appendix C. Below, we summarize the main aspects of the gauge-field production at zero temperature. The interested reader can find more details in Refs. [31][32][33][34].
We consider a pseudo-scalar inflaton, φ, coupled to an Abelian gauge field, in a Friedmann Robertson Walker (FRW) metric, with t the cosmic time and τ the conformal time. The Lagrangian reads -3 -where F µν = ∂ µ A ν − ∂ ν A µ ,F µν = 1 2 µνσρ F σρ , with A µ the gauge field, and 0123 = 1 √ −g . The potential V (φ) will be specified in the next section. The equation of motion for φ is given byφ where the dot denotes a derivative with respect to cosmic time t and the mean field approximation is used for E · B. The inflaton φ is assumed to dominate the energy density, withφ 2 V (φ), so the Hubble parameter is given by where M Pl is the reduced Planck mass. The equations of motion for the gauge field are more conveniently written using the conformal time τ ≡ t dt /a(t ), which during inflation is τ −(aH) −1 . Note that τ < 0.
Choosing the Coulomb gauge ∇ · A = 0, we have A 0 = 0 and 2 where a prime denotes a derivative with respect to τ . We promote the classical field A(τ, x) to an operator Â (τ, x) and decompose Â into annihilation and creation operators where the helicity vectors ± are such that k · ± = 0 and k × ± = ∓i| k| ± . Then, A ± must satisfy the equation where we have defined ξ ≡ c γφ 2 f H . (2.8) The parameter ξ is convenient because it stays almost constant when the term E · B is the dominant dissipative force in the inflaton dynamics [31]. To set our conventions, we will assume φ rolls from positive to negative [i.e. V (φ) > 0], soφ < 0 and ξ < 0. Furthermore, τ < 0 by definition, and we take c γ > 0. Eq. (2.7) implies that low-momentum (long wavelength) modes of the A − polarization, satisfying develop a tachyonic instability and grow exponentially. This condition can be rewritten as (aH) −1 . (2.10) Here k −1 is the comoving wavelength of the mode A − , while (aH) −1 is the comoving horizon, which shrinks during inflation. We see that, as inflation proceeds, modes with shorter and shorter wavelength become tachyonic. Since typically |ξ| O(10), the comoving wavelength of the exponentially enhanced modes has a typical size comparable to the comoving horizon. Note that only one polarization of the photon experiences exponential enhancement, a consequence of parity violation in the system. The signatures of parity violation in the CMB power spectrum have been discussed in Ref. [39]. Eq. (2.7) can be solved analytically. However, it is more illuminating to use an approximate solution, which can be derived from the WKB approximation, valid as long as Ω /Ω 2 1. The WKB solution for the tachyonic modes holds in the range (8|ξ|) −1 < |kτ | < 2|ξ|, where it can be written as  13) and the photon energy density (2.14) In the last expression, we took A + 0. Using Eq. (2.12), one finds [31]  Incidentally, one can show that E 2 8 7 |ξ| 2 B 2 , and therefore ρ γ is dominated by the electric field contribution.
The evolution of the inflaton φ is dictated by the equation of motion (2.3), with theφ term typically negligible. In Ref. [31], the authors considered the regime where the term -5 -cγ f E · B balances the slope V , meaning that the dissipation mechanism that ensures slow roll is due to gauge-boson production. In such a case, the backreaction of the gauge quanta on the inflaton produces perturbations that are too large, and excluded by CMB observations. On the other hand, in Ref. [32], the authors considered the regime in which the term cγ f E · B is negligible for most of the observable e-folds and slow-roll is solely due to Hubble friction. They showed that even in this case the photon production can leave imprints on the CMB that can be measured.
In the scenario we investigate in this paper, inflation proceeds in the following steps: 1. Initially, the photon production is negligible, |ξ| 1, and φ slow-rolls because of a nearly flat potential. In this regime, |φ| H|φ| and the equation of motion is given by Note that |φ| increases slowly since V (φ) is roughly constant and H decreases as φ rolls down its potential.
2. Eventually, |φ| increases to the point where |ξ| grows large enough for the backreaction of the photons to become important in Eq. (2.3). This is when we enter the second regime described by the equation of motion with E · B given by Eq. (2.15). The approximate solution is where we have neglected a factor of ξ 4 inside the logarithm. We see that ξ is roughly constant in this regime, and we havė (2.21) Unlike the previous regime, |φ| now decreases with decreasing H. The produced photons have an energy density that remains roughly constant 3 Here we have used Eqs. (2.15), (2.16) and (2.19). When the potential of φ drops below the value V (φ) ∼ ρ γ , the photon energy density becomes dominant, and we exit inflation.
The problem with this scenario is that the produced photons have extremely long wavelength and do not thermalize via perturbative scattering processes to reheat the universe. From Eq. (2.10) it follows that their typical physical momentum, q γ = k a , is As we describe in more detail in later section, the relaxation mechanism requires values of H MeV close to the end of inflation, which in turn implies q γ MeV. At the same time these photons have a high occupation number in the Hubble volume ρ γ /q γ H 3 1 due to the large exponential e 2π|ξ| in Eq. (2.16). This system is best described classically as an electromagnetic field. One can show that the photons add up coherently to form a constant electric field within the horizon with magnitude | E| ∼ √ ρ γ (see Appendix D for further discussion). This electric field grows strong enough to allow for vacuum e + e − production via the Schwinger mechanism. This changes dramatically the picture in the second regime described above. We discuss it in detail in Section 5. So far, we have described the generalities of φ playing the role of the inflaton. Our main purpose is to use this inflaton to relax the electroweak scale and, to do so, we need to add the relaxion ingredients, that come next. In the rest of the paper, we explain in detail the whole cosmological evolution of the relaxion/inflaton field.

A relaxed inflation model
The first model we consider consists of an axion field on a very flat potential, V roll , and coupled to SM photons. This pseudoscalar dominates the energy density of the universe during inflation and acts both as the inflaton and as the field that scans the Higgs mass. Additionally, there is a periodic potential V wig that plays a crucial role in setting the VEV of the Higgs after reheating. The effective Lagrangian for our model is given by Here, H is the SM Higgs doublet, φ is the relaxion/inflaton field, and is the φ-dependent squared mass parameter of the Higgs potential. The Higgs bare mass Λ is the cutoff of the effective Lagrangian, g h is a dimensionless parameter of order one, and m Λ. We comment on the parameter Λ wig at the end of this section. We omit terms -7 -with W ± and Z for simplicity. In particular, there are (φ/f )Z µνF µν and (φ/f )Z µνZ µν terms that are gauge invariant but the photon production is not affected by them because their effect only appears from dimension 8 operators suppressed by m 2 Z f 2 after integrating out Z. Similarly, W terms do not affect the photon production.
The relaxion potential here is the same as in Ref. [1] and our model inherits some of the properties of that scenario. These include a trans-Planckian field range for φ, a nearlyflat potential V roll (φ), and the periodic ("wiggle") potential V wig (φ). On the other hand, there are sharp differences that lead to stark contrast with the original proposal. First, in our case the relaxion is the inflaton itself, which allows the energy density of the universe to be of the same order as ∼ Λ 4 . Second, since the dynamics of inflation and relaxation end at almost the same time, the classical rolling is automatically a good description when the electroweak scale is settled. Finally, the relaxion stops after the end of inflation, and therefore we require a form of dissipation distinct from the Hubble friction. The coupling to photons provides this extra source of dissipation, and offers a novel opportunity for reheating.
The smallness of m is justified by the fact that as m → 0, the model possesses the discrete shift symmetry φ → φ+2πnf . A potential of this kind was first used by Abbott [40] in an attempt to explain dynamically the smallness of the cosmological constant. Here, following [1], we use it for the EW scale instead. As written, the model poses some theoretical issues [5] that can be circumvented with a clockwork axion model [9,10,41], which we present in Appendix A, where we also show how to map its parameters to the ones used in this section and in the rest of the paper.
For the given coordinate, a special point in field space is where µ 2 (φ 0 ) = 0. It separates the unbroken EW phase, φ > φ 0 , from the broken phase φ < φ 0 . For field values of order φ 0 , the small m expansion in Eq. (3.3) is not well defined, as each term is of order ∼ Λ 4 and generically order one corrections are expected. In what follows, we will only keep the term linear in m and φ, a choice that is justified only once we consider a UV completion of this model, such as the clockwork axion discussed in Appendix A.
As most of the interesting dynamics happens near φ 0 , it is convenient to expand the potential around this point. We define 4 We keep only the φ linear term in the potential (3.3) and expand around φ 0 . The potential then reads 8) 4 We stress that δφ is still a classical field, not a quantum fluctuation.
where h is the radial mode of H and µ 2 (δφ) = g h mδφ. We choose V 0 such that the cosmological constant has the observed value V obs cc ∼ meV 4 once h and φ settle to their VEVs: (3.10) We have then The contribution of V obs cc to V 0 is a lot smaller compared to the other two terms. In what follows we take V obs cc to be effectively zero. The parameter Λ wig can be written generically as with n > 0 and M some fixed mass scale. The fact that Λ 4 wig depends on the Higgs VEV, v, is crucial: as v grows, the amplitude of the wiggles becomes larger and larger up to the point where they stop the rolling of δφ. This must happen when v attains the observed value of 246 GeV. The case of the QCD axion corresponds to n = 1, y ∼ y u (the lightest quark yukawa), and M ∼ f π (the pion decay constant) 5 . This case is excluded [1] because it results in θ QCD ∼ 1 and thus is plagued by the strong CP problem. In the n = 2 case, the sector responsible for generating V wig does not break the electroweak symmetry, we have a two-loop wiggle-potential [2,5] also in the unbroken electroweak phase, and the relaxation mechanism works, provided that M < v.

Dynamics
In this section, we discuss the cosmological evolution of the fields δφ, h, and A µ . For the purpose of our study, we can treat δφ as a homogeneous classical field, but we must treat h and A µ as quantum fields. The equations of motion are Here, an overdot denotes a derivative with respect to cosmic time t, and H ≡ȧ a is the Hubble parameter, with a the scale factor. Since inflation is driven by δφ and the energy 5 In this case, Λ 4 wig ∼ M 3 (mδφ) 1/2 , there is a singular term in the first derivative of the potential, ∂Λ 4 wig /∂δφ, at δφ = 0. The singularity is evaded thanks to the quark condensate, qLqR , which provides a tadpole for the Higgs potential and results in a small, but non zero VEV even for µ 2 > 0.  5 is the -dependent squared mass parameter of the Higgs potential. The Higgs bare mass ⇤ is the cuto↵ for our model, we assume m ⌧ ⇤, and g h is dimensionless parameter. The goal is to generate dynamically a small µ 2 , that is |µ 2 | ⇠ m 2 W ⌧ ⇤ 2 , where we took the W mass, m W , as representative of the electroweak scale The larger the hierarchy between m W and ⇤, the more successful is this mechanism i addressing the SM hierarchy problem.
The relaxion potential here is the same as the one used in Ref. [5], and the smallnes of m is justified by the fact that as m ! 0 the model possesses the discrete shift symmetr ! + 2⇡ k f. As written the model poses some theoretical issues [6], that can b circumvented with a clockwork axion model [7]. We present the clockwork model in th Appendix and we show there how to map its parameters to the ones we use in this sectio and in the rest of the paper.
A special point in field space is where µ 2 = 0. It separates the unbroken EW phase, > 0 , from the broken phase < 0 As most of the interesting dynamics happen near this point, it is convenient to expand th potential around 0 . We define 1 We write the Higgs doublet as where h is the real degree of freedom that eventually gets a vacuum expectation valu (VEV) equal to v = 246 GeV, while h a (a = 1, 2, 3) are the three goldstone modes that ge eaten by the W and Z gauge bosons. ⌧ a are the Pauli matrices. 1 We stress that is still a classical field, not a quantum fluctuation. Excuse the notation Jack. with µ 2 ( ) = g h m . We choose V 0 such that the cosmologica and settle to their VEVs: Choosing this V 0 corresponds to tuning the cosmological constan success of the mechanism we describe in the rest of the paper. For future convenience we rewrite the potential as The parameter ⇤ 0 can be written generically as with n > 0 and M some fixed mass scale. The fact that ⇤ 4 0 dep v, is crucial: as v grows the amplitude of the wiggles becomes la point where they stop the rolling of . This must happen whe value of 246 GeV. The case of the QCD axion corresponds to n quark yukawa), and M ⇠ f ⇡ (the pion decay constant). This cas it results in ✓ QCD ⇠ 1 and thus is plagued by the strong CP pro the sector responsible for generating V wig does not break the ele we have a two-loop wiggle-potential [6,8] also in the unbroken e case the relaxation mechanism works provided that M < 4⇡v.
We choose V 0 such that the cosmological constant is zero once h and settle to their VEVs: (3.12) We have Choosing this V 0 corresponds to tuning the cosmological constant. This is crucial for the success of the mechanism we describe in the rest of the paper. For future convenience we rewrite the potential as The parameter ⇤ 0 can be written generically as with n > 0 and M some fixed mass scale. The fact that ⇤ 4 0 depends on the Higgs VEV, v, is crucial: as v grows the amplitude of the wiggles becomes larger and larger up to the point where they stop the rolling of . This must happen when v attains the observed value of 246 GeV. The case of the QCD axion corresponds to n = 1, y ⇠ y u (the lightest quark yukawa), and M ⇠ f ⇡ (the pion decay constant). This case is excluded [5] because it results in ✓ QCD ⇠ 1 and thus is plagued by the strong CP problem. In the n = 2 case the sector responsible for generating V wig does not break the electroweak symmetry, and we have a two-loop wiggle-potential [6,8] also in the unbroken electroweak phase. In this case the relaxation mechanism works provided that M < 4⇡v.
-5 -This is a regime of warm inflation, with constant photon temperature ⇠ ⇤ 0 . Once we reheat and the universe becomes radiation dominated. Before reheating the slow-roll conditions are This is satisfied for The first term is small due to (4.34), the second is small for that is until we exit inflation and reheat.
In our model V 00 ⇠ 0. Then Using the definition of ✏ above it is easy to very that this condition is again satisfied given (4.34).
we reheat and the universe becomes radiation dominated. Before reheating the slow-roll conditions are This is satisfied for The first term is small due to (4.25), the second is small for that is until we exit inflation and reheat.
ligible. The motion is described by (2.16) and (2.17), with a velocity the field rolls. We stay in this regime for a very large number of efolds he way into the broken EW phase ( < 0). When the parameter |⇠|, ows larger than one, we smoothly switch into the second regime, which is ton production and described by (2.18) and (2.19). We enter this regime y very close to the end of its run, EW , and we stay only for the last the inflaton potential reaches ⇠ ⇤ 4 0 the energy density of the produced dominant and we reheat to a temperature just slightly above ⇤ 0 . What of radiation domination in which the wiggles at first disappear and the ling. As the universe cools down, the wiggles reappear and the relaxion e point in which it sets the correct electroweak scale. bsections we are going to give the quantitative details of this picture.
on the slope m rolling the relaxion/inflaton must stop on the wiggles. This implies two slope.
ve at least one wiggle between = 0 and EW : long as the kinetic is small enough, when the condition |V 0 roll ( )| ⇠ satisfied, that is We keep only the linear term in (3.3), expand around 0 and abso in V 0 . The potential then reads We choose V 0 such that the cosmological c and settle to their VEVs: We have Choosing this V 0 corresponds to tuning the cosmological constant. success of the mechanism we describe in the rest of the paper. For future convenience we rewrite the potential as The parameter ⇤ 0 can be written generically as with n > 0 and M some fixed mass scale. The fact that ⇤ 4 0 depend v, is crucial: as v grows the amplitude of the wiggles becomes large point where they stop the rolling of . This must happen when v value of 246 GeV. The case of the QCD axion corresponds to n = 1 quark yukawa), and M ⇠ f ⇡ (the pion decay constant). This case i it results in ✓ QCD ⇠ 1 and thus is plagued by the strong CP proble the sector responsible for generating V wig does not break the electr we have a two-loop wiggle-potential [6,8] also in the unbroken elect case the relaxation mechanism works provided that M < 4⇡v.
-5 -A qualitative overview of the dynamics, shown in the cartoon of Fig. , goes as follows. The inflaton field starts rolling from a point i > 0, in the unbroken electroweak phase. In the first regime the slow roll is due to the smallness of the slope m and the photon production is negligible. The motion is described by (2.16) and (2.17), with a velocity that increases as the field rolls. We stay in this regime for a very large number of efolds (N > 10 30 ), all the way into the broken EW phase ( < 0). When the parameter |⇠|, defined in (2.7), grows larger than one, we smoothly switch into the second regime, which is dominated by photon production and described by (2.18) and (2.19). We enter this regime when is already very close to the end of its run, EW , and we stay only for the last few e-folds. When the inflaton potential reaches ⇠ ⇤ 4 0 the energy density of the produced photons becomes dominant and we reheat to a temperature just slightly above ⇤ 0 . What follows is a period of radiation domination in which the wiggles at first disappear and the relaxion keeps rolling. As the universe cools down, the wiggles reappear and the relaxion finally stops at the point in which it sets the correct electroweak scale.
In the next subsections we are going to give the quantitative details of this picture.

Conditions on the slope m
At the end of the rolling the relaxion/inflaton must stop on the wiggles. This implies two conditions on the slope.
1. We must have at least one wiggle between = 0 and EW : We stop, as long as the kinetic is small enough, when the condition |V 0 r e g im e 1 As most of the interesting dynamics happens near 0 , it is convenient to expand the potential around this point. We define 3 We keep only the linear term in the potential (3.3) and expand around 0 . The potential then reads where h is the radial mode of H and µ 2 ( ) = g h m . We choose V 0 such that the cosmological constant is to be the observed value V obs cc ⇠ meV 4 once h and settle to their VEVs: We have then [TV: Changed V 0 . Check! Agree?] Choosing this V 0 corresponds to tuning the cosmological constant. This is crucial for the success of the mechanism we describe in the rest of the paper. We discuss this tuning further in Section 9. The contribution of V obs cc to V 0 is a lot smaller compared to the other two terms. In what follows we take V obs cc to be e↵ectively zero. 3 We stress that is still a classical field, not a quantum fluctuation.
-8 - Figure 1. Sketch of the different stages in our relaxation mechanism. The first (black) is the standard slow-roll regime, as described in subsection 4.2. In the second (blue), the photons are responsible for the dissipation in the inflaton/relaxion dynamics, which is discussed in subsection 4.3. Finally, the last stage of relaxation occurs after reheating (red), with the photons still providing dissipation and allowing the relaxion to get trapped in the wiggle potential (see subsection 4.4).
density of the universe is dominated by V (h, δφ), the Friedmann equation yields The qualitative overview of the dynamics is similar to that described in Sec. 2 and is illustrated in the cartoon of Fig. 1. The inflaton field starts rolling from a point δφ i > 0, in the unbroken electroweak phase. In the first regime, the slow roll is due to the smallness of the slope m and the photon production is negligible. The motion is described by δφ −V (φ)/3H, with a speed |δφ| that increases as the field rolls. We stay in this regime for a very large number of efolds (N > 10 30 ), all the way into the broken EW phase (δφ < 0). Eventually, the parameter |ξ| = |c γ δφ/2f H| grows larger than one and we smoothly switch into the second regime, which is dominated by photon production and described by V c γ E · B /f . We enter this regime when δφ is already very close to the end of its run, δφ EW , and we remain there only for the last O(20) e-folds. When the inflaton potential reaches ∼ |ξ|Λ 4 wig /c γ , with ξ roughly constant, the energy density of the produced photons becomes dominant and we exit inflation. What follows is a period of radiation domination in which the relaxion keeps slowing down due to photon dissipation, until it stops on the growing wiggles to set the observed electroweak scale.
-10 -We now give some quantitative details of each stage in this simplified picture. In this section we neglect thermal effects, with the aim of keeping the discussion clearer. As we will see, these effects have significant implications which require a careful treatment, that we postpone to Section 5.

Conditions on the slope m
At the end of the rolling, the relaxion must stop on the wiggles. This implies two conditions on the parameters of the model: 1. At least one wiggle between δφ = 0 and δφ EW must exist, 2. Assuming significant dissipation, the inflaton must halt when the condition |V roll (δφ)| |V wig (δφ)| is satisfied. Consequently, the relation is implied. As we mention below Eq. (3.12), Λ wig can never exceed m W , thus For Λ m W , this bound is stronger than (4.5).

Regime 1: slow-roll on a gentle slope
We assume that the rolling starts from δφ i > 0. In this first regime δφ is small, |ξ| 1, and we can ignore the photon production, dropping the term We can also safely drop the last term: for δφ > 0, h 2 is zero, while for δφ < 0 it never grows larger than m 2 W , which is much smaller than Λ 2 . During the slow-roll we also haveφ 3Hφ and therefore .
As the field rolls down the potential, H decreases so |δφ| increases. We can introduce as usual the slow-roll parameters One slow-roll condition, (δφ) < 1, is satisfied so long as The second condition, |η(δφ)| < 1, is also satisfied when (δφ) < 1, as we show in Appendix B.
Once δφ crosses 0, an important phenomenon happens: the Higgs field experiences an instability, known as tachyonic or spinodal instability [42][43][44][45], that eventually results in the spontaneous breaking of the EW symmetry. The instability develops fast and drives the field to the minimum of its mexican hat potential, while δφ has not moved much from δφ = 0. From that point the dynamics of the Higgs are well captured by the evolution of its zero mode, which oscillates around the minimum. Meanwhile, the minimum grows deeper, as δφ rolls to more negative values. The energy density associated with Higgs oscillations grows at the expense of the relaxion energy density. One might wonder if in the end we store enough energy in the Higgs to allow for reheating via its decays into SM particles. The answer is negative: the relaxion dissipates most of its energy via Hubble friction, and at the end of the run the energy density of the Higgs is still several orders of magnitude smaller than Λ 4 wig , insufficient to reheat above the BBN temperature. Another consequence of the instability at δφ = 0 is that the exponential production of tachyonic modes of the Higgs field happens at the expense of the relaxion kinetic energy, and provides another source of friction for the relaxion. This friction is active for a very short time because, as we mentioned above, the Higgs is quickly driven to the minimum of its potential, at which point the tachyonic production, and therefore the friction, switches off. The energy dissipated by the relaxion via this mechanism is absolutely negligible compared to the potential energy available at that point, that is ∼ m 2 W Λ 2 , so it does not affect the dynamics.
The slow-roll motion described by Eq. (4.9) continues into the broken EW phase, Hf grows larger than one. At that point photon production becomes important. Neglecting thermal effects, we enter a second regime of slow-roll, where the dissipation is provided by photon production rather than Hubble friction. We describe the associated dynamics next.

Regime 2: slow-roll via photon production
We switch from the first to the second regime of inflation when |ξ| becomes larger than one and the increasing |δφ| V roll /3H from Eq. (4.9) matches the |δφ| derived assuming the photon-driven friction, Eq. (2.21), This happens when the potential is The inequality is dictated by the condition (4.17) below, and implies that we switch to the second regime while we are still slow-rolling from the first ( < 1).
In the second regime, the dissipation from photon production is important and the equation of motion, Eq. (4.1), becomes The dependence on ξ is largely through ln[V 2 (δφ)], and therefore ξ varies little from the beginning to the exit of the second regime. To be more accurate, we find this value (≡ ξ 2 ) by using the potential (4.14) in Eq.

Inflation exit and relaxation
From Eqs. (4.16) and (2.16), we obtain that the energy density of the produced photons is where we have used Eq. (4.6) for the last expression. The photon energy density remains roughly constant (up to a logarithmic variation of |ξ|) as the result of the approximate balance between the exponential production of photons and the Hubble dilution of this radiation. Once the potential of the inflaton drops below the value the energy density is no longer dominated by φ, we exit inflation and enter a radiation dominated universe. However, the photons have very low momentum and are not thermalized, hence we cannot talk about a reheat temperature yet. We address the reheating mechanism in the next section. The motion of δφ is still described by Eq. (4.16), so |δφ| keeps decreasing as the relaxion rolls. When the increasing amplitude of the wiggle potential reaches Λ 4 wig with the correct value of the EW VEV, the slope of the wiggles counterbalances the linear slope of V roll (δφ) and the relaxion stops at From the end of inflation to this point, δφ has changed approximately Given that δφ ξ 2 Hf /c γ , this implies that about one Hubble time has elapsed and the energy density ρ γ has only changed by an order-one amount. Note this is an important difference with respect to the initial proposal of Ref. [1]. In that work, the relaxation of the EW scale occurs during inflation, while in ours φ settles down after the end of inflation. For this reason the friction provided by gaugeboson production is crucial in this last stage. Without it, the kinetic energy 1 2 δφ 2 would inevitably grow larger than Λ 4 wig and the relaxion would overshoot the EW minimum, causing the whole mechanism to fail.

Schwinger reheating
The picture described in the previous section is good for a successful dynamical relaxation of the EW scale, but fails to reheat the universe. Each produced photon carries very little energy and the system cannot be thermalized via perturbative scattering processes. The large occupation number of the photons implies that they form a classical electromagnetic field, as we explained in Section 2. In order to discuss thermalization in this case, we have to take into account an important non-perturbative phenomenon: the Schwinger effect. We discuss it in this section and proceed to point out a problem that arises when trying to reheat via SM photons. In the next section we propose a resolution with a dark photon.
Quantum electrodynamics predicts that a strong electric field, e| E| m 2 e , can create electron-positron pairs, provided that the characteristic wavelength of the photons is larger than the Compton wavelength of the electron m −1 e . The virtual pairs, produced in the vacuum polarization of the photon, can be accelerated apart and become real asymptotic states if they can borrow enough energy from the electric field itself. This is known as the Schwinger effect [36,37]. In the presence of a constant electric field, the number of pairs produced per unit volume per unit time is [46] where k is the electron (or positron) momentum, and k ⊥ is the component orthogonal to E.
In axion inflation scenarios, like ours, one typically has very strong electric fields e| E| m 2 e . So, in a Hubble time, a large number of pairs per unit volume ∼ (e| E|) 2 /4π 3 H is produced. In the model we consider, close to the end of the first regime, with 1 < |ξ| 10, the electric field grows exponentially and reaches e E ∼ πm 2 e , prompting the pair production 6 . These electrons and positrons inherit an energy of order (e| E|) 1/2 , so the energy density transferred to the e + e − pairs per unit time via the Schwinger effect is roughly (e| E|) 5/2 . This is a very efficient process: an order one fraction of the electric field energy density is transferred to e + e − . The thermalization of the produced e + e − pairs proceeds via annihilations, e + e − → γγ, and inverse Compton scatterings on the long-wavelength photons, eγ → eγ. The rate of such processes is faster than the Hubble expansion. Consequently, the electrons and positrons thermalize very fast and the temperature quickly reaches T ∼ m e .
The finite temperature changes the dispersion relation of the photon, due to in-medium effects, and the tachyonic instability is suppressed, especially when the Debye mass, m D = eT / √ 6, is larger than the characteristic momentum of the instability, m D k a ∼ |ξ|H. Accounting for these thermal effects, we arrive at different expressions for the electric and magnetic fields (see Appendix C for details), Here m D H, and thus a big suppression of order H 4 /m 4 D in the exponent is present when compared to the zero temperature case. This tells us that the intensity of the electric field cannot go much above | E| ∼ m 2 e /e because, once this threshold is crossed, the temperature reaches T ∼ m e through the Schwinger effect, and thermal effect suppress the photon production. On top of that, since the size of the backreaction E · B is correlated with E 2 , the photon friction does not grow enough, unless |ξ| reaches the very large value (m D /H) 2/3 . Now we have two issues: (1) because of the suppressed backreaction, the relaxion does not slow down enough and does not stop on the wiggles (its kinetic energy at the end is larger than the height of the barriers, 1 2 δφ 2 > Λ 4 wig ), (2) the reheat temperature would be of order m e , which is below BBN temperature. One way to fix both problems is to introduce a dark photon, as we describe in the next section.

A model with a dark photon
We have seen that the scenario where the relaxion couples to the SM photon is not viable due to thermal effects. In this section we show that by coupling, instead, the relaxion to a dark photon, we can avoid those issues and successfully achieve relaxation of the EW scale and reheating. We consider the following Lagrangian where the index D denotes the massless dark photon. Here, ψ e is the visible electron, and we assume there is no light matter content in the dark sector besides the dark photon. The field redefinition A µ → A µ − κA D µ removes the kinetic mixing and introduces a coupling of the dark photon to the visible electrons, eκA D µψ e γ µ ψ e . Note that the coupling of the dark photon to φ distinguishes it from the visible photon. Since during the cosmic evolution only dark photons are produced in the time-dependent φ background, our choice of shifting only the visible photon in order to remove the mixing proves convenient.
The relevance of the photons being dark clarifies when describing the end of inflation and reheating. They are produced in the same fashion as described in the first part of the paper, and give rise to a constant dark electric field | E D | ∼ √ ρ γ D . The equations derived in Sections 2 and 4 can be used for this model simply with the replacements: c γ → c γ D , e → κe. In particular, because the coupling to electrons is suppressed by κ, the Schwinger production rate is now To avoid the complication we encountered with the suppressed tachyonic production of visible photons, we wish to ensure that there is no thermal mass associated with the dark photon. To do so, we require the dark photon to be sufficiently weakly coupled as to stay out of thermal equilibrium. Equivalently, the dark photon's mean free path, m.f.p. , must be larger than the Hubble radius, and therefore it cannot be refracted. Such a condition reads and needs to hold until the relaxion settles down. This is satisfied as long as In Eq. (6.5), we took the electrons to be relativistic and in thermal equilibrium at a temperature T > m e , so that their number density n e scales as T 3 . We considered the cross section σ eγ D →eγ ∼ κ 2 α 2 T 2 , rather than σ eγ D →eγ D ∼ κ 4 α 2 T 2 , since the latter is suppressed by two extra powers of κ. Also, we took H ∼ T 2 M Pl and used the reheating temperature T ∼ (|ξ 2 |/c γ D ) 1/4 Λ wig , since it changes only by an order-one amount between reheating and the end of relaxation, as explained in Sec. 4.4 . Note that with these choices the bound (6.6) is conservative.
The absence of a thermal mass for the dark photons implies that we keep producing efficiently the dark electric field as we enter the second regime of slow roll for the relaxion, described in Section 4, where the main friction force arises from dark photon production. We saw that in this regime the amount of energy available in the dark electric field is The energy transfer from the dark electric field to e + e − directly by the Schwinger effect is inefficient, unlike in the SM photon case. In a Hubble time this can be estimated as where the typical e ± energy is (κe| E D |) 1/2 , and we used Eq. (6.6) in the inequality. However, the electric field can transfer an amount (κe| E D |)d of energy to each electron, by accelerating it over a distance d. Shortly after Schwinger creation, the number density of electrons is n e = Thus, the energy density transferred can be estimated as where we took H 2 ∼ ρ γ D /M 2 Pl ∼ Λ 4 wig /M 2 Pl , with Λ wig = 100 GeV, and we again used Eq. (6.6). This is very efficient, provided that 10) and implies that an order one fraction of ρ γ D can be quickly transferred to the SM radiation, so the reheating temperature can reach One can show that for values of the kinetic mixing bounded by Eqs. (6.4) and (6.6), the dark photons never reach thermal equilibrium with the visible sector, after reheating of the latter, and remain cold. So far we have assumed a massless dark photon to maximize its production via the relaxion. However, one can give it a small mass. Its mass would have to be small enough in order not to suppress significantly its production, otherwise the relaxation mechanism could be spoiled.
-17 -Our model is similar to those of natural inflation, where the axion field couples to Abelian gauge bosons. The associated cosmological perturbations have been largely investigated in the literature [31-34, 38, 47-49]. The coupling φFF leads to several features, which include the generation of curvature perturbations and nongaussianities, the production of gravitational waves, and the formation of primordial black holes (PBH). See Ref. [34] for a review of these topics.
In most models of natural inflation the Hubble scale is of order 10 13 GeV, and the number of e-folds is roughly 60. The important difference in our model is that the potential is much shallower. At the beginning we can also have H ∼ 10 13 GeV, but then inflation proceeds for more than 10 30 e-folds and most of the potential energy initially stored in the scalar field is dissipated. The energy density V * available close to the end of inflation is of order Λ 4 wig < m 4 W . For this reason, our model should be regarded as a low-scale inflation model. The number of observable e-folds is given by [50] where V * is the energy density when the mode k left the horizon, V end the energy density at the end of inflation, ρ RH the energy density at reheating, and the subscript 0 refers to today's value. In our case we have V * ∼ V end ∼ ρ RH ∼ Λ 4 wig . Taking the highest value for Λ wig , that is Λ wig ∼ m W , we have ln 10 16 GeV are roughly zero. Therefore the observable number of observable e-folds in our model is N (k) 30.
We have two sources for curvature perturbations: one is from vacuum quantum fluctuations δϕ of the inflaton, proportional to H, the other is from fluctuations induced by the inverse decay of photons [32], δA + δA → δϕ. The first one gives a power spectrum which, as we show in Appendix E, is largely insufficient to explain the observed perturbations: The smallness of P here is a consequence of low-scale H, combined with a very shallow potential. Including the second contribution in regime 1, we have [32] P ζ (k) = P k k 0 where k 0 = 0.002 Mpc −1 and f 2 (ξ) 10 −4 /ξ 6 . The second equality in (7.3) holds for large |ξ|, and for the sake of the estimate we took n s 1. As |ξ| increases, the power spectrum (7.3) increases exponentially. When we enter regime 2, ξ remains quasi-constant with value ξ 2 [see (4.19)], and the power spectrum saturates to [31,33] P ζ (k) 1 (2πξ 2 ) 2 ∼ 5 × 10 −5 P COBE (regime 2) (7.4)  [33,49]. Such a bound is typically very stringent for natural inflation models, but here it is easily evaded due to the significantly lower inflation scale. Note that while it is numerically possible to explain the observed P ζ = P COBE around N e = 30, we see that the lines fall very steeply, due to the exponent in Eq. (7.3). This indicates that it is difficult for the model as it stands to predict the observed curvature perturbations.
This allows, in principle, to have a period around 30 e-folds from the end of inflation where we are still in regime 1, but with a large |ξ| such that, thanks to the exponential in (7.3), we match the observed power spectrum for curvature perturbations, P ζ (k) ∼ P COBE . In practice, it still does not mean that this model is agreement with CMB observations, unfortunately. The same exponential implies, as we see in Fig. 2, that P ζ changes by many orders of magnitude within a couple of e-folds, which is in contradiction with CMB measurements of higher multipoles. Therefore, we need to roughly have less than 25 e-folds in regime 2 to comply with CMB bounds, the consequence being that we do not produce the observed amount of curvature perturbations in the model as it stands. We note that the addition of another field, like a curvaton, can help in matching the CMB power spectrum.
Our current estimate does not take into account the modulation effects due to the wiggles [51]. Naively they are negligible, because P is so small, but a conclusive statement requires a dedicated study, beyond the scope of this paper. We leave a more detailed study of the parameter space in relation to CMB constraints, and a possible extension of this model, to future study.

Constraints and relevant scales
We are now in the position of summarizing the constraints on the model with the dark photon. A summary plot is given in Figure 3. The independent parameters in our construction are m, Λ, g h , Λ wig , f, c γ D , κ.
The first 3 parameters are related to the shallow rolling potential, Λ wig and f are related to the wiggle potential, c γ D and κ to the hidden photon coupling to φ and the visible sector respectively. For the sake of simplicity, we take g h = O(1). Since φ is the Goldstone of a global symmetry spontaneously broken at the scale f , we must impose that the scale Λ, which explicitly breaks the symmetry, be smaller than f , This implies a lower bound on m from Eq. (4.6). We also require the presence of many wiggles between δφ = 0 and δφ EW , that is we impose |δφ EW | > f . This implies an upper bound on m. The two conditions together give the window with Λ wig < m W , as discussed at the end of Section 3. The combination f /c γ D is constrained to the window The upper bound comes from the requirement that we enter the photon-dominated slowroll regime, while the lower bound comes from asking that such a regime does not last more than the last 25 e-folds, see Fig. 2. The number of e-folds in regime 2 is with V switch given by Eq. (4.14) and V RH by Eq. (4.21). Requiring this to last for 25 e-folds at most, and treating ξ as a constant, gives from which we obtain the lower bound of (8.4). Incidentally, in this window we have Pl |ξ 2 | 2 ∼ O(10), which confirms, following Eq. (4.18), that ξ varies very little during this regime. We stress that since c γ D is a free parameter, this rather narrow window leaves a significant viable parameter space, nonetheless.
The goal of the whole mechanism is to achieve a cutoff Λ as large as possible. As Λ f , the cutoff is only limited by the upper bound on f from Eq. (8.4). We have seen -20 -that |ξ| varies only logarithmically in the short photon-dominated regime, and its value is typically |ξ 2 | ∼ O (20). To increase the allowed value of f one would like a value of c γ D as big as possible. Large values of c γ D can possibly be achieved in the clockwork framework, see Appendix A, but for now we restrict our attention to the case c γ D < 10. Note that once we fix c γ D we get directly an upper bound on f and on the cutoff Λ, independently of the other parameters of the model.
Finally, as discussed in Section 6, we need to ensure that the dark photons create e + e − pairs, Eq. (6.4), while not acquiring thermal mass, Eq. (6.6). We also impose that the dark electric field transfers sufficient energy to the e + e − , Eq. (6.10). Together, these requirements contrain κe to the window, Any value of κ in this range will be equally good for reheating. At the same time, they yield a lower bound on Λ wig , The reheating temperature we get is (see Eq. (6.7)) We provide a benchmark point to give an idea of the scales and numbers involved. First, we fix f /c γ D = 0.2M Pl /|ξ 2 |, Λ = 0.1f (which means m ∼ 100Λ 4 wig /f 3 ), and take for instance, Λ wig = 1 GeV and f = 10 11 GeV. Thus, we have More generically, we show in Fig. 3 the allowed (white) region on the plane f vs Λ wig . Note that for Λ wig ∼ 10 GeV, values of f very close to M Pl are allowed, which in turn can accommodate a cutoff as high as 10 16 GeV. The relaxion mass and its mixing angle with the Higgs are given by Here the contributions from V roll are negligible: they are small because they break the discrete shift symmetry. For high values of f , say above 10 10 GeV, the relaxion mass is smaller than 1 keV, and its couplings to matter, suppressed by θ 10 −8 , are tiny. In this range it is hard to detect it experimentally as a particle. However, via its mixing with the Higgs, it can be the mediator of a long-range force. Experimental tests for fifth force 5th force astrophysics cosmology beam dump and collider ). The blue region is excluded by 5th force constraints, while the orange region corresponds to a set of constraints from astrophysics, cosmology, beam dump experiments, and colliders; these are explained in detail in [22]. The dark gray area corresponds to values of Λ wig > m W and is excluded as it implies an unacceptable electroweak breaking scale. The light gray region is defined by the bound Eq. (8.8), combined with Eq. (8.4). In such a region there is no viable value of the mixing κ to allow at the same time for reheating via the Schwinger effect and for the dark photon to avoid a thermal mass. Note that of the dimensionful parameters listed in (8.1), Λ, is fixed by Eq. (4.6), and m does not need to be specified as long as Eq. (8.3) is satisfied, which is always the case. [22,52] (blue in Fig. 3), provide interesting bounds for high f . To cover the whole region with f > 10 14 GeV their sensitivity would have to improve by a few orders of magnitude. For f 10 9 GeV, the mass of the relaxion is above 10 keV. In this region of parameter space, the relaxion can be probed via cosmological and astrophysical processes, or in the laboratories, and there are various constraints studied in Refs. [21,22] (orange in Fig. 3).
Concerning the dark photon, there are almost no experimental constraints in our scenario. This is because the dark photon has to be massless or extremely light, m γ D < 10 −14 GeV, and the mixing very small, κ < 10 −8 (see e.g. [53] for bounds that extend to this region of parameter space).
CMB observables represent perhaps the most interesting arena for testing this framework. The dark photon production can lead to the generation of nongaussianities, primordial black holes and gravitational waves, while the wiggles of the relaxion potential -22 -can produce measurable modulations. These features deserve a dedicated study, which is beyond the scope of the current work.
In this relaxed inflation scenario, we can achieve a higher cutoff than in Ref. [1]. The limiting factors in the original model were the conditions: 1. that the vacuum energy be dominated by the inflaton, 2. that the evolution of the relaxion be dominated by classical rolling rather than quantum fluctuations, 3. that the Hubble parameter during inflation be smaller than Λ wig for the wiggles to appear.
In the framework presented in this paper, these three conditions are not relevant, so we can achieve a cutoff Λ ∼ 10 16 GeV. It is obvious why condition 1 does not apply, as in our case the relaxion is the inflaton itself. Condition 2 is not necessary since δφ settles down when the universe is not de Sitter anymore but radiation dominated. Condition 3 is not necessary either, as our wiggles reappear after reheating once the temperature drops below Λ wig .

Summary
We have investigated a model in which the relaxion, originally proposed in Ref. [1], is also the inflaton. Two key ingredients of the original proposal were a very shallow slope of the potential and the presence of a periodic potential (wiggles), with amplitude growing proportionally to the Higgs VEV. The wiggles provide the backreaction necessary to stop the motion of the relaxion and set the observed EW scale. A shallow slope suggests that the relaxion itself could be the inflaton, as it automatically satisfies the slow-roll conditions. The EW scale must be set after the end of inflation and to avoid overshooting it is necessary to introduce an additional dissipation mechanism. We have shown that this can be accomplished by coupling the relaxion to gauge bosons. In the last stages of inflation, the gauge-boson production becomes significant, slowing down the relaxion and allowing for a new reheating mechanism. The reheating process is an important novelty of this work. It first starts with the production of very strong electric and magnetic fields, which allow for vacuum electronpositron pair production via the Schwinger mechanism. The e + e − pairs quickly thermalize, reheating the universe. To achieve a sufficiently high reheat temperature, the produced gauge bosons cannot be coupled strongly to the thermal bath, as thermal effects quickly shut off the non-perturbative photon production. Here we considered the production of dark photons which are only weakly coupled to the visible sector. We find that this allows to reheat safely above BBN temperature, while the unsuppressed production of dark photons provides enough dissipation for the relaxion, which slows down and settles on the correct EW minimum. A detailed study of this reheating mechanism is under study and will be presented in future work.
-23 -We have studied the phenomenologically viable parameter space, and showed that while our scenario can evade CMB constraints from primordial black hole formation, typically quite stringent, it is difficult to generate the observed amount of curvature perturbations. An extra ingredient, like a curvaton field, is likely needed to match the measured power spectrum. We find that the promotion of the relaxion to an inflaton can accommodate a cutoff close to the Planck scale, significantly above the one found in the original proposal [1].
We couple Φ 1 to fermions charged under a non-Abelian gauge group that confines at the scale Λ wig . Via the one-loop triangle diagram the relaxion obtains the coupling α 1 8π φ f G 1G1 , which gives rise to the periodic wiggle potential. We couple Φ N +1 to fermions -24 -charged under another gauge group with confining scale Λ N Λ wig . Because of the suppressed overlap of the relaxion with the N + 1 field, the operator leads to the coupling f . Below the confining scale, the potential Λ 4 N cos φ F , responsible for the rolling, emerges. By controlling which of the scalars couple to the dark photon, one may control the strength of the photon coupling to the relaxion, namely one can set the value of c γ (D) over a large range [63]. For example, by charging the fermions at the (j + 1)th site under the Abelian gauge symmetry, the relaxion-photon coupling would be c The full clockwork-inspired Lagrangian for the relaxion that we consider is then Here, α cc is a dimensionless constant that we use to tune the cosmological constant to zero.
To make sense of the notion of pNGB, all the scales corresponding to explicit breaking have to be smaller than the spontaneous breaking scale, so we have the following hierarchy 7 The dimensionless parameter g h > 1 determines the point at which we switch from the unbroken to the broken EW phase: With these conventions, we imagine that the rolling starts from φ/F between 0 and π and rolls down to the left. We define φ 0 as the point where m h = 0: Expanding around this point, φ = φ 0 + δφ, we have In the absence of tuning, the scale ΛN is expected to be of order Λ (up to a loop factor), as the h 2 term is going to generate g h Λ 4 16π 2 cos φ F anyway.
-25 - We want to tune the cosmological constant at this point: which after expanding around φ 0 reads Putting all the pieces together we have We see that, by identifying we can match this potential to the one given at the end of Section 3.

B.1 Regime 1
In Section 4.2, we discussed the slow-roll conditions in regime 1, where the barriers from the wiggles are not yet large, namely the condition V roll + V wig > 0 is satisfied. We saw that the parameter (δφ) = −Ḣ/H 2 remains smaller than 1 for values of the potential down to V ∼ M Pl V roll . In this appendix, we discuss in detail the other slow-roll parameter: The last equality holds as long as < 1.
We start from the equation of motion and define the small parameter In regime 1, ϑ typically does not grow larger than 0.1. We expand δφ as This proves that |η| < 1 also in this limit. Note that δφ stays roughly constant because the relaxion does not gain net kinetic energy from the wiggles. Indeed, the maximum deviation from δφ =φ 0 can be estimated by taking one period t f = 2πf /φ 0 ,

B.2 Regime 2
Let us rewrite the full EOM as where we have defined f γ ≡ f cγ , and C 0 2.4 × 10 −4 . Recall that in our conventions V > 0 and δφ < 0. In Section 4.3 we claimed that in regime 2 the EOM is well approximated by With this we check the following conditions: • The kinetic energy is smaller than the potential • H is slowly varying, that is = −Ḣ H 2 < 1. Using the Friedmann equations we can bring to the form [31] The second term in parentheses is smaller than one for V > 2ρ γ = 2V RH , that is roughly until reheating. Then for the first term we have to impose 2|ξ|fγ , see Eq. (4.14).
• The term δφ is negligible. Taking the time derivative of Eq. (B.18) we find Here, the first term is smaller than one for V < V switch , the second vanishes as V roll = 0, the third is small as long as Eq. (B.21) is satisfied.
We see that the condition of Eq. (B.21) is enough to guarantee slow-roll in this approximation. Next we examine what happens when we take into account also the wiggles, the main worry being that δφ could grow larger than V because of the nonzero V wig . We proceed as in the previous section, by defining the small parameter ϑ = V wig V roll and expanding δφ as in Eq. (B.4). We have already solved the zeroth order EOM, that is Eq. (B.17). At first order in ϑ the EOM is We linearized the equation in the second line assuming πcγ ϑδφ (1) f H is another small parameter. One can check this assumption is correct after finding the solution. Now the photon friction is just a modification of the Hubble friction term, and the EOM reduces to Eq. (B.7) with the replacementsφ The solution then is that of Eq. (B.9). It is easy to verify that ϑδφ (1) δφ (0) < 1 and πϑδφ (1) Hfγ < 1, which confirms the consistency of our expansion in Eq. (B.24). Again we can check what happens to the acceleration ϑδφ (1) in two limits: Here we have Here we have We conclude that even when taking into account the wiggle potential, the acceleration δφ in Eq. (B.16) remains negligible.
So far we have checked the self-consistency conditions, |3Hδφ| V , |δφ| V < 1, based on the solution of Eq. (B.18). At last, we examine a possible O(ϑ 0 ) correction to |δφ| V . This is because one could worry that δφ is large when we transition from regime 1 to 2, and perhaps it is not a good approximation to neglect it in the EOM. We show that even if δφ is large at the beginning of regime 2, the solution quickly converges to the one of Eq. (B.18). Keeping the δφ term in the EOM introduces extra time-dependence of δφ (0) leading to a different contribution to |δφ| V . To see this effect, we consider the following differential equation For simplicity we drop the 3Hδφ term, and neglect again the time-dependence of the prefactor on the RHS, keeping only the stronger time-dependence in the exponent. The solution is ).

C Thermal effects on gauge-field production
At zero temperature, the equation of motion (EOM) for the polarization that gets exponentially enhanced reads Here τ is the conformal time, defined as adτ = dt, with t the cosmic time. An overdot denotes a derivative with respect to t. In our conventionsφ < 0. Written in terms of t, the EOM isÄ from which we can read off explicitly the dispersion relation The mode A − experiences tachyonic enhancement when ω 2 < 0. The easiest way to find A − that solves Eq. (C.1) is via the WKB approximation: where, Ω ≡ iω. This approximation holds as long as we satisfy the adiabatic condition (C.8) At finite temperature, in the long wavelength limit, the dispersion relation (C.3) is modified to [64]  with g the U (1) gauge coupling. We want to find tachyonic solutions, Ω = iω > 0, of the equations above.

D Electric field
We discuss here some properties of the classical electric field formed by the exponential number of photons.

D.1 Coherence
The comoving momentum of photons with the largest tachyonic enhancement is k * = −|ξ|/τ (the physical momentum is q γ * = |ξ|H). The occupation number is given by the number of photons in the coherent volume (within the de Broglie wavelentgh), V coh ∼ (|ξ|aH) −3 , This number is significantly larger than 1, implying that the photons are coherent and form a classical field.

D.2 Size and Direction
Even if numerous photons are produced, one might wonder if their random directions result in a zero net electric field. Randomized photons in a microscopic scale must have high momentum, but we have seen that those produced exponentially in our model have low momentum, k * , instead. Thus, at a comoving scale larger than k −1 * , roughly, we expect zero electric field, but we will have a non-zero field when we zoom into scales smaller than k −1 * . We can make these statements more explicit by using an averaged electric field within a radius R, 8 where V R is volume inside a sphere with the radius R from x. We study the dispersion of the averaged electric field, V R e i k· x 2 = 9 [sin(kR) − kR cos(kR)] 2 /(kR) 6 . Since f R (k) is a function damping quickly for k R −1 , for simplicity we treat it as a step function, f R (k) → 8 We thank Masahiro Takimoto for suggesting this quantity. because the production of non-tachyonic photons, with momentum k > 2k * , is negligible. The IR cutoff is needed because there are no zero momentum photons. We examine two cases, with microscopic and macroscopic scales R, Averaging over large scales (R > k −1 IR ), there is no net electric field, while at small scales (R < (2k * ) −1 ), there is a strong electric field E 2 ∼ ρ γ , as given in Eq. (2.16). The transition from E 2 to zero, going from small to large scales, is expected to be smooth.
The direction of the electric field can appear as a consequence of quantum fluctuations which grow exponentially. Our analytic approach is limited to estimating quadratic quantities, such as E 2 , but cannot probe directions. In order to observe the direction, one needs a lattice simulation, which is beyond the scope of this paper. For a similar situation of tachyonic instability, simulations were performed in Refs. [42,43]. They studied a potential V = λ 4 (φ 2 − v 2 ) 2 with a homogenous initial condition in the symmetric phase (φ = 0) and initial quantum fluctuations. Then a tachyonic instability drives the inhomogeneity: some patches have φ = v and other patches have φ = −v. The appearance of a direction of the electric field is analogous to this inhomogeneity.

E Estimate of curvature perturbations
The power spectrum from the usual vacuum fluctuations of the inflaton, neglecting the contribution from gauge fields, is We want to check if at 30 e-folds from the end of inflation P can match the observed one, P COBE = 2.5 × 10 −9 . First ,we need to estimate H andφ at that time. We are in regime 1, withφ The number of e-folds N 1 in this regime, before we switch to the one dominated by photon backreaction, is The curvature perturbations generated only by the linear slope in our model are many orders of magnitude below what is measured.