betaFeedback

WorksPhysics and Astronomy

General δNδN formalism in Single Field Inflation and Graph Theory

Abstract

We study higher order correlation functions in single field inflation using the δNδN formalism and methods from graph theory. By representing connected tree level contributions as graphs, we systematically classify their independent topologies, momentum shapes and amplitudes. We show that the amplitudes of arbitrary trees factorize into products of primary amplitudes, providing a general framework for relations among higher order correlators, including the Suyama-Yamaguchi equality. As the number of external legs increases, the number of distinct tree topologies grows exponentially, whereas the number of algebraically distinct amplitude products grows considerably more slowly. As an application, we consider ultra-slow-roll inflation followed by a transition to a slow-roll attractor. For transitions without a potential step, we derive an explicit all-order expression for the vertex amplitudes and establish their positivity and dependence on the sharpness of the transition. When a potential step is included, we prove that all nonlinear vertex amplitudes remain positive for upward steps, whereas downward steps allow cancellations at specific parameter values. These cancellations can eliminate entire families of tree contributions, including configurations in which the leading local bispectrum and trispectrum vanish simultaneously while higher order correlations remain nonzero.

Introduction

Cosmological perturbation theory is the fundamental tool to study cosmological observables such as the cosmic microwave background (CMB) temperature anisotropies and the power spectrum or bispectrum of primordial curvature perturbations [1, 2, 3, 4]. The general view is that these perturbations are generated from quantum fluctuations during inflation in early universe which eventually provide the seeds of CMB anisotropies and large scale structure. There are two main approaches to study primordial perturbations within cosmological perturbation theory, namely the standard quantum field theory approach and the alternative δN\delta N formalism. In the standard QFT approach one promotes the cosmological observables, such as the curvature perturbations on comoving slice R\mathcal{R} or the Bardeen potential, into quantum operators. These quantum operators are further endowed with appropriate vacuum (initial conditions) deep inside the horizon such as the Baunch-Davies (Minkowski) vacuum and with the corresponding creation and annihilation operators. One then look at the evolution of the corresponding mode functions and calculate the expectation values such the two-point functions (power spectrum) and three-point functions (bispectrum) etc at the surface of end of inflation. This is achieved within the machinery of QFT in-in formalism [5, 6, 7, 8] where all information is encoded in the corresponding interaction Hamiltonian.

On the other hand, in the alternative approach based on δN\delta N formalism one employs the separate universe approach in which perturbations on superhorizon scales follow the background trajectories [9, 10, 11, 12, 13, 14]. Different patches of Hubble size H−1H^{-1} are viewed as independent FLRW backgrounds which are slightly different from each other by their initial conditions which are induced by the corresponding large scale perturbations. To calculate curvature perturbations, one only has to read the differences in the background e-fold number between an initial hypersurface and a final hypersurface. The initial hypersurface is spatially flat while the final surface is a surface of constant energy. The advantage of δN\delta N formalism is that one only needs to follow the background trajectory and calculate the number of e-folds NN as a function of background fields and their velocities. Another important advantage of δN\delta N approach is that it is a non-linear formalism so it does not need the perturbative expansion as is required in the standard QFT approach. Correspondingly, δN\delta N is a powerful tool in calculating non-Gaussianities and the full probability density function for primordial black holes (PBHs) formation [15, 16, 17, 18]. δN\delta N formalism has been employed extensively in literature to calculate bispectrum and trispectrum in single field and multiple fields setups, for an incomplete list see [19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38]. In addition, stochastic δN\delta N formalism proved very fruitful to study cosmological correlators in the large diffusion limit within the context of stochastic inflation [39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55].

In this work we take a different approach and look at the δN\delta N formalism and the corresponding analysis of cosmological correlators from the graph theory point of view. It turns out that general and powerful mathematical conclusions can be established for the nn-point correlation functions of comoving curvature perturbation in single field setup within the context of δN\delta N formalism which can be nicely represented via graph theory.

Correlation Functions

Our goal in this section is to express the nn-point correlation functions of curvature perturbations in single field inflation using δN\delta N formalism. Here we present the shapes of correlation functions for the simple cases of power spectrum (n=2n = 2), bispectrum (n=3n = 3) and trispectrum (n=4n = 4). After this review we set the stage for the general nn-point correlation function. The background is given by a flat FLRW metric,

ds2=−dt2+a(t)2dx2(1)ds^2=-dt^2+a(t)^2dx^2 \tag*{(1)}

in which tt is the cosmic time and a(t)a(t) is the scale factor. Related to the scale factor, the Hubble expansion parameter is defined via H≡a˙/aH \equiv\dot{a}/a in which a dot indicates the derivative with respect to tt. We assume that inflation is driven by a single scalar field, the inflaton field ϕ\phi, with the potential V(ϕ)V(\phi). It is assumed that during inflation the energy density responsible for the expansion is predominantly provided by the potential and the inflaton kinetic energy is subleading. Defining the first slow-roll (SR) parameter ϵ≡−H˙/H2\epsilon\equiv-\dot{H}/H^2, this assumption corresponds to ϵ≪1\epsilon\ll1. However, we do not assume that the system is always in the SR regime. Indeed, there can be periods during inflation in which the second SR parameter η=ϵ˙/Hϵ\eta=\dot{\epsilon}/H\epsilon may not be small such as in the ultra slow-roll (USR) setup. However, the assumption ϵ≪1\epsilon\ll1 ensures that we have a near exponential expansion for the scale factor a(t)∼eHta(t)\sim e^{Ht} with HH being nearly constant during inflation.

In our analysis below, it is more useful to use the number of e-fold NN as the clock which is related to the cosmic time via dN=H(t)dtdN=H(t)dt. One may chose the convention that N=0N=0 at the start of inflation (or when the CMB observables leave the Hubble horizon) while the time of end of inflation is denoted by N=NfN=N_f. To solve the flatness and the horizon problems in big bang cosmology, one typically requires about 60 number of e-folds, i.e. Nf∼60N_f\sim60.

To employ δN\delta N formalism, one has to count the differences in NN between two hypersurfaces. The initial hypersurface is spatially flat where the inflaton perturbations δϕ\delta\phi at the moment of horizon crossing have the amplitude H/2πH/2\pi. On the other hand, the final hypersurface is the surface of constant energy density which in simple SR setups is determined by the time of reheating. Between these two hypersurfaces, one obtains NN as a function of the background field and its velocity, N=N(ϕ,ϕ˙)N=N(\phi,\dot{\phi}). Correspondingly, the curvature perturbation R\mathcal{R} on superhorizon scales is related to δN\delta N via R=δN\mathcal{R}=\delta N [12, 14].

In general, N(ϕ,ϕ˙)N(\phi,\dot{\phi}) can be calculated non-perturbatively. However, in many situations we are interested in perturbative regimes in which, after obtaining N(ϕ,ϕ˙)N(\phi,\dot{\phi}), one expands it perturbatively to calculate the correlations functions such as the power spectrum, bispectrum, trispectrum etc. Expanding N(ϕ,ϕ˙)N(\phi,\dot{\phi}) in this way, yields

δN=N′δϕ+N′′2!δϕ2+N′′′3!δϕ3+⋯=∑j=1∞N(j)(ϕ,ϕ˙)δϕjj!.(2)\begin{aligned} \delta N &= N'\delta\phi+\frac{N''}{2!}\delta\phi^2+\frac{N'''}{3!}\delta\phi^3+\cdots\\ &=\sum_{j=1}^{\infty}N^{(j)}(\phi,\dot{\phi})\frac{\delta\phi^j}{j!}. \tag*{(2)} \end{aligned}

Here and below, a prime on NN indicates the derivative with respect to the background ϕ\phi, i.e. N′=N,ϕN'=N_{,\phi}, N′′=N,ϕϕN''=N_{,\phi\phi} and so on. Similarly, N(j)N^{(j)} represents the jj-th derivative with respect to ϕ\phi.

In writing the expansion in Eq. (2) we have neglected δϕ˙\delta\dot{\phi} because it vanishes on superhorizon scales. Having said this, it is crucial to note that we should treat NN as a function of both ϕ\phi and ϕ˙\dot{\phi} so it is a function in the two-dimensional phase space. Only in the case of attractor regime one can interchangeably use t↔ϕt\leftrightarrow\phi and consider NN as function of ϕ\phi. However, in general, as in USR setup when the system is in non-attractor regime, this assumption is not valid and one has to solve NN in the full phase space.

With the relation R=δN\mathcal{R}=\delta N and using the latter expansion given in Eq. (2), one can provide a direct link with the correlations of R\mathcal{R} and δϕ\delta\phi to all orders in perturbation theory. This can be done either in configuration space or in Fourier space. Usually, it is more convenient to work in the Fourier space (k space) in which the amplitude of δϕk\delta\phi_k is known at the moment of horizon crossing,

⟨δϕk1δϕk2⟩=(2π)3δ3(k1+k2)H22k3≡(2π)3δ3(k1+k2)Pδϕ.(3)\langle\delta\phi_{k_1}\delta\phi_{k_2}\rangle=(2\pi)^3\delta^3(\mathbf{k}_1+\mathbf{k}_2)\frac{H^2}{2k^3}\equiv(2\pi)^3\delta^3(\mathbf{k}_1+\mathbf{k}_2)P_{\delta\phi}. \tag*{(3)}

In using δN\delta N formalism it is understood that all modes of interest kik_i are superhorizon, ki<aHk_i<aH. Starting with the power spectrum,

⟨Rk1Rk2⟩≡(2π)3δ3(k1+k2)PR(4)\langle\mathcal{R}_{k_1}\mathcal{R}_{k_2}\rangle\equiv(2\pi)^3\delta^3(\mathbf{k}_1+\mathbf{k}_2)P_{\mathcal{R}} \tag*{(4)}

we obtain,

PR(k)=(N′)2Pδϕ=(N′)2H22k3.(5)P_{\mathcal{R}}(k)=(N')^2P_{\delta\phi}=(N')^2\frac{H^2}{2k^3}. \tag*{(5)}

Going to higher order correlation functions, the bispectrum is given by the three-point function via,

⟨Rk1Rk2Rk3⟩=(2π)3δ3(k1+k2+k3)BR(k1,k2,k3).(6)\langle\mathcal{R}_{k_1}\mathcal{R}_{k_2}\mathcal{R}_{k_3}\rangle=(2\pi)^3\delta^3(\mathbf{k}_1+\mathbf{k}_2+\mathbf{k}_3)B_{\mathcal{R}}(k_1,k_2,k_3). \tag*{(6)}

As the inflaton perturbations δϕ\delta\phi is Gaussian, to calculate the bispectrum one has to consider the second order expansion in Eq. (2) in which two δϕ\delta\phi’s are contracted with one (δϕ)2(\delta\phi)^2. Performing the corresponding contractions associated to Gaussian perturbations δϕ\delta\phi, one obtains

BR(k1,k2,k3)=(N′)2N′′(H2)4S(3)(k1,k2,k3),(7)B_{\mathcal{R}}(k_1,k_2,k_3)=(N')^2N''\left(\frac{H}{2}\right)^4S_{(3)}(k_1,k_2,k_3), \tag*{(7)}

in which the bispectrum shape is defined via,

S(3)≡1k13k23+1k13k33+1k23k33.(8)S_{(3)}\equiv\frac{1}{k_1^3k_2^3}+\frac{1}{k_1^3k_3^3}+\frac{1}{k_2^3k_3^3}. \tag*{(8)}

This is the standard local-shape non-Gaussianity. An immediate conclusion is that the bispectrum has only one independent shape (the local shape)

It is customary to define the amplitude of local-type non-Gaussianity fNLf_{NL} by factoring out two powers of PRP_{\mathcal{R}} from Eq. (7). Noting from Eq. (5) that PRP_{\mathcal{R}} is related to PδϕP_{\delta\phi} by an additional factor N′2N'^2, the bispectrum is written in the usual form as follows [56],

BR(k1,k2,k3)=65fNL[PR(k1)PR(k2)+PR(k1)PR(k3)+PR(k2)PR(k3)],(9)B_{\mathcal{R}}(k_1,k_2,k_3)=\frac{6}{5}f_{NL}\left[P_{\mathcal{R}}(k_1)P_{\mathcal{R}}(k_2)+P_{\mathcal{R}}(k_1)P_{\mathcal{R}}(k_3)+P_{\mathcal{R}}(k_2)P_{\mathcal{R}}(k_3)\right], \tag*{(9)}

in which fNLf_{NL} is obtained to be,

fNL=56N′′(N′)2(N′)4=56N′′(N′)2≡56FNL.(10)f_{NL}=\frac{5}{6}\frac{N''(N')^2}{(N')^4}=\frac{5}{6}\frac{N''}{(N')^2}\equiv\frac{5}{6}F_{NL}. \tag*{(10)}

For the later convenience, we have defined FNLF_{NL} by absorbing the numerical factor 5/6. The last example is the trispectrum which is given by the four-point function as follows,

⟨Rk1Rk2Rk3Rk4⟩=(2π)3δ3(k1+k2+k3+k4)TR(k1,k2,k3,k4).(11)\langle\mathcal{R}_{k_1}\mathcal{R}_{k_2}\mathcal{R}_{k_3}\mathcal{R}_{k_4}\rangle=(2\pi)^3\delta^3(\mathbf{k}_1+\mathbf{k}_2+\mathbf{k}_3+\mathbf{k}_4)T_{\mathcal{R}}(k_1,k_2,k_3,k_4). \tag*{(11)}

To calculate the trispectrum we have two options in higher order contractions. Either contract two δϕ\delta\phi with one δϕ2\delta\phi^2, or contact three δϕ\delta\phi with one δϕ3\delta\phi^3. Correspondingly, in Fourier space, we have two distinct local shapes in trispectrum, parameterized as follows [56],

TR(k1,k2,k3,k4)=τNL[PR(k13)PR(k3)PR(k4)+(11 perms)]+5425gNL[PR(k2)PR(k3)PR(k4)+(3 perms)],(12)T_{\mathcal{R}}(k_1,k_2,k_3,k_4) = \tau_{NL}\left[P_{\mathcal{R}}(k_{13})P_{\mathcal{R}}(k_3)P_{\mathcal{R}}(k_4) + (11\ \text{perms})\right] + \frac{54}{25}g_{NL}\left[P_{\mathcal{R}}(k_2)P_{\mathcal{R}}(k_3)P_{\mathcal{R}}(k_4) + (3\ \text{perms})\right], \tag*{(12)}

where we have defined kij≡∣ki+kj∣k_{ij} \equiv\lvert\mathbf{k}_i+\mathbf{k}_j\rvert. The two distinct shapes of trispectrum from the above expression are defined as follows,

S(4)(a)(k1,k2,k3,k4)=(1k13k23k33+3 Permutations),(13)S^{(a)}_{(4)}(\mathbf{k}_1,\mathbf{k}_2,\mathbf{k}_3,\mathbf{k}_4) = \left(\frac{1}{k_1^3 k_2^3 k_3^3} + 3\ \text{Permutations}\right), \tag*{(13)}
S(4)(b)(k1,k2,k3,k4)=(1k13k23k133+11 Permutations).S^{(b)}_{(4)}(\mathbf{k}_1,\mathbf{k}_2,\mathbf{k}_3,\mathbf{k}_4) = \left(\frac{1}{k_1^3 k_2^3 k_{13}^3} + 11\ \text{Permutations}\right).

In Eq. (12), τNL\tau_{NL} and gNLg_{NL} are two new parameters associated to the local-shape trispectrum. After taking into account the additional factor (N′)2(N')^2 relating PδϕP_{\delta\phi} to PRP_{\mathcal{R}}, we obtain

gNL=2554N′′′(N′)3(N′)6=2554N′′′(N′)3≡2554GNL,(14)g_{NL} = \frac{25}{54}\frac{N'''(N')^3}{(N')^6} = \frac{25}{54}\frac{N'''}{(N')^3} \equiv\frac{25}{54}G_{NL}, \tag*{(14)}

and

τNL=(N′′)2(N′)2(N′)6=(N′′)2(N′)4≡TNL.(15)\tau_{NL} = \frac{(N'')^2(N')^2}{(N')^6} = \frac{(N'')^2}{(N')^4} \equiv T_{NL}. \tag*{(15)}

Again, to absorb the numerical prefactors and to unify the notation, we have defined GNLG_{NL} and TNLT_{NL}. Upon inspection, one observes that τNL\tau_{NL} and fNLf_{NL} in Eqs. (15) and (10) are related to each other via TNL=FNL2T_{NL}=F_{NL}^2. This is the special case of the Suyama-Yamaguchi (SY) inequality which states that TNL≥FNL2T_{NL}\geq F_{NL}^2 for general multiple field setups [57, 58]. The equality in the above inequality is valid for any single field model of inflation in the tree level i.e. in the absence of loop corrections [59].

Motivated by the above examples, in this work we study the nn-point correlation function of curvature perturbation at the tree level for a general single field model,

⟨Rk1Rk2…Rkn⟩.(16)\langle\mathcal{R}_{k_1}\mathcal{R}_{k_2}\ldots\mathcal{R}_{k_n}\rangle. \tag*{(16)}

As the example of trispectrum has shown, one expects to have more independent shapes for higher values of nn. In addition, one may expect a generalized form of the SY equality to emerge. In order to study the higher order correlation functions, in the next section we present a graph theory representation of the general nn-point correlation function. This approach, as in QFT Feynman diagrams, helps us to visualize the important ingredients in higher order correlations and to classify independent shapes based on their physical and mathematical structures.

Before presenting our graph theory diagrams for δN\delta N analysis, we comment that diagrammatic representations of cosmological correlators and non-Gaussianity were studied in [59, 60, 61, 62] as well. In particular, Refs. [59, 60] employ Feynman-type diagrammatic approach to study nn-point functions with emphasis on loop corrections. On the other hand, Refs. [61, 62] study nn-point functions using diagrammatic approach similar to ours. Ref. [61] presents reduced and tree diagrams with the aim to calculate the contributions to the nn-point correlation functions at a given time NFN_F in the tree diagrams. The vertices start from an initial flat hypersurfaces and, by use of interacting vertices as well as propagators, they are connected to the time NFN_{\mathrm{F}}. The contribution of each tree to the nn-point correlation function is then given as an integral over the production of the propagators and interacting vertices [61]. On the other hand, Ref. [62] studies the SY inequality for higher nn-point correlation functions in which one can write inequalities between the amplitudes of some certain diagrams and two sub diagrams obtained from them.

δN\delta N Diagrams

Here we present our graph theory representation of δN\delta N formalism in calculating the nn-point correlation function.

Looking at the δN\delta N expansion in terms of δϕ\delta\phi in Eq. (2), we see that due to the nonlinear form of the expansion, various powers of δϕ\delta\phi's can contract with each other to yield a nonzero value for the nn-point correlation function. This is clearly seen in the example of bispectrum studied in the previous section in which one has to contract two δϕ\delta\phi's with external momentums ki\mathbf{k}_i and kj\mathbf{k}_j with a quadratic combination (δϕ2)kr(\delta\phi^2)_{\mathbf{k}_r} subject to ki+kj+kr=0\mathbf{k}_i+\mathbf{k}_j+\mathbf{k}_r=0. On the other hand, the situation for trispectrum becomes more non-trivial as we have two different non-linear combinations δϕ2\delta\phi^2 and δϕ3\delta\phi^3 in which the remaining δϕ\delta\phi's can contract. Correspondingly, we obtain two different shapes for the trispectrum controlled by the parameters GNLG_{\mathrm{NL}} and TNLT_{\mathrm{NL}}. As one can see, going to higher order this pattern repeats and more an more independent shapes will emerge. To study the general case of correlation functions, it proved very helpful to look at the correlation functions as in graph theory diagrams which we will elaborate below.

Definitions and basic graph-theoretic results

We begin by introducing the graph-theoretic terminology needed to classify the connected tree level contributions to the nn-point correlation functions.

In our diagrammatic representation, each curvature perturbation Rki\mathcal{R}_{\mathbf{k}_i} appearing in the nn-point function is represented by a vertex. Expanding Rki\mathcal{R}_{\mathbf{k}_i} in powers of the Gaussian field perturbation δϕ\delta\phi, as in Eq. (2), each vertex is associated with a term proportional to N(d)(δϕ)d/d!N^{(d)}(\delta\phi)^d/d!. Wick contractions between field perturbations belonging to different vertices are represented by edges connecting the corresponding vertices. Consequently, a vertex associated with the dd-th order term in the δN\delta N expansion has dd incident edges. The resulting diagrams can therefore be studied using the standard concepts of graph theory. In the following, we restrict ourselves to connected tree level diagrams. These are represented by graphs with vertices corresponding to curvature perturbation insertions and edges corresponding to Gaussian contractions. The momentum labels are temporarily suppressed when classifying the graph topologies.

Definition 1. (Degree and degree sequence): The degree dvd_v of a vertex vv is the number of edges incident to it. The degree sequence of a graph with nn vertices is the sequence (d1,…,dn)(d_1,\ldots,d_n) of vertex degrees arranged in non-increasing order, d1≥d2≥⋯≥dnd_1\ge d_2\ge\cdots\ge d_n. In our δN\delta N representation, a vertex of degree dd corresponds to the term N(d)(δϕ)d/d!N^{(d)}(\delta\phi)^d/d! in the expansion of the curvature perturbation.

Definition 2. (Path). A path is a sequence of distinct vertices v0,v1,…,vrv_0, v_1, \ldots, v_r such that each pair of consecutive vertices is connected by an edge.

Definition 3. (Connected graph): A graph is connected if there exists a path between every pair of distinct vertices.

Definition 4. (Cycle): A cycle is a closed sequence of edges connecting distinct vertices, except that the initial and final vertices coincide.

Definition 5. (Tree): A tree is a connected graph containing no cycles.

Definition 6. (Leaf): A leaf is a vertex of degree one. In our δN\delta N representation, a leaf corresponds to the linear term N′δϕN'\delta\phi.

Definition 7. (Star tree): A star tree SnS_n is a tree with n≥3n \ge3 vertices in which one vertex is adjacent to all the remaining n−1n-1 vertices. Its degree sequence is (n−1,1,…,1)(n-1,1,\ldots,1). In our follow up δN\delta N representation, we refer to the star tree as the “primary diagram.” The significance of this terminology will become clear in § 4, where we study the factorization of tree amplitudes.

It is instructive to examine the simple tree diagrams associated with n=2,3,4,5n = 2, 3, 4, 5 before proceeding to the general nn-point correlation functions. These diagrams are illustrated in Figure 1. For the power spectrum (n=2)(n = 2), the graph consists of two vertices of degree one (leaves), connected by a single edge. For the bispectrum (n=3)(n = 3), there are three vertices and two edges. The central vertex has degree two, corresponding to the quadratic term N′′(δϕ)2/2!N''(\delta\phi)^2/2! in the δN\delta N expansion, while the remaining two vertices are leaves associated with the linear term N′δϕN'\delta\phi. There is only one tree diagram with three vertices, corresponding to the unique local bispectrum shape. For the trispectrum (n=4)(n = 4), there are two independent trees, each containing four vertices and three edges. The first is the star tree S4S_4, with degree sequence (3,1,1,1)(3,1,1,1), whose central vertex corresponds to the cubic term in the δN\delta N expansion. This diagram generates the momentum shape S(4)(a)S_{(4)}^{(a)} in Eq. (13), with amplitude GNLG_{\mathrm{NL}}. The second is the path tree P4P_4, with degree sequence (2,2,1,1)(2,2,1,1), containing two vertices of degree two and two leaves. It generates the momentum shape S(4)(b)S_{(4)}^{(b)} in Eq. (13), with amplitude TNLT_{\mathrm{NL}}. Finally, for n=5n=5 we have three independent diagrams identified by the sequences (4,1,1,1,1)(4,1,1,1,1) which is the star diagram S5S_5, (3,2,1,1,1)(3,2,1,1,1) which is denoted by T5T_5 and (2,2,2,1,1)(2,2,2,1,1) which is the path tree P5P_5.

Non-isomorphic independent tree diagrams for $n=2,3,4,5$

Figure 1. Non-isomorphic (independent) tree diagrams for n=2,3,4,5n = 2, 3, 4, 5. Each vertex represents a curvature perturbation insertion, and the number above each vertex denotes its degree. The degree sequence is displayed beneath each diagram.

The notation in Figure 1 follows standard graph-theoretic terminology: SnS_n denotes a star tree with nn vertices, while PnP_n denotes a path graph with nn vertices, in which the vertices form a single chain. The symbol T5T_5 is used in Figure 1 as a label for the remaining five-vertex tree, which is neither a star nor a path; it has degree sequence (3,2,1,1,1)(3,2,1,1,1). All diagrams in Figure 1 are connected and contain no cycles, and therefore qualify as trees. This restriction is important for our analysis, since we consider the connected tree level contributions to the δN\delta N correlation functions and do not study quantum loop corrections in correlation functions. For comparison, Figure 2 illustrates graphs containing cycles. In standard graph theory, CnC_n denotes a cycle graph with nn vertices, each of degree two. Figure 2(a) shows the triangle C3C_3, while Figure 2(b) shows a cycle C4C_4 with an additional leaf attached to one of its vertices. These graphs are not trees because they contain cycles, and hence lie outside the class of diagrams considered in our tree level analysis.

Examples of graphs containing cycles

Figure 2. Examples of graphs containing cycles. (a) The triangle C3C_3, consisting of three vertices of degree two. (b) A connected graph consisting of a four-vertex cycle C4C_4 with an additional leaf attached to one vertex. The numbers next to the vertices indicate their degrees.

With the above definitions, we first consider the relation between the number of vertices and edges in a tree. A standard result in graph theory states the following [63].

Theorem 1. A tree with nn vertices has exactly n−1n-1 edges.

This statement can be verified directly from Figure 1, where each tree with n=2,3,4,5n=2,3,4,5 vertices contains precisely n−1n-1 edges. In our δN\delta N representation, each edge corresponds to a contraction of two field perturbations, or equivalently to a propagator of δϕ\delta\phi. Thus, a connected tree level nn-point function contains exactly n−1n-1 propagators.

The above theorem also implies the following characterization of trees.

Lemma 1. Among all connected graphs with nn vertices, the minimum possible number of edges is n−1n-1. This minimum is attained if and only if the graph is a tree.

Indeed, every connected graph contains a spanning tree, which has n−1n-1 edges. If the graph contains additional edges, it necessarily contains a cycle. The graphs in Figure 2 illustrate this distinction in which both diagrams are connected but contain cycles and therefore have more than n−1n-1 edges.

Next, we consider the relation between the number of propagators and the degrees of the vertices. Let dvd_v denote the degree of vertex vv. Then one has [63],

∑v=1ndv=2(n−1).(17)\sum_{v=1}^{n} d_v = 2(n-1). \tag*{(17)}

In our δN\delta N representation, the degree dvd_v determines the order of the field perturbation at vertex vv. Consequently, Eq. (17) shows that the total number of δϕ\delta\phi fields appearing in the corresponding tree level nn-point contribution is 2(n−1)2(n-1), with the fields contracted pairwise to form n−1n-1 propagators.

This relation can be checked explicitly in Figure 1. For the power spectrum, n=2n=2 and the degrees are (1,1)(1,1), giving 1+1=21+1=2. For the star diagram S4S_4, associated with one of the trispectrum shapes, the degrees are (3,1,1,1)(3,1,1,1), giving 3+1+1+1=6=2(4−1)3+1+1+1=6=2(4-1). The same relation holds for the path diagram P4P_4, whose degree sequence is (2,2,1,1)(2,2,1,1).

We now ask under what conditions a given sequence of positive integers (d1,…,dn)(d_1,\ldots,d_n) can represent the vertex degrees of a tree diagram. Before addressing this question, we present a useful property of trees:

Lemma 2. Every tree with at least two vertices has at least two leaves, namely vertices of degree one [63].

For example, all trees in Figure 1 contain at least two leaves. The path trees PnP_n have exactly two leaves, whereas the star trees SnS_n have n−1n-1 leaves.

Using the preceding results, we obtain the following characterization of the tree degree sequences.

Theorem 2. A given sequence of n≥2n \ge2 positive integers (d1,…,dn)(d_1,\ldots,d_n) is the degree sequence of at least one tree if and only if,

∑i=1ndi=2(n−1).(18)\sum_{i=1}^{n} d_i = 2(n-1). \tag*{(18)}

Proof. Necessity follows immediately from Eq. (17). To establish sufficiency, we proceed by induction on nn. For n=2n=2, the only possible sequence of positive integers satisfying the condition is (1,1)(1,1), which corresponds to the single edge tree representing the power spectrum. Now, suppose the statement holds for sequences of nn positive integers. Consider a sequence (d1,…,dn+1)(d_1,\ldots,d_{n+1}) satisfying ∑i=1n+1di=2n\sum_{i=1}^{n+1} d_i = 2n. As in the proof of Lemma 2, this condition implies that at least two entries are equal to one. Choose one such entry, say dj=1d_j=1, and another entry dk≥2d_k \ge2. Remove djd_j from the sequence and replace dkd_k by dk−1d_k-1. The resulting sequence contains nn positive integers and satisfies

∑i=1ndi′=2(n−1).\sum_{i=1}^{n} d_i' = 2(n-1).

By the induction hypothesis, this reduced sequence is realized by a tree with nn vertices. Attaching a new leaf to the vertex whose degree was reduced from dkd_k to dk−1d_k-1 restores the original degree sequence without introducing a cycle. The resulting graph is therefore a tree with n+1n+1 vertices and the prescribed degrees. This completes the induction. □\square

Theorem 2 establishes that every sequence of positive integers satisfying Eq. (17) can be realized by at least one tree. However, the degree sequence does not necessarily determine the tree uniquely. In particular, for sufficiently large nn, distinct tree topologies may share the same degree sequence. This distinction becomes relevant for the classification of higher order correlation functions. We therefore introduce the standard graph-theoretic notion of isomorphism:

Definition 8: (Graph isomorphism). Two trees T1=(V1,E1)T_1=(V_1,E_1) and T2=(V2,E2)T_2=(V_2,E_2) are isomorphic if there exists a bijection f:V1→V2f:V_1\to V_2 such that, for every pair of vertices u,v∈V1u,v\in V_1,

{u,v}∈E1⟺{f(u),f(v)}∈E2,(19)\{u,v\}\in E_1 \quad\Longleftrightarrow\quad\{f(u),f(v)\}\in E_2, \tag*{(19)}

in which {x,y}\{x,y\} represents the edge connecting the vertices xx and yy. In other words, an isomorphism preserves adjacency between vertices. Two trees are non-isomorphic if no such bijection exists.

An immediate consequence is that isomorphic trees have the same vertex degrees, up to permutation. The converse, however, is not generally true. As an example, consider the two trispectrum diagrams in Figure 1. The star tree S4S_4 has degree sequence (3,1,1,1)(3,1,1,1), while the path tree P4P_4 has degree sequence (2,2,1,1)(2,2,1,1). Since graph isomorphisms preserve vertex degrees, these two trees cannot be isomorphic. They therefore represent two distinct tree topologies, associated with the two independent local trispectrum shapes discussed in § 2.

This distinction motivates the counting of non-isomorphic trees, which we consider in the next subsection.

Number of Independent Diagrams

The first question we address is the number of independent tree diagrams for a given nn, where independent diagrams are understood as non-isomorphic trees. Although there is no simple closed form expression for this number, the following generating function relations, associated with the Otter theorem [64], allow us to determine it at arbitrary order.

Theorem 3 (Otter’s tree counting theorem). Let T(n)T(n) denote the number of non-isomorphic trees with nn vertices. Define the generating function via,

ϕ(x)=T(1)+T(2)x+T(3)x2+⋯ .(20)\phi(x)=T(1)+T(2)x+T(3)x^2+\cdots. \tag*{(20)}

Then,

ϕ(x)=ψ(x)−x2[ψ(x)2−ψ(x2)],(21)\phi(x)=\psi(x)-\frac{x}{2}\left[\psi(x)^2-\psi(x^2)\right], \tag*{(21)}

where

ψ(x)=A1+A2x+A3x2+⋯ ,(22)\psi(x)=A_1+A_2x+A_3x^2+\cdots, \tag*{(22)}

is the corresponding generating function, with AnA_n denoting the number of such trees with nn vertices. It satisfies

ψ(x)=exp⁡[∑ν=1∞xννψ(xν)].(23)\psi(x)=\exp\left[\sum_{\nu=1}^{\infty}\frac{x^\nu}{\nu}\psi(x^\nu)\right]. \tag*{(23)}

Solving the above recurrence relations together yields the number of non-isomorphic trees associated with nn.

To illustrate this procedure, let us expand ψ(x)\psi(x) to the first few orders,

ψ(x)=A1+A2x+A3x2+O(x3).(24)\psi(x)=A_1+A_2x+A_3x^2+O(x^3). \tag*{(24)}

On the other hand, expanding Eq. (23), we obtain,

ψ(x)=1+A1x+A12+A1+2A22x2+O(x3).(25)\psi(x)=1+A_1x+\frac{A_1^2+A_1+2A_2}{2}x^2+O(x^3). \tag*{(25)}

Matching the coefficients of equal powers of xx in the above two equations yields,

A1=1,A2=1,A3=2.(26)A_1=1,\qquad A_2=1,\qquad A_3=2. \tag*{(26)}

Substituting these results into Eq. (21), we find,

ϕ(x)=1+x+x2+O(x3),(27)\phi(x)=1+x+x^2+O(x^3), \tag*{(27)}

which gives T(1)=T(2)=T(3)=1T_{(1)}=T_{(2)}=T_{(3)}=1. Therefore, there is one independent tree diagram for the power spectrum (n=2n=2) and one tree diagram for the bispectrum (n=3n=3) as illustrated in Figure 1. Here, T(1)=1T_{(1)}=1 counts the trivial tree consisting of a single vertex.

Continuing this procedure to higher orders, we obtain

T(4)=2,T(5)=3,T(6)=6,T(7)=11,T(8)=23,T(9)=47,T(10)=106.(28)T_{(4)}=2,\qquad T_{(5)}=3,\qquad T_{(6)}=6,\qquad T_{(7)}=11,\qquad T_{(8)}=23,\qquad T_{(9)}=47,\qquad T_{(10)}=106. \tag*{(28)}

The two independent diagrams for the trispectrum (T(4)=2T_{(4)}=2) and the three independent diagrams for the five-point function (T(5)=3T_{(5)}=3) are displayed in Figure 1.

As can be seen from the above examples, the number of independent diagrams increases rapidly with nn. For large nn, Otter’s asymptotic result gives [64],

T(n)∼0.535(2.956)nn−5/2,(n≫1).(29)T_{(n)}\sim0.535(2.956)^n n^{-5/2},\qquad(n\gg1). \tag*{(29)}

For example, at n=20n=20, the exact number of non-isomorphic trees is T(20)=823 065T_{(20)}=823\,065, whereas the asymptotic expression gives approximately 776 000776\,000, differing from the exact result by about 5.7%5.7\%.

Tree Amplitudes and Their Factorization

In this section, we study the amplitudes associated with the connected tree diagrams contributing to the nn-point correlation functions. As in the examples of the bispectrum and trispectrum discussed in § 2, the contribution of each tree diagram can be separated into an amplitude and a momentum dependent shape function. Our main goal here is to establish the general structure of the amplitudes and their relations at different orders.

Tree Amplitudes

Consider a tree TT with nn vertices, and denote the degree of vertex vv by dvd_v. As established in § 3.1, the degrees satisfy,

∑v∈V(T)dv=2(n−1).(30)\sum_{v\in V(T)} d_v = 2(n-1). \tag*{(30)}

The contribution of this tree to the connected nn-point function is proportional to

⟨Rk1⋯Rkn⟩T′=(H2)2(n−1)∏v∈V(T)N(dv)ST(k1,…,kn),(31)\left\langle\mathcal{R}_{k_1}\cdots\mathcal{R}_{k_n}\right\rangle'_T = \left(\frac{H}{2}\right)^{2(n-1)} \prod_{v\in V(T)} N^{(d_v)} \mathcal{S}_T(k_1,\ldots,k_n), \tag*{(31)}

where the prime indicates that the overall momentum conserving delta function has been omitted. Here, ST(k1,…,kn)\mathcal{S}_T(k_1,\ldots,k_n) represents the shape function which will be studied in § 5. Our goal in this section is to concentrate on the amplitude of the correlation function.

As in the cases of bispectrum and trispectrum, we can define the amplitude of the nn-point function associated with the tree TT with the degree sequence (d1,…,dn)(d_1,\ldots,d_n). Noting that PRP_{\mathcal{R}} scales like N′2PδϕN^{'2}P_{\delta\phi}, from Eq. (31) we define the amplitude fTf_T associated to that shape function as follows,

fT≡∏v∈V(T)N(dv)(N′)2(n−1).(32)f_T \equiv\frac{\prod_{v\in V(T)}N^{(d_v)}}{(N')^{2(n-1)}}. \tag*{(32)}

In addition, it is convenient to introduce,

ad≡N(d)(N′)d,a1=1.(33)a_d \equiv\frac{N^{(d)}}{(N')^d}, \qquad a_1=1. \tag*{(33)}

We refer to ada_d as the normalized vertex amplitude associated with a vertex of degree dd. The normalization ensures that a1=1a_1=1, so that leaves contribute a unit factor to the tree amplitude.

Using the graph-theoretic degree sum relation Eq. (30), the tree amplitude takes the factorized form,

fT=∏v∈V(T)adv.(34)f_T = \prod_{v\in V(T)} a_{d_v}. \tag*{(34)}

This is a fundamental result of our graph-theoretic formulation: the amplitude of any connected tree diagram factorizes into the product of the vertex amplitudes associated with its vertices.

For a tree with degree sequence (d1,…,dn)(d_1,\ldots,d_n), we may equivalently denote its amplitude fTf_T by f(d1,…,dn)f_{(d_1,\ldots,d_n)}. For example, comparing with the bispectrum and trispectrum amplitudes introduced in § 2, we have in particular,

f(2,1,1)=FNL=a2,(35)f_{(2,1,1)}=F_{NL}=a_2, \tag*{(35)}

and

f(3,1,1,1)=GNL=a3,f(2,2,1,1)=TNL=a22.(36)f_{(3,1,1,1)}=G_{NL}=a_3, \qquad f_{(2,2,1,1)}=T_{NL}=a_2^2. \tag*{(36)}

An important consequence of Eq. (34) is that the tree amplitude depends only on the degree sequence of the tree. Therefore, two non-isomorphic trees with the same degree sequence have identical amplitudes, even though their momentum dependent shape functions may differ.

Leaf gluing and amplitude factorization

We now introduce an operation that relates the amplitudes of trees at different orders. Consider two disjoint trees T1T_1 and T2T_2, each containing at least three vertices. Choose one leaf in each tree and remove the selected leaves together with their incident edges. We then connect their former neighbouring vertices by a new edge. We refer to this operation as leaf gluing. The resulting graph is again a tree. Furthermore, the degrees of all surviving vertices remain unchanged: each of the two neighbouring vertices loses one edge and gains one new edge. This leads to the following theorem.

Theorem 4 (Leaf-gluing factorization). Let T1T_1 and T2T_2 be two trees with m≥3m \ge3 and n≥3n \ge3 vertices respectively. Let T3T_3 be the tree obtained by leaf gluing T1T_1 and T2T_2. Then T3T_3 contains m+n−2m+n-2 vertices, and its amplitude satisfies,

fT3=fT1fT2.(37)f_{T_3}=f_{T_1}f_{T_2}. \tag*{(37)}

Proof. By construction, the two removed leaves have degree one and hence contribute factors a1=1a_1=1 to the amplitudes of the original trees. All surviving vertices preserve their degrees. Consequently, the product of vertex factors in T3T_3 is precisely the product of the vertex factors in T1T_1 and T2T_2. Applying Eq. (34), we obtain

fT3=∏v∈V(T3)adv=(∏v∈V(T1)adv)(∏v∈V(T2)adv)=fT1fT2.(38)f_{T_3}=\prod_{v\in V(T_3)}a_{d_v}=\left(\prod_{v\in V(T_1)}a_{d_v}\right)\left(\prod_{v\in V(T_2)}a_{d_v}\right)=f_{T_1}f_{T_2}. \tag*{(38)}

This completes the proof.

The above theorem establishes an exact factorization relation among tree amplitudes at different orders. The simplest example reproduces the familiar single field SY equality.

In Figure 3, we have illustrated the leaf-gluing operation with two examples. In both cases, the red vertices represent the leaves selected for removal. The operation preserves the degrees of all remaining vertices, and consequently the amplitude of the resulting tree is the product of the original amplitudes. In the first example, for the four-point function, the path tree P4P_4 is obtained by gluing two bispectrum trees, each with degree sequence (2,1,1)(2,1,1). Applying the above theorem yields,

Examples of the leaf-gluing operation and the factorization of tree amplitudes

Figure 3. Examples of the leaf-gluing operation and the factorization of tree amplitudes. The red vertices denote the leaves selected for removal. In each case, the two selected leaves and their incident edges are removed, and their former neighbouring vertices are connected by a new edge. The degrees of all surviving vertices are preserved. (a) Gluing two bispectrum trees with degree sequence (2,1,1)(2,1,1) produces the four-point path tree P4P_4, with degree sequence (2,2,1,1)(2,2,1,1). The corresponding amplitude satisfies fP4=FNL2=a22=TNLf_{P_4}=F_{\mathrm{NL}}^2=a_2^2=T_{\mathrm{NL}}, reproducing the SY equality. (b) Gluing the four-point star S4S_4, with degree sequence (3,1,1,1)(3,1,1,1), and a bispectrum tree produces the five-point tree with degree sequence (3,2,1,1,1)(3,2,1,1,1). Its amplitude satisfies fT5=GNLFNL=a3a2f_{T_5}=G_{\mathrm{NL}}F_{\mathrm{NL}}=a_3a_2.

TNL=f(2,2,1,1)=f(2,1,1)2=FNL2=a22.(39)T_{NL}=f_{(2,2,1,1)}=f_{(2,1,1)}^2=F_{NL}^2=a_2^2. \tag*{(39)}

This reproduces the familiar SY equality in our normalization.

As the second example, consider the five-point tree with degree sequence (3,2,1,1,1)(3,2,1,1,1), shown in the lower panel of Figure 3. It can be constructed by gluing the four-point star S4S_4 and a bispectrum tree. Its amplitude therefore satisfies,

f(3,2,1,1,1)=f(3,1,1,1)f(2,1,1)=a2a3=GNLFNL.(40)f_{(3,2,1,1,1)}=f_{(3,1,1,1)}f_{(2,1,1)}=a_2a_3=G_{NL}F_{NL}. \tag*{(40)}

These examples demonstrate how the amplitudes of higher order tree diagrams can be expressed in terms of lower order amplitudes. Theorem 4 extends this factorization structure to arbitrary tree order.

Primary diagrams.

The leaf gluing operation allows us to construct all non-star trees from lower order trees. However, the star diagram is the unique exception. To see this, consider a tree TT that is not a star. Such a tree contains at least one edge connecting two vertices of degree greater than one. Removing this edge separates the tree into two connected components. By attaching a new leaf to each of the two exposed endpoints, we obtain two smaller trees, each containing at least three vertices. The original tree can then be reconstructed by leaf gluing these two components. In contrast, every edge of a star is incident to a leaf. Therefore, the star cannot be decomposed into two smaller trees by reversing a nontrivial leaf gluing operation.

We refer to the star tree SnS_n as the primary diagram. It has degree sequence (n−1,1,…,1)(n-1,1,\ldots,1), and its amplitude is,

fSn=f(n−1,1,…,1)=N(n−1)(N′)n−1=an−1.(41)f_{S_n}=f_{(n-1,1,\ldots,1)}=\frac{N^{(n-1)}}{(N')^{n-1}}=a_{n-1}. \tag*{(41)}

In comparison, the path tree PnP_n, containing n−2n-2 vertices of degree two and two leaves, has the amplitude fPn=a2n−2f_{P_n}=a_2^{n-2}.

Since every non-primary tree can be obtained by repeated leaf gluing, Eq. (37) implies that its amplitude can be expressed as a product of primary amplitudes. More explicitly, let md(T)m_d(T) denotes the number of degree-dd vertices in a tree TT. Then, Eq. (34) gives,

fT=∏d=2n−1admd(T)=∏d=2n−1(fSd+1)md(T).(42)f_T = \prod_{d=2}^{n-1} a_d^{m_d(T)} = \prod_{d=2}^{n-1} \left(f_{S_{d+1}}\right)^{m_d(T)} . \tag*{(42)}

The degree-one vertices do not contribute because a1=1a_1 = 1.

Equation (4.13) is the central factorization result of this section. It shows that the amplitudes of all connected tree diagrams are determined by the amplitudes of the primary diagrams. The corresponding results for 3≤n≤73 \le n \le7 are summarized in Table 1. Notice that different tree topologies can have the same degree sequence and hence the same amplitude. However, the factorization theorem does not identify their momentum dependent shape functions, which generally depend on the full topology.

nntree amplitudesProducts of primary amplitudes
3f(2,1,1)f_{(2,1,1)}a2a_2
4f(3,1,1,1)f_{(3,1,1,1)}, f(2,2,1,1)f_{(2,2,1,1)}a3a_3, a22a_2^2
5f(4,1,1,1,1)f_{(4,1,1,1,1)}, f(3,2,1,1,1)f_{(3,2,1,1,1)}, f(2,2,2,1,1)f_{(2,2,2,1,1)}a4a_4, a3a2a_3a_2, a23a_2^3
6f(5,1,1,1,1,1)f_{(5,1,1,1,1,1)}, f(4,2,1,1,1,1)f_{(4,2,1,1,1,1)}, f(3,3,1,1,1,1)f_{(3,3,1,1,1,1)}, f(3,2,2,1,1,1)f_{(3,2,2,1,1,1)}, f(2,2,2,2,1,1)f_{(2,2,2,2,1,1)}a5a_5, a4a2a_4a_2, a32a_3^2, a3a22a_3a_2^2, a24a_2^4
7f(6,1,1,1,1,1,1)f_{(6,1,1,1,1,1,1)}, f(5,2,1,1,1,1,1)f_{(5,2,1,1,1,1,1)}, f(4,3,1,1,1,1,1)f_{(4,3,1,1,1,1,1)}, f(4,2,2,1,1,1,1)f_{(4,2,2,1,1,1,1)}, f(3,3,2,1,1,1,1)f_{(3,3,2,1,1,1,1)}, f(3,2,2,2,1,1,1)f_{(3,2,2,2,1,1,1)}, f(2,2,2,2,2,1,1)f_{(2,2,2,2,2,1,1)}a6a_6, a5a2a_5a_2, a4a3a_4a_3, a4a22a_4a_2^2, a32a2a_3^2a_2, a3a23a_3a_2^3, a25a_2^5

Table 1. Factorization of tree amplitudes into primary amplitudes for 3≤n≤73 \le n \le7. Each degree sequence is listed once. Distinct tree topologies may have the same degree sequence and hence the same amplitude.

Counting distinct tree amplitudes

Having established the factorization of tree amplitudes into primary amplitudes, we now compare the number of distinct amplitude products with the number of independent tree topologies.

Let Namp(n)N_{\mathrm{amp}}(n) denote the number of algebraically distinct products of the primary amplitudes ada_d at order nn. Intuitively speaking, Namp(n)N_{\mathrm{amp}}(n) counts the number of distinct products of the primary amplitudes ada_d that can be formed at a given order nn, subject to the degree-sum constraint. This should not be confused with the number of available nontrivial primary amplitudes ada_d, which is simply n−2n - 2. According to Eq. (42), each tree amplitude is specified by the nonnegative integers md(T)m_d(T), which count the number of vertices of degree dd. The degree-sum relation Eq. (17) then implies,

∑d=2n−1(d−1)md=n−2.(43)\sum_{d=2}^{n-1}(d-1)m_d=n-2. \tag*{(43)}

Thus, each primary amplitude ada_d contributes a weight d−1d-1, and the distinct amplitude products Namp(n)N_{\mathrm{amp}}(n) correspond to the nonnegative integer solutions of this constraint. These solutions are in one-to-one correspondence with the unrestricted integer partitions of n−2n-2. As an example, consider n=6n=6. Then Eq. (43) gives m2+2m3+3m4+4m5=4m_2+2m_3+3m_4+4m_5=4. Solving the possible integer solutions yields (m5=1,mi=0)(m_5=1,m_i=0); (m4=1,m2=1,mi=0)(m_4=1,m_2=1,m_i=0); (m3=2,mi=0)(m_3=2,m_i=0); (m3=1,m2=2,mi=0)(m_3=1,m_2=2,m_i=0); (m2=4,mi=0)(m_2=4,m_i=0) so in total there are five distinct solutions and therefore Namp(6)=5N_{\mathrm{amp}}(6)=5.

Since every admissible degree sequence is realized by at least one tree, as established in Theorem 2, we therefore obtain,

Namp(n)=p(n−2),(44)N_{\mathrm{amp}}(n)=p(n-2), \tag*{(44)}

where p(m)p(m) is the integer partition function, denoting the number of unrestricted integer partitions of mm. The results for 3≤n≤103\le n\le10 are summarized in Table 2. The distinction first appears at n=6n=6, where six non-isomorphic trees give only five distinct amplitude products as we demonstrated above. This difference increases rapidly with nn.

nnT(n)T(n)Namp(n)N_{\mathrm{amp}}(n)T(n)−Namp(n)T(n)-N_{\mathrm{amp}}(n)
3110
4220
5330
6651
71174
8231112
9471532
101062284

Table 2. Comparison between the number T(n)T(n) of non-isomorphic trees and the number Namp(n)N_{\mathrm{amp}}(n) of distinct primary amplitude products. The difference in the fourth column counts the excess of tree topologies over the number of distinct primary amplitude products.

The large-nn behaviour of Namp(n)N_{\mathrm{amp}}(n) follows from the Hardy–Ramanujan asymptotic formula [65],

Namp(n)∼143(n−2)exp⁡[π2(n−2)3].(45)N_{\mathrm{amp}}(n)\sim\frac{1}{4\sqrt{3}(n-2)}\exp\left[\pi\sqrt{\frac{2(n-2)}{3}}\right]. \tag*{(45)}

Comparing this expression with Otter’s asymptotic formula (29), we obtain

lim⁡n→∞Namp(n)T(n)=0.(46)\lim_{n\to\infty}\frac{N_{\mathrm{amp}}(n)}{T(n)}=0. \tag*{(46)}

Thus, although the number of independent momentum shapes T(n)T(n) grows exponentially with nn, the number of distinct amplitude products grows considerably more slowly. The important conclusion is that at high orders, many different tree topologies therefore share the same amplitude.

This completes our discussion of the tree amplitudes and their factorization. In the next section, we turn to the momentum dependent shape functions associated with individual tree topologies.

Momentum Shapes in Fourier Space

In the previous section we have shown that the amplitude of a tree depends only on its vertex degrees. Now, as defined in Eq. (31), we consider the momentum dependent shape functions, which retain information about the full connectivity of the tree. This distinction first becomes particularly important at n=6n = 6 where T(n)≠Namp(n)T_{(n)} \ne N_{\mathrm{amp}}(n) so non-isomorphic trees can have the same degree sequence.

Consider a tree TT with nn vertices carrying external momenta k1,…,kn\mathbf{k}_1,\ldots,\mathbf{k}_n, with ∑i=1nki=0\sum_{i=1}^{n}\mathbf{k}_i = 0. Removing an edge ee, partitions its vertices into two connected sets AeA_e and AecA_e^c. The momentum magnitude carried by that edge is,

qe=∣∑i∈Aeki∣=∣∑i∈Aecki∣.(47)q_e = \left|\sum_{i\in A_e}\mathbf{k}_i\right| = \left|\sum_{i\in A_e^c}\mathbf{k}_i\right|. \tag*{(47)}

In particular, an edge incident to a leaf labeled ii carries qe=kiq_e = k_i.

For a fixed assignment of external labels, the momentum dependence is the product of the n−1n-1 edge propagators. Summing over distinct assignments gives the shape function as follows,

ST(k1,…,kn)=∑inequivalent labelings∏e∈E(T)1qe3.(48)S_T(\mathbf{k}_1,\ldots,\mathbf{k}_n) = \sum_{\text{inequivalent labelings}} \prod_{e\in E(T)} \frac{1}{q_e^3}. \tag*{(48)}

Here E(T)E(T) denotes the edge set of the tree TT, containing n−1n-1 edges. Labelings which are related by an automorphism that preserves the graph connectivity are equivalent and are counted only once. Consequently, the number of inequivalent labelings in a tree diagram is n!/∣Aut⁡(T)∣n!/|\operatorname{Aut}(T)|. As a simple example, consider the five-point star S5S_5, whose central vertex has degree four and whose four leaves have degree one. Its edge set is E(S5)={(1,2),(1,3),(1,4),(1,5)}E(S_5) = \{(1,2),(1,3),(1,4),(1,5)\}, for a representative labeling with vertex 1 at the center. Since the four leaves can be permuted arbitrarily, the automorphism group has order ∣Aut⁡(S5)∣=4!=24|\operatorname{Aut}(S_5)| = 4! = 24. Therefore, there are only 5!/4!=55!/4! = 5 inequivalent assignments of the external momenta, corresponding to the five possible choices of the central vertex. The resulting shape function is given in Eq. (49) below, while its amplitude is fS5=a4f_{S_5} = a_4.

In contrast, the amplitude fTf_T is given by Eq. (34) as established in § 4. The complete tree contribution carries the corresponding factor fTSTf_T S_T.

The associated shapes for the bispectrum and trispectrum were given in § 2, here we concentrate on the shape functions for n=5n = 5 and n=6n = 6.

Five-point functions

As we have seen previously, there are three non-isomorphic trees with five vertices: the star S5S_5, the path P5P_5, and the branched tree with degree sequence (3,2,1,1,1)(3,2,1,1,1). These are the three graphs already displayed in the last column in Figure 1.

For compactness, let us define ki1⋯ir≡∣ki1+⋯+kir∣k_{i_1\cdots i_r} \equiv\lvert k_{i_1}+\cdots+k_{i_r}\rvert. The following expressions give one representative labeling for the total n!/∣Aut⁡(T)∣n!/\lvert\operatorname{Aut}(T)\rvert labeling of each topology, followed by the remaining distinct permutations of the external labels.

For the star with center 1, the edge momenta are k2k_2, k3k_3, k4k_4, k5k_5, so

S5(a)=(1k23k33k43k53+4 permutations),f5(a)=a4.(49)S_5^{(a)}=\left(\frac{1}{k_2^3 k_3^3 k_4^3 k_5^3}+4\ \text{permutations}\right),\qquad f_5^{(a)}=a_4. \tag*{(49)}

For the path 1−2−3−4−51-2-3-4-5, the edge momenta are k1k_1, k12k_{12}, k45k_{45}, k5k_5, giving,

S5(b)=(1k13k123k453k53+59 permutations),f5(b)=a23.(50)S_5^{(b)}=\left(\frac{1}{k_1^3 k_{12}^3 k_{45}^3 k_5^3}+59\ \text{permutations}\right),\qquad f_5^{(b)}=a_2^3. \tag*{(50)}

Finally, for the edges (1,2),(1,3),(1,4),(4,5)(1,2),(1,3),(1,4),(4,5), the edge momenta are k2k_2, k3k_3, k5k_5, k45k_{45}, and hence,

S5(c)=(1k23k33k53k453+59 permutations),f5(c)=a3a2.(51)S_5^{(c)}=\left(\frac{1}{k_2^3 k_3^3 k_5^3 k_{45}^3}+59\ \text{permutations}\right),\qquad f_5^{(c)}=a_3a_2. \tag*{(51)}

The three topologies therefore give three distinct momentum shapes with their amplitudes in each case given above.

Six-point functions

For n=6n=6 there are six non-isomorphic trees N(6)=6N(6)=6, but only five distinct degree sequences, Namp(6)=5N_{\mathrm{amp}}(6)=5. Figure 4 displays all six trees with a representative assignment of the external labels. The panel letters (a)–(f) agree with the shape functions given below. In particular, panels (a) and (b) have the same degree sequence (3,2,2,1,1,1)(3,2,2,1,1,1) but different edge connectivity.

Six non-isomorphic trees associated to $n=6$

Figure 4. The six non-isomorphic trees associated to n=6n = 6. Panels (a)(a) and (b)(b) have identical degree sequences (3,2,2,1,1,1)(3, 2, 2, 1, 1, 1) and identical tree amplitudes, but different momentum shapes. The remaining panels are respectively the degree (4,2,1,1,1,1)(4, 2, 1, 1, 1, 1) tree, the double-star (3,3,1,1,1,1)(3, 3, 1, 1, 1, 1), the path (2,2,2,2,1,1)(2, 2, 2, 2, 1, 1), and the star (5,1,1,1,1,1)(5, 1, 1, 1, 1, 1).

Reading the edge partitions directly from the figure, we obtain

S6(a)=(1k13k53k63k453k1263+359 permutations),f6(a)=a3a22.(52)S_6^{(a)}=\left(\frac{1}{k_1^3 k_5^3 k_6^3 k_{45}^3 k_{126}^3}+359\ \text{permutations}\right),\qquad f_6^{(a)}=a_3a_2^2. \tag*{(52)}
S6(b)=(1k13k53k63k123k453+359 permutations),f6(b)=a3a22.(53)S_6^{(b)}=\left(\frac{1}{k_1^3 k_5^3 k_6^3 k_{12}^3 k_{45}^3}+359\ \text{permutations}\right),\qquad f_6^{(b)}=a_3a_2^2. \tag*{(53)}
S6(c)=(1k13k43k53k63k123+119 permutations),f6(c)=a4a2.(54)S_6^{(c)}=\left(\frac{1}{k_1^3 k_4^3 k_5^3 k_6^3 k_{12}^3}+119\ \text{permutations}\right),\qquad f_6^{(c)}=a_4a_2. \tag*{(54)}
S6(d)=(1k13k43k53k63k1263+89 permutations),f6(d)=a32.(55)S_6^{(d)}=\left(\frac{1}{k_1^3 k_4^3 k_5^3 k_6^3 k_{126}^3}+89\ \text{permutations}\right),\qquad f_6^{(d)}=a_3^2. \tag*{(55)}
S6(e)=(1k53k63k1263k123k453+359 permutations),f6(e)=a24.(56)S_6^{(e)}=\left(\frac{1}{k_5^3 k_6^3 k_{126}^3 k_{12}^3 k_{45}^3}+359\ \text{permutations}\right),\qquad f_6^{(e)}=a_2^4. \tag*{(56)}
S6(f)=(1k13k23k33k43k53+5 permutations),f6(f)=a5.(57)S_6^{(f)}=\left(\frac{1}{k_1^3 k_2^3 k_3^3 k_4^3 k_5^3}+5\ \text{permutations}\right),\qquad f_6^{(f)}=a_5. \tag*{(57)}

The equality f6(a)=f6(b)f_6^{(a)}=f_6^{(b)} illustrates the difference between the amplitude and the shape. The two graphs have identical vertex degrees, but their internal edges produce different momentum partitions. Consequently, their momentum dependence is distinct even though their amplitudes are equal. More generally, the classification by tree topology determines the distinct momentum structures, while the factorization rule of § 4 determines their amplitudes.

Examples of Single-Field Inflation

The graph-theoretic analysis developed in the previous sections were general and applicable to all single field models of inflation employing δN\delta N formalism. In this section, we study USR inflation as an interesting setup where the previous results can be applied.

USR Inflation with Transition to SR

The USR setup is a single field model of inflation in which the potential is exactly flat V=V0V = V_0 [28, 66, 67, 68, 69]. As a result, the first SR parameter scales like ϵ∝a(t)−6\epsilon\propto a(t)^{-6} while the second SR parameter is large, η=−6\eta= -6. Since ϵ\epsilon falls off exponentially, the curvature perturbation grows on superhorizon scales like R∝a(t)3\mathcal{R} \propto a(t)^3. This is the key reason why the USR setup violates the Maldacena consistency condition, predicting fNL=5/2f_{\mathrm{NL}} = 5/2 [28]. Since the curvature perturbations grow exponentially during the USR phase, the system becomes non-perturbative if the duration of USR phase is unlimited. Therefore, one has to terminate the USR phase after which the system enters a SR phase. The transition from the non-attractor USR phase to the final attractor SR phase can be either sharp or mild. The original USR model studied in [28, 70] was an idealized setup in which the transition from a USR phase to the attractor phase is infinitely sharp so the system enters the attractor phase immediately after the transition. However, in a realistic setup one expects that the transition from the non-attractor phase to the attractor phase to take some time. This indeed is controlled by a dimensionless parameter, hh, which is defined as [26],

h≡62ϵVϕ˙(te)MPH=−6ϵVϵe,(58)h \equiv\frac{6\sqrt{2\epsilon_{\mathrm{V}}}}{\dot{\phi}(t_{\mathrm{e}})}M_{\mathrm{P}}H=-6\sqrt{\frac{\epsilon_{\mathrm{V}}}{\epsilon_{\mathrm{e}}}}, \tag*{(58)}

in which ϵe\epsilon_{\mathrm{e}} is the value of ϵ\epsilon at the end of USR phase t=tet=t_{\mathrm{e}} while ϵV\epsilon_{\mathrm{V}} is its value during the final attractor phase. Here, without loss of generality, we have assumed that ϕ˙<0\dot{\phi}<0 so h<0h<0. The infinitely sharp transition considered in [28, 70] corresponds to h→−∞h\to-\infty.

During the USR phase the system is in a non-attractor phase so the evolution of the system should be considered in the full phase space in which N=N(ϕ,ϕ˙)N=N(\phi,\dot{\phi}). This is the important difference between δN\delta N formalism in USR setup compared to the standard SR setups in which the system is in the attractor phase and N=N(ϕ)N=N(\phi).

In this subsection we consider the setup in which the potential is continuous across the transition from USR to the SR phase but its derivative has a jump. In the next subsection we relax the first assumption as well in which the potential itself has a jump (step) across the transition.

The total number of e-fold N(ϕ,ϕ˙)N(\phi,\dot{\phi}) was calculated in [26], yielding,

Ntot=NUSR+NSR=−13ln⁡[1+3(ϕ−ϕe)π]+1ηVln⁡[−2ηVπe−62ϵVMP]+constant,(59)N_{\mathrm{tot}}=N_{\mathrm{USR}}+N_{\mathrm{SR}}=-\frac{1}{3}\ln\left[1+\frac{3(\phi-\phi_{\mathrm{e}})}{\pi}\right]+\frac{1}{\eta_{\mathrm{V}}}\ln\left[-2\eta_{\mathrm{V}}\pi_{\mathrm{e}}-6\sqrt{2\epsilon_{\mathrm{V}}}M_{\mathrm{P}}\right]+\text{constant}, \tag*{(59)}

in which NUSRN_{\mathrm{USR}} and NSRN_{\mathrm{SR}} are the contributions in total number of e-folds during the USR phase and the second SR phase respectively while the constant piece is independent of the initial values in phase space. Furthermore, π≡ϕ˙(N)=2ϵMP\pi\equiv\dot{\phi}(N)=\sqrt{2\epsilon}M_{\mathrm{P}} and (ϕe,πe)(\phi_{\mathrm{e}},\pi_{\mathrm{e}}) are the values of the field and its momentum at the end of USR. Also, ηV≪1\eta_{\mathrm{V}}\ll1 is the value of the second SR parameter (related to second derivative of V(ϕ)V(\phi)) during the final attractor phase. In addition, the following constraint between (πe,ϕe)(\pi_{\mathrm{e}},\phi_{\mathrm{e}}) and the initial values (ϕ,π)(\phi,\pi) holds,

πe=π+3(ϕ−ϕe).(60)\pi_{\mathrm{e}}=\pi+3(\phi-\phi_{\mathrm{e}}). \tag*{(60)}

When employing δN\delta N formalism, it is important to realize that the variations are with respect to initial values, (ϕ,π)(\phi,\pi) while the position of the end of USR is fixed in the field space at ϕ=ϕe\phi=\phi_{\mathrm{e}}. On the other hand, πe\pi_{\mathrm{e}} is not fixed in phase space and is related to the initial values (ϕ,π)(\phi,\pi) and ϕe\phi_{\mathrm{e}} via Eq. (60). Finally, after taking the variation, only δϕ\delta\phi is relevant in δN\delta N expansion since δπ\delta\pi is negligible as it approaches zero for light fields on superhorizon scales.

We are interested in the limit ηV→0\eta_{\mathrm{V}}\to0 in which the final phase of inflation is in the attractor regime as in conventional SR models. Taking the variation of Eq. (59) with respect to the initial field value, we obtain

δN=6hδϕπe−13ln⁡(1+3δϕπe).(61)\delta N=\frac{6}{h}\frac{\delta\phi}{\pi_{\mathrm{e}}}-\frac{1}{3}\ln\left(1+\frac{3\delta\phi}{\pi_{\mathrm{e}}}\right). \tag*{(61)}

Note that hh and πe\pi_{\mathrm{e}} depends on the initial data in the phase space. In particular, πe\pi_{\mathrm{e}} is related to the initial data via the constraint Eq. (60).

Expanding δN\delta N in Eq. (61) formally in powers of δϕ\delta\phi via,

δN=∑d=1∞N(d)d!(δϕ)d,(62)\delta N=\sum_{d=1}^{\infty}\frac{N^{(d)}}{d!}(\delta\phi)^d, \tag*{(62)}

we find,

N′=−1γπe,N(d)=(−1)d3d−1(d−1)!πed,d≥2,(63)N' = -\frac{1}{\gamma\pi_e}, \qquad N^{(d)} = \frac{(-1)^d 3^{d-1}(d-1)!}{\pi_e^d}, \qquad d \ge2, \tag*{(63)}

where for convenience we have defined,

γ=hh−6,0<γ<1.(64)\gamma= \frac{h}{h-6}, \qquad0 < \gamma< 1. \tag*{(64)}

Using the definition of normalized vertex amplitudes ada_d introduced in § 4, the above expressions yield,

a1=1,ad=N(d)(N′)d=3d−1(d−1)!γd,d≥2.(65)a_1 = 1, \qquad a_d = \frac{N^{(d)}}{(N')^d} = 3^{d-1}(d-1)!\gamma^d, \qquad d \ge2. \tag*{(65)}

Since γ>0\gamma> 0, an immediate conclusion from the above result is that ad>0a_d > 0 for all d≥2d \ge2.

The above expression determines the complete hierarchy of normalized vertices for the USR to SR transition. In particular, the amplitude of bispectrum is obtained to be,

FNL=a2=3h2(h−6)2=3γ2.(66)F_{\mathrm{NL}} = a_2 = \frac{3h^2}{(h-6)^2} = 3\gamma^2. \tag*{(66)}

We realize that the maximum value of FNLF_{\mathrm{NL}} is obtained in the limit of an infinitely sharp transition h→−∞h \to-\infty. Taking the conventional factor 5/65/6 into account, in this limit we obtain fNL=5FNL/6=5/2f_{\mathrm{NL}} = 5F_{\mathrm{NL}}/6 = 5/2 as obtained originally in [28]. On the other hand, in mild transitions with ∣h∣≲1\lvert h\rvert\lesssim1, much of the non-Gaussianity accumulated during the USR phase is washed out during the subsequent attractor phase as observed in [26].

We can now apply the graph-theoretic results of the previous sections. As we have seen, for a tree TT with nn vertices and vertex degrees dvd_v, the amplitude fTf_T of the tree is given by the products of adva_{d_v} as given in Eq. (34). Now, let L(T)L(T) denotes the number of leaves of the tree. Since the leaves have a1=1a_1 = 1, using Eq. (65) and the tree identities,

∑vdv=2(n−1),∑v: dv≥2(dv−1)=n−2,(67)\sum_v d_v = 2(n-1), \qquad\sum_{v:\,d_v\ge2}(d_v-1) = n-2, \tag*{(67)}

the amplitude fTf_T from Eq. (34) is obtained to be,

fT=3n−2γ2n−2−L(T)∏v∈V(T)(dv−1)!.(68)f_T = 3^{n-2}\gamma^{2n-2-L(T)} \prod_{v\in V(T)}(d_v-1)!. \tag*{(68)}

This is the general result for the amplitude of an arbitrary tree in the USR to SR model. An interesting consequence is that the transition parameter affects different tree topologies differently. At a fixed number of vertices nn, the power of γ\gamma depends on the number of leaves, while the remaining combinatorial factor depends on the vertex degrees. Finally, we see that the amplitudes are positive for all shapes. This is a strong and interesting result, extending the result of [71] established for the particular case dv=2d_v = 2.

As an example, consider the trispectrum n=4n = 4. For the star tree with the degree sequence (3,1,1,1)(3,1,1,1) we obtain,

f(3,1,1,1)=GNL=18h3(h−6)3.(69)f_{(3,1,1,1)} = G_{\mathrm{NL}} = \frac{18h^3}{(h-6)^3}. \tag*{(69)}

On the other hand, for the path tree with the degree sequence (2,2,1,1)(2,2,1,1) corresponding to TNLT_{NL}, we obtain

f(2,2,1,1)=9h4(h−6)4.(70)f_{(2,2,1,1)}=\frac{9h^4}{(h-6)^4}. \tag*{(70)}

Both of the above results agree with the results of [72].

As the second example, let us compare the amplitudes of the star graph SnS_n and path graph PnP_n for a general value of nn. The star graph contains n−1n-1 leaves and one vertex of degree n−1n-1, so its amplitude is,

fSn=3n−2(n−2)!γn−1.(71)f_{S_n}=3^{n-2}(n-2)!\gamma^{n-1}. \tag*{(71)}

On the other hand, the path graph PnP_n contains only two leaves, and its amplitude is,

fPn=3n−2γ2n−4.(72)f_{P_n}=3^{n-2}\gamma^{2n-4}. \tag*{(72)}

As a result, their ratio is given by,

fSnfPn=(n−2)!γ3−n.(73)\frac{f_{S_n}}{f_{P_n}}=(n-2)!\gamma^{3-n}. \tag*{(73)}

For a mild transition, γ≪1\gamma\ll1, the path amplitude is more strongly suppressed than the star amplitude. The difference becomes increasingly pronounced at higher orders. On the other hand, in the infinitely sharp transition limit, γ→1\gamma\to1, the relative amplitudes reduce to purely combinatorial factors.

More generally, for any two trees T1T_1 and T2T_2 with the same number of vertices nn, Eq. (68) gives

fT1fT2=γL(T2)−L(T1)∏v∈V(T1)(dv−1)!∏w∈V(T2)(dw−1)!.(74)\frac{f_{T_1}}{f_{T_2}}=\gamma^{L(T_2)-L(T_1)}\frac{\prod_{v\in V(T_1)}(d_v-1)!}{\prod_{w\in V(T_2)}(d_w-1)!}. \tag*{(74)}

In particular, we see that trees with the same degree sequence have identical amplitudes even when their momentum shapes are different. As we have observed in previous Section, this phenomenon first occurs in the six-point functions.

The factorization identities derived in § 4 also hold for this model. For example, one can check that all the factorization products listed in Table 1 apply here as well. For example, for the five-point amplitudes we have the following relations, f(3,2,1,1,1)=a3a2f_{(3,2,1,1,1)}=a_3a_2 and f(2,2,2,1,1)=a23f_{(2,2,2,1,1)}=a_2^3 and so on for higher order correlations. These relations follow directly from the graph structure and do not require additional dynamical assumptions. The USR dynamics enters through the specific values of the vertices ada_d in Eq. (65).

USR Transition with a Potential Step

We now extend the preceding analysis to the setup in which the potential has a jump ΔV\Delta V across the transition. Following the analysis of Ref. [17], the transition is characterized by two parameters, hh and gg in which hh as before measures the sharpness of the transition while gg measures the discontinuity in the potential across the transition. As we shall see, the additional parameter gg changes the hierarchy of nonlinear vertices and introduces the possibility of cancellations among different contributions. These cancellations have direct consequences for the tree amplitudes and their momentum shapes.

Let πe\pi_e and πd\pi_d denote the inflaton velocities immediately before and after the potential step, respectively. As in previous example, we assume π<0\pi< 0 and the constraint Eq. (60) does hold. Assuming that the Hubble parameter is approximately constant across the transition, from the energy conservation we obtain [17],

πd=−πe2−6MPl2ΔVV0,(75)\pi_d=-\sqrt{\pi_e^2-\frac{6M_{\mathrm{Pl}}^2\Delta V}{V_0}}, \tag*{(75)}

where ΔV\Delta V is the change in potential energy across the step. Let us define,

g≡πdπe>0,h≡62ϵVMPlπd<0.(76)g\equiv\frac{\pi_d}{\pi_e}>0,\qquad h\equiv\frac{6\sqrt{2\epsilon_V}M_{\mathrm{Pl}}}{\pi_d}<0. \tag*{(76)}

An upward step with ΔV>0\Delta V>0 corresponds to 0<g<10<g<1, whereas a downward step with ΔV<0\Delta V<0 gives g>1g>1. The case g=1g=1 recovers the transition studied in previous subsection.

As before, the total number of e-folds contains the USR and SR contributions which is given by [17],

N(ϕ,π)=−13ln⁡(πeπ)+1ηVln⁡(−2ηVπd−62ϵVMPl)+constant.(77)N(\phi,\pi)=-\frac{1}{3}\ln\left(\frac{\pi_e}{\pi}\right)+\frac{1}{\eta_V}\ln\left(-2\eta_V\pi_d-6\sqrt{2\epsilon_V}M_{\mathrm{Pl}}\right)+\text{constant}. \tag*{(77)}

In calculating δN\delta N note that we take the variations with respect to δϕ\delta\phi and δπ\delta\pi while holding the point of transition ϕ=ϕe\phi=\phi_e and the point of end of inflation ϕ=ϕf\phi=\phi_f fixed. In particular, the parameters gg and hh are evaluated on the fiducial background only after performing these variations.

As in previous example, we are interested in the limit where ηV→0\eta_V\to0 so the second phase is in the SR regime. Neglecting the initial momentum fluctuation as before, the resulting curvature perturbation is given by,

R=δN=−13ln⁡(1+3δϕπe)+2h[1+6δϕg2πe+9(δϕ)2g2πe2−1].(78)\mathcal{R}=\delta N=-\frac{1}{3}\ln\left(1+\frac{3\delta\phi}{\pi_e}\right)+\frac{2}{h}\left[\sqrt{1+\frac{6\delta\phi}{g^2\pi_e}+\frac{9(\delta\phi)^2}{g^2\pi_e^2}}-1\right]. \tag*{(78)}

The logarithmic contribution above originates from the USR evolution whereas the square root contribution reflects the change in velocity across the potential step and the subsequent SR evolution.

Expanding Eq. (78) in powers of δϕ\delta\phi and calculating the normalized vertex amplitudes ada_d as in Eq. (33), for the amplitudes relevant to n=6n=6, we obtain,

a2=3h[g4h+6(g2−1)](g2h−6)2,a3=18h2[g6h+9(g2−1)](g2h−6)3,a4=162h3[g8h−3(g2−1)(g2−5)](g2h−6)4,a5=486h4[4g10h−15(g2−1)(3g2−7)](g2h−6)5.(79)\begin{aligned} a_2&=\frac{3h\left[g^4h+6\left(g^2-1\right)\right]}{\left(g^2h-6\right)^2},\\ a_3&=\frac{18h^2\left[g^6h+9\left(g^2-1\right)\right]}{\left(g^2h-6\right)^3},\\ a_4&=\frac{162h^3\left[g^8h-3\left(g^2-1\right)\left(g^2-5\right)\right]}{\left(g^2h-6\right)^4},\\ a_5&=\frac{486h^4\left[4g^{10}h-15\left(g^2-1\right)\left(3g^2-7\right)\right]}{\left(g^2h-6\right)^5}. \tag*{(79)} \end{aligned}

One can check that for g=1g = 1, the above results reduce to ada_d in Eq. (65).

From the above results for ada_d we see that there can exist situations in which some amplitudes ada_d vanish. This corresponds to non-trivial competition between the parameters (h,g)(h, g). In our setup h<0h < 0, but depending on whether the jump in potential is upward (downward) then 0<g<10 < g < 1 (g>1g > 1). So let us consider these two cases in turn.

Positivity of Amplitudes for 0<g<10 < g < 1

We first consider an upward potential step, for which the inflaton loses kinetic energy when crossing the discontinuity, corresponding to 0<g<10 < g < 1. As can be checked explicitly from Eq. (79), the nonlinear vertices a2,…,a5a_2, \ldots, a_5 are positive in this case so there is no cancellation in the amplitudes. We have confirmed in Appendix § 9 that this conclusion does hold for all values of ada_d. Therefore,

ad>0,(0<g<1).(80)a_d > 0, \qquad(0 < g < 1). \tag*{(80)}

An immediate consequence follows from the graph-theoretic factorization relation Eq. (34). For any connected tree TT, its amplitude is positive for 0<g≤10 < g \le1,

fT=∏v∈V(T)adv>0,(0<g<1).(81)f_T = \prod_{v \in V(T)} a_{d_v} > 0, \qquad(0 < g < 1). \tag*{(81)}

This is a strong and interesting result. An upward potential step cannot generate a vanishing tree amplitude through the cancellation of nonlinear vertices at any order. The parameter gg modifies the relative magnitudes of different tree amplitudes, but not their signs.

Cancellation of Amplitudes for g>1g > 1

Now consider the case of a downward potential step, g>1g > 1. In this case the inflaton gains kinetic energy at the discontinuity, and the nonlinear vertices can receive contributions with opposite signs. Unlike the upward step case, the amplitudes ada_d can take either signs and in some situations they can vanish. This effect was observed in [73] for bispectrum and here we extend this study to higher orders. The details of the analysis where the cancellation of the amplitudes occur in the parameter space of (h,g)(h, g) are presented in Appendix § 9.

The first two cancellation curves follow immediately from Eq. (79), yielding,

a2=0:h=−6(q2−1)q4,a3=0:h=−9(q2−1)q6.(82)\begin{aligned} a_2 &= 0: \qquad h = -\frac{6(q^2 - 1)}{q^4}, \\ a_3 &= 0: \qquad h = -\frac{9(q^2 - 1)}{q^6}. \tag*{(82)} \end{aligned}

Both lie in the region h<0h < 0 for g>1g > 1. They intersect at,

g=32,h=−43,(83)g = \sqrt{\frac{3}{2}}, \qquad h = -\frac{4}{3}, \tag*{(83)}

where a2=a3=0a_2 = a_3 = 0. At this point, the tree level bispectrum and both of trispectrum amplitudes vanish (FNL=GNL=TNL=0F_{\mathrm{NL}} = G_{\mathrm{NL}} = T_{\mathrm{NL}} = 0), while higher order non Gaussianity can survive. This is a unique situation that can occur in the downward step transition. Of course, there can be subleading SR contributions in bispectrum and trispectrum which are not captured by our leading δN\delta N analysis. But their magnitudes are considerably small and may not be detectable observationally.

The graph-theoretic interpretation is particularly simple. Since the amplitude of each tree fTf_T is given by the products of vertex amplitudes ada_d as in Eq. (34), the condition ad=0a_d = 0 eliminates every tree containing at least one vertex of degree dd. This selection rule holds at arbitrary order and is independent of the momentum configuration.

As some case examples, consider first the curve a2=0a_2 = 0. The bispectrum amplitude vanishes, FNL=a2=0F_{\mathrm{NL}} = a_2 = 0. Similarly, the path trispectrum amplitude vanishes as well, f(2,2,1,1)=a22=0f_{(2,2,1,1)} = a_2^2 = 0. However, the star trispectrum amplitude f(3,1,1,1)=a3f_{(3,1,1,1)} = a_3 is generally nonzero. Therefore, as expected, the vanishing of bispectrum does not imply the absence of higher order non-Gaussianity. On the other hand, at five points, f(4,1,1,1,1)=a4f_{(4,1,1,1,1)} = a_4, f(3,2,1,1,1)=a3a2=0f_{(3,2,1,1,1)} = a_3a_2 = 0 and f(2,2,2,1,1)=a22=0f_{(2,2,2,1,1)} = a_2^2 = 0. As a result, only the star topology survives generically. Finally, at six points, both the star and double star topologies contain no a2a_2 vertices. Consequently, f(5,1,1,1,1,1)=a5f_{(5,1,1,1,1,1)} = a_5, f(3,3,1,1,1,1)=a32f_{(3,3,1,1,1,1)} = a_3^2 can survive while all remaining six-point tree amplitudes vanish on the curve of a2=0a_2 = 0.

As the second example, now consider the curve, a3=0a_3 = 0. The star trispectrum vanishes, whereas the path trispectrum generally survives, f(3,1,1,1)=0f_{(3,1,1,1)} = 0, f(2,2,1,1)=a22f_{(2,2,1,1)} = a_2^2. At five points, the mixed topology (3, 2, 1, 1, 1) disappears, while the star and path topologies generally remain. At six points, the two distinct topologies with degree sequence (3, 2, 2, 1, 1, 1) vanish simultaneously, despite their different momentum shapes. The double-star topology (3, 3, 1, 1, 1, 1) also vanishes. The remaining amplitudes are proportional to a5a_5, a4a2a_4a_2 and a24a_2^4.

At the intersection a2=a3=0a_2 = a_3 = 0, these selection rules combine. All three- and four-point tree amplitudes vanish, whereas the five- and six-point star amplitudes a4a_4 and a5a_5 remain nonzero. This illustrates how lower-order non-Gaussianity can be suppressed while higher order correlations are retained.

We conclude that downward potential steps admit a hierarchy of possible vertex zero curves. Whenever a vertex vanishes, entire families of tree topologies are eliminated according to their degree sequences. This provides a direct connection between the inflationary transition parameters and the graph theoretic organization of higher order correlation functions. Having said this, we comment that we have idealized the setup by an instant and sharp transition. One may consider a relaxed situation where the transition is smooth. Whether or not the cancellation of some topologies occur needs to be examined in these realistic situations.

Summary and Discussions

In this work we studied the δN\delta N formalism in general single field models of inflation and investigated the nn-point correlation functions using graph theory methods. A great advantage of the δN\delta N formalism is that one only needs to follow the background trajectory N=N(ϕ,ϕ˙)N = N(\phi,\dot{\phi}) in phase space. The curvature perturbation on superhorizon scales is related to the perturbation in the number of e-folds, which can be expanded perturbatively in terms of the initial field fluctuations. This provides a natural starting point for organizing higher order correlation functions using established results from graph theory.

Using graph theory, we studied connected tree diagrams and classified their independent, non-isomorphic topologies. Each vertex is characterized by its degree, and the corresponding normalized vertex amplitude is ad=N(d)/(N′)da_d = N^{(d)}/(N')^d, with a1=1a_1 = 1. We showed that the amplitude of an arbitrary tree TT factorizes as fT=∏v∈V(T)advf_T = \prod_{v \in V(T)} a_{d_v}. This result is independent of the details of the inflationary model and allows the amplitudes of higher order diagrams to be constructed directly from their vertex degrees. In particular, the star diagrams play the role of primary diagrams, while the amplitudes of other trees are the products of the corresponding primary amplitudes. The factorization also provides a simple graph-theoretic interpretation of the familiar Suyama-Yamaguchi equality at tree level and its higher order counterparts. The relations between amplitudes are algebraic consequences of the vertex structure, whereas the corresponding relations between complete momentum channels require the appropriate sums over external momentum assignments.

An important distinction is that the degree sequence determines the amplitude but does not necessarily determine the topology of a tree. Two non-isomorphic trees may have the same degree sequence and therefore the same amplitude, while their momentum dependent shapes are different. The number of non-isomorphic trees is governed by the well-known Otter counting problem and increases rapidly with the number of vertices. The graph representation therefore becomes particularly useful at higher orders, where a direct calculation of all momentum permutations becomes increasingly cumbersome. It also makes clear which properties of the correlation functions are universal and which depend on the underlying inflationary model.

To illustrate the applications of this framework, we considered a transient USR phase followed by a SR attractor. The sharpness of the transition from the USR phase to the final attractor regime is controlled by the parameter h<0h < 0. For a transition without a potential step (g=1g = 1), the nonlinear vertex amplitudes can be obtained analytically at arbitrary order as given in Eq. (65). This in turn gives the amplitude of every connected tree diagram at arbitrary order as given in Eq. (68). We have shown that the transition parameter affects different topologies differently, so that the relative strengths of the higher point correlation functions depend on the details of the USR to SR transition. An important conclusion was that all amplitudes are positive.

We then generalized the analysis by including a potential step, introducing the additional parameter gg which measures the change in the inflaton velocity across the transition. Both upward steps, 0<g<10 < g < 1, and downward steps, g>1g > 1, are allowed. An important result of our analysis is that, for upward steps, all nonlinear vertex amplitudes ada_d are strictly positive at arbitrary order. Consequently, all connected tree amplitudes are positive, and no cancellation of individual tree amplitudes can occur. For downward steps, the situation is qualitatively different. The competition between the USR evolution and the velocity change across the step allows the nonlinear vertex amplitudes to change sign or vanish. Their graph-theoretic interpretation is particularly simple: whenever ad=0a_d = 0, every tree containing a vertex of degree dd has vanishing amplitude, while trees without such vertices may survive. Thus, the cancellation of a single nonlinear response coefficient eliminates entire families of higher-point tree topologies. As explicit examples, we studied the cancellation curves a2=0a_2 = 0 and a3=0a_3 = 0 and examined their consequences for five- and six-point correlation functions. These results demonstrate that lower order non-Gaussianity can be suppressed without necessarily eliminating higher order correlations. More generally, the pattern of surviving and vanishing tree amplitudes provides information about the underlying inflationary transition that cannot be inferred from the bispectrum and trispectrum alone.

There are several directions in which the current study can be extended. A natural extension is to include loop diagrams and investigate how the resulting structures are related to corresponding theorems in graph theory. In particular, it would be interesting to examine which tree-level factorization relations and SY-type equalities survive after loop corrections are included. Another direction is to consider multifield models of inflation. In such scenarios, the graph structure becomes considerably richer because vertices and edges may carry different field labels, and the single field SY equality is replaced, in general, by the SY inequality. Finally, to perform the analytical calculations, here we considered an instantaneous and sharp transition from USR to SR phase. It would be desirable to extend the current analysis to the realistic setups with smooth transition and examine whether or not the current conclusions about the sign and the amplitude cancelation hold.

Acknowledgments: We thank Amjad Ashoorioon, Mohammad Hossein Namjoo and Farnaz Bahmani for useful discussions and correspondences. The author acknowledges the use of Maple and Mathematica software packages for analytical and numerical analysis. The work of A. N. is partially supported by the Iran National Science Foundation (INSF). The work of H. F. is supported by the INSF of Iran under the grant number 4045105.

Use of artificial intelligence: During the preparation of this manuscript, the authors used ChatGPT (GPT-6) to assist with the development of figures, mathematical analysis, manuscript organization and editing. The research questions and underlying theoretical framework were developed by the authors. All AI-assisted derivations, figures, and interpretations were independently verified and critically examined by the authors, who take full responsibility for this work.

Signs of Amplitudes

In this Appendix we study the signs of the vertex amplitudes ada_d. We show that for the upward step with 0<g<10 < g < 1 all ada_d are positive while for the downward step with g>1g > 1 there are corners of parameter space (h,g)(h,g) where ada_d can vanish.

Let us introduce the dimensionless variable x≡−3δϕ/πex \equiv-3\delta\phi/\pi_e. Equation (80) then takes the following form,

δN(x)=−13ln⁡(1−x)+2h(1−2xg2+x2g2−1).(84)\delta N(x)=-\frac{1}{3}\ln(1-x)+\frac{2}{h}\left(\sqrt{1-\frac{2x}{g^2}+\frac{x^2}{g^2}}-1\right). \tag*{(84)}

For convenience, let us define the following Taylor expansion,

1−2tz+z2=∑d=0∞cd(t)zd,t=1g,z=xg,(85)\sqrt{1-2tz+z^2}=\sum_{d=0}^{\infty}c_d(t)z^d,\qquad t=\frac{1}{g},\qquad z=\frac{x}{g}, \tag*{(85)}

where cd(t)c_d(t) is the coefficients of Taylor expansion. Using the generating function of Legendre polynomials [74], the coefficients for d≥2d\geq2 can be expressed as,

cd(t)=Pd(t)−2tPd−1(t)+Pd−2(t).(86)c_d(t)=P_d(t)-2tP_{d-1}(t)+P_{d-2}(t). \tag*{(86)}

Expanding δN\delta N and using N′=−πe−1(1−6/hg2)N'=-\pi_e^{-1}(1-6/hg^2), we obtain the vertex amplitude as follows,

ad=3d−1(d−1)!(1−6hg2)d[1+6dhgdcd(1g)],d≥2.(87)a_d=\frac{3^{d-1}(d-1)!}{\left(1-\frac{6}{hg^2}\right)^d}\left[1+\frac{6d}{hg^d}c_d\left(\frac{1}{g}\right)\right],\qquad d\geq2. \tag*{(87)}

This expression reproduces the explicit results in Eq. (79). In particular, for g=1g=1 it reduces to our previous result Eq. (65).

Upward Step (0<g<10 < g < 1)

The above analysis were general. We now establish the sign of cd(t)c_d(t) in the case of upward step, corresponding to 0<g<10 < g < 1 and t>1t > 1. Differentiating the generating function twice gives,

d2dz21−2tz+z2=(1−t2)(1−2tz+z2)−3/2.(88)\frac{d^2}{dz^2}\sqrt{1-2tz+z^2} = (1-t^2)(1-2tz+z^2)^{-3/2}. \tag*{(88)}

For t>1t > 1 the quadratic polynomial factorizes as,

1−2tz+z2=(1−λ+z)(1−λ−z),λ±=t±t2−1>0.(89)1-2tz+z^2=(1-\lambda_+z)(1-\lambda_-z), \qquad\lambda_\pm=t\pm\sqrt{t^2-1}>0. \tag*{(89)}

Both factors (1−λ±z)−3/2(1-\lambda_\pm z)^{-3/2} have strictly positive Taylor coefficients. Correspondingly, their product has strictly positive coefficients as well. Since 1−t2<01-t^2<0, this establishes that the Taylor coefficients of the left hand side of Eq. (A.2) for zmz^m, m≥2m\geq2 are all negative. Consequently, comparison of the right hand side of Eq. (A.2) yields,

cd(t)<0,(d≥2,t>1).(90)c_d(t)<0, \qquad(d\geq2,\quad t>1). \tag*{(90)}

In conclusion, for h<0h<0, the correction proportional to cd(1/g)c_d(1/g) in the square brackets of the general expression for ada_d in Eq. (A.4) is positive. Moreover, 1−6/(hg2)>01-6/(hg^2)>0 so we conclude the all order positivity relation,

ad>0,(d≥2,0<g≤1).(91)a_d>0, \qquad(d\geq2,\quad0<g\leq1). \tag*{(91)}

Note that the limiting case g=1g=1 follows directly from Eq. (65).

Using the graph-theoretic factorization relation Eq. (34), the above results indicate that for any connected tree TT, its amplitude is positive,

fT=∏v∈V(T)adv>0,(0<g≤1).(92)f_T=\prod_{v\in V(T)}a_{d_v}>0, \qquad(0<g\leq1). \tag*{(92)}

Downward Step (g>1g>1)

Now consider the downward step with g>1g>1. Following the same step as in previous section, the Taylor expansion Eq. (A.2) and the formula for ada_d as in Eq. (A.4) are still valid. However, since now t=1/g<1t=1/g<1, then the characteristics λ±\lambda_\pm become complex and the arguments following Eq. (A.6) do not proceed. As a result, cd(1/g)c_d(1/g) can take either signs which, depending on the model parameters, may yield to vanishing ada_d.

To handle this properly, note that the coefficients cdc_d can be expressed in terms of Gegenbauer polynomials. Indeed, using the second derivative given in Eq. (A.5) and the generating function [74],

(1−2tz+z2)−3/2=∑m=0∞Cm(3/2)(t)zm,(93)(1-2tz+z^2)^{-3/2}=\sum_{m=0}^{\infty}C_m^{(3/2)}(t)z^m, \tag*{(93)}

we obtain,

cd(t)=1−t2d(d−1)Cd−2(3/2)(t),d≥2,(94)c_d(t)=\frac{1-t^2}{d(d-1)}C_{d-2}^{(3/2)}(t), \qquad d\geq2, \tag*{(94)}

where Cd−2(α)(t)C_{d-2}^{(\alpha)}(t) are the ultraspherical polynomials (proportional to the Gegenbauer polynomials). Substituting this value of cd(1/g)c_d(1/g) into the general expression for ada_d in Eq. (A.4), we obtain the requirement for the cancellation of the vertex amplitude as a condition on the value of hh,

ad=0⟺h=−6(g2−1)(d−1)gd+2Cd−2(3/2)(1g),(d≥2).(95)a_d = 0 \quad\Longleftrightarrow\quad h = -\frac{6(g^2-1)}{(d-1)g^{d+2}} C_{d-2}^{(3/2)}\left(\frac{1}{g}\right), \qquad(d \ge2). \tag*{(95)}

Since h<0h < 0 and g>1g > 1, a physically admissible cancellation curve exists where,

Cd−2(3/2)(1g)>0.(96)C_{d-2}^{(3/2)}\left(\frac{1}{g}\right) > 0. \tag*{(96)}

This provides a general criterion for the existence of vertex zero curves at arbitrary order. Indeed, one can examine that the Gegenbauer coefficient Cd−2(3/2)(t)C_{d-2}^{(3/2)}(t) has a root only for t<1t < 1 so the cancellation of vertex amplitude can occur for g>1g > 1 while it is not possible for g<1g < 1.

References

  1. [1]H. Kodama and M. Sasaki, Cosmological Perturbation Theory, Prog. Theor. Phys. Suppl. 78 (1984) 1–166.
  2. [2]V. F. Mukhanov, H. A. Feldman and R. H. Brandenberger, Theory of cosmological perturbations. Part 1. Classical perturbations. Part 2. Quantum theory of perturbations. Part 3. Extensions, Phys. Rept. 215 (1992) 203–333.
  3. [3]S. Weinberg, Cosmology. 2008.
  4. [4]D. Baumann, Inflation, in Theoretical Advanced Study Institute in Elementary Particle Physics: Physics of the Large and the Small, pp. 523–686, 2011, 0907.5424, DOI.
  5. [5]S. Weinberg, Quantum contributions to cosmological correlations, Phys. Rev. D 72 (2005) 043514, [hep-th/0506236].DOI
  6. [6]J. M. Maldacena, Non-Gaussian features of primordial fluctuations in single field inflationary models, JHEP 05 (2003) 013, [astro-ph/0210603].DOI
  7. [7]P. Adshead, R. Easther and E. A. Lim, The ’in-in’ Formalism and Cosmological Perturbations, Phys. Rev. D 80 (2009) 083521, [0904.4207].DOI
  8. [8]X. Chen and Y. Wang, Quasi-Single Field Inflation and Non-Gaussianities, JCAP 04 (2010) 027, [0911.3380].DOI
  9. [9]M. Sasaki and E. D. Stewart, A General analytic formula for the spectral index of the density perturbations produced during inflation, Prog. Theor. Phys. 95 (1996) 71–78, [astro-ph/9507001].
  10. [10]M. Sasaki and T. Tanaka, Superhorizon scale dynamics of multiscalar inflation, Prog. Theor. Phys. 99 (1998) 763–782, [gr-qc/9801017].
  11. [11]D. Wands, K. A. Malik, D. H. Lyth and A. R. Liddle, A New approach to the evolution of cosmological perturbations on large scales, Phys. Rev. D 62 (2000) 043527, [astro-ph/0003278].DOI
  12. [12]D. H. Lyth, K. A. Malik and M. Sasaki, A General proof of the conservation of the curvature perturbation, JCAP 05 (2005) 004, [astro-ph/0411220].DOI
  13. [13]D. H. Lyth and Y. Rodriguez, The Inflationary prediction for primordial non-Gaussianity, Phys. Rev. Lett. 95 (2005) 121302, [astro-ph/0504045].DOI
  14. [14]A. A. Abolhasani, H. Firouzjahi, A. Naruko and M. Sasaki, Delta N Formalism in Cosmological Perturbation Theory. WSP, 2, 2019, 10.1142/10953.
  15. [15]S. Hooshangi, M. H. Namjoo and M. Noorbala, Rare events are nonperturbative: Primordial black holes from heavy-tailed distributions, Phys. Lett. B 834 (2022) 137400, [2112.04520].DOI
  16. [16]Y.-F. Cai, X.-H. Ma, M. Sasaki, D.-G. Wang and Z. Zhou, One small step for an inflaton, one giant leap for inflation: A novel non-Gaussian tail and primordial black holes, Phys. Lett. B 834 (2022) 137461, [2112.13836].DOI
  17. [17]Y.-F. Cai, X.-H. Ma, M. Sasaki, D.-G. Wang and Z. Zhou, Highly non-Gaussian tails and primordial black holes from single-field inflation, JCAP 12 (2022) 034, [2207.11910].DOI
  18. [18]R. Kawaguchi, T. Fujita and M. Sasaki, Highly asymmetric probability distribution from a finite-width upward step during inflation, JCAP 11 (2023) 021, [2305.18140].DOI
  19. [19]D. H. Lyth, Generating the curvature perturbation at the end of inflation, JCAP 11 (2005) 006, [astro-ph/0510443].DOI
  20. [20]M. Sasaki, Multi-brid inflation and non-Gaussianity, Prog. Theor. Phys. 120 (2008) 159–174, [0805.0974].DOI
  21. [21]A. Naruko and M. Sasaki, Large non-Gaussianity from multi-brid inflation, Prog. Theor. Phys. 121 (2009) 193–210, [0807.0180].
  22. [22]D. Wands, Multiple field inflation, Lect. Notes Phys. 738 (2008) 275–304, [astro-ph/0702187].
  23. [23]D. Wands, Local non-Gaussianity from inflation, Class. Quant. Grav. 27 (2010) 124002, [1004.0818].DOI
  24. [24]D. Cruces, S. Pi and M. Sasaki, δn formalism: A new formulation for the probability density of the curvature perturbation, Phys. Rev. D 114 (2026) 023549, [2505.24590].DOI
  25. [25]S. Pi and M. Sasaki, Logarithmic Duality of the Curvature Perturbation, Phys. Rev. Lett. 131 (2023) 011002, [2211.13932].DOI
  26. [26]Y.-F. Cai, X. Chen, M. H. Namjoo, M. Sasaki, D.-G. Wang and Z. Wang, Revisiting non-Gaussianity from non-attractor inflation models, JCAP 05 (2018) 012, [1712.09998].DOI
  27. [27]X. Chen, H. Firouzjahi, E. Komatsu, M. H. Namjoo and M. Sasaki, In-in and δN calculations of the bispectrum from non-attractor single-field inflation, JCAP 12 (2013) 039, [1308.5341].DOI
  28. [28]M. H. Namjoo, H. Firouzjahi and M. Sasaki, Violation of non-Gaussianity consistency relation in a single field inflationary model, EPL 101 (2013) 39001, [1210.3692].
  29. [29]A. A. Abolhasani, H. Firouzjahi, S. Khosravi and M. Sasaki, Local Features with Large Spiky non-Gaussianities during Inflation, JCAP 11 (2012) 012, [1204.3722].DOI
  30. [30]A. A. Abolhasani and M. Sasaki, Single-field consistency relation and δN-formalism, JCAP 08 (2018) 025, [1805.11298].
  31. [31]A. Talebian-Ashkezari, N. Ahmadi and A. A. Abolhasani, δ M formalism: a new approach to cosmological perturbation theory in anisotropic inflation, JCAP 03 (2018) 001, [1609.05893].
  32. [32]A. A. Abolhasani, R. Emami, J. T. Firouzjaee and H. Firouzjahi, δN formalism in anisotropic inflation and large anisotropic bispectrum and trispectrum, JCAP 08 (2013) 016, [1302.6986].DOI
  33. [33]T. Tanaka, T. Suyama and S. Yokoyama, Use of delta N formalism - Difficulties in generating large local-type non-Gaussianity during inflation -, Class. Quant. Grav. 27 (2010) 124003, [1003.5057].DOI
  34. [34]T. Tanaka and Y. Urakawa, δN formalism, JCAP 07 (2025) 045, [2408.13673].DOI
  35. [35]T. Tanaka and Y. Urakawa, Statistical Anisotropy of Primordial Gravitational Waves from Generalized δN Formalism, Phys. Rev. Lett. 132 (2024) 231003, [2309.08497].DOI
  36. [36]T. Tanaka and Y. Urakawa, Anisotropic separate universe and Weinberg’s adiabatic mode, JCAP 07 (2021) 051, [2101.05707].
  37. [37]D. Cruces, C. Germani and A. Palomares, An update on adiabatic modes in cosmology and δN formalism, JCAP 06 (2023) 002, [2212.05112].DOI
  38. [38]H. Firouzjahi and A. Riotto, Primordial Black Holes and loops in single-field inflation, JCAP 02 (2024) 021, [2304.07801].arxiv.org/abs/2304.07801
  39. [39]T. Fujita, M. Kawasaki, Y. Tada and T. Takesako, A new algorithm for calculating the curvature perturbations in stochastic inflation, JCAP 12 (2013) 036, [1308.4754].DOI
  40. [40]T. Fujita, M. Kawasaki and Y. Tada, Non-perturbative approach for curvature perturbations in stochastic δN formalism, JCAP 10 (2014) 030, [1405.2187].DOI
  41. [41]V. Vennin and A. A. Starobinsky, Correlation Functions in Stochastic Inflation, Eur. Phys. J. C 75 (2015) 413, [1506.04732].DOI
  42. [42]V. Vennin, H. Assadullahi, H. Firouzjahi, M. Noorbala and D. Wands, Critical Number of Fields in Stochastic Inflation, Phys. Rev. Lett. 118 (2017) 031301, [1604.06017].DOI
  43. [43]H. Assadullahi, H. Firouzjahi, M. Noorbala, V. Vennin and D. Wands, Multiple Fields in Stochastic Inflation, JCAP 06 (2016) 043, [1604.04502].DOI
  44. [44]J. Grain and V. Vennin, Stochastic inflation in phase space: is slow roll a stochastic attractor?, JCAP 05 (2017) 045, [1703.00447].DOI
  45. [45]M. Noorbala, V. Vennin, H. Assadullahi, H. Firouzjahi and D. Wands, Tunneling in Stochastic Inflation, JCAP 09 (2018) 032, [1806.09634].DOI
  46. [46]H. Firouzjahi, A. Nassiri-Rad and M. Noorbala, Stochastic Ultra Slow Roll Inflation, JCAP 01 (2019) 040, [1811.02175].DOI
  47. [47]H. Firouzjahi, A. Nassiri-Rad and M. Noorbala, Stochastic nonattractor inflation, Phys. Rev. D 102 (2020) 123504, [2009.04680].DOI
  48. [48]N. Ahmadi, M. Noorbala, N. Feyzabadi, F. Eghbalpoor and Z. Ahmadi, Quantum diffusion in sharp transition to non-slow-roll phase, JCAP 08 (2022) 078, [2207.10578].DOI
  49. [49]M. Noorbala, Classicality of stochastic noise away from quasi-de Sitter inflation, JCAP 10 (2024) 053, [2408.11640].DOI
  50. [50]Z. Ahmadi and M. Noorbala, Deviations from Gaussian white noise in stochastic inflation, JCAP 07 (2026) 054, [2512.17070].DOI
  51. [51]D. Cruces, C. Germani and T. Prokopec, Failure of the stochastic approach to inflation beyond slow-roll, JCAP 03 (2019) 048, [1807.09057].
  52. [52]K. Miyamoto and Y. Tada, Calculating the squeezed bispectrum in stochastic inflation by Monte Carlo simulation, 2609.02076.arxiv.org/abs/2609.02076
  53. [53]T. Murata and Y. Tada, STOchastic LAttice Simulation of hybrid inflation, JCAP 09 (2026) 061, [2603.04850].DOI
  54. [54]K. Miyamoto and Y. Tada, Calculating the power spectrum in stochastic inflation by Monte Carlo simulation and least squares curve fitting, JCAP 03 (2026) 060, [2508.17654].DOI
  55. [55]T. Murata and Y. Tada, Stochastic tail of the curvature perturbation in hybrid inflation, Phys. Rev. D 113 (2026) 023552, [2507.22439].DOI
  56. [56]C. T. Byrnes, M. Sasaki and D. Wands, The primordial trispectrum from inflation, Phys. Rev. D 74 (2006) 123519, [astro-ph/0611075].DOI
  57. [57]T. Suyama and M. Yamaguchi, Non-Gaussianity in the modulated reheating scenario, Phys. Rev. D 77 (2008) 023505, [0709.2545].DOI
  58. [58]T. Suyama, T. Takahashi, M. Yamaguchi and S. Yokoyama, On Classification of Models of Large Local-Type Non-Gaussianity, JCAP 12 (2010) 030, [1009.1979].DOI
  59. [59]G. Tasinato, C. T. Byrnes, S. Nurmi and D. Wands, Loop corrections and a new test of inflation, Phys. Rev. D 87 (2013) 043512, [1207.1772].DOI
  60. [60]C. T. Byrnes, K. Koyama, M. Sasaki and D. Wands, Diagrammatic approach to non-Gaussianity from inflation, JCAP 11 (2007) 027, [0705.4096].
  61. [61]S. Yokoyama, T. Suyama and T. Tanaka, Efficient diagrammatic computation method for higher order correlation functions of local type primordial curvature perturbations, JCAP 02 (2009) 012, [0810.3053].DOI
  62. [62]T. Suyama and S. Yokoyama, Extension of local-type inequality for the higher order correlation functions, JCAP 07 (2011) 033, [1105.5851].DOI
  63. [63]J. A. Bondy and U. Murty, Graph Theory, Springer 2008.
  64. [64]R. Otter, Richard Otter, The number of trees, The Annals of Mathematics, 2nd Ser., Vol. 49, No. 3. (Jul., 1948), pp. 583-599, .DOI
  65. [65]G. H. Hardy and S. Ramanujan, Asymptotic formulae in combinatory analysis, Proceedings of the London Mathematical Society s2-17 (1918) 75–115.
  66. [66]W. H. Kinney, Horizon crossing and inflation with large eta, Phys. Rev. D 72 (2005) 023515, [gr-qc/0503017].DOI
  67. [67]J. Martin, H. Motohashi and T. Suyama, Ultra Slow-Roll Inflation and the non-Gaussianity Consistency Relation, Phys. Rev. D 87 (2013) 023514, [1211.0083].DOI
  68. [68]M. J. P. Morse and W. H. Kinney, Large-η constant-roll inflation is never an attractor, Phys. Rev. D 97 (2018) 123519, [1804.01927].DOI
  69. [69]K. Dimopoulos, Ultra slow-roll inflation demystified, Phys. Lett. B 775 (2017) 262–265, [1707.05644].DOI
  70. [70]X. Chen, H. Firouzjahi, M. H. Namjoo and M. Sasaki, A Single Field Inflation Model with Large Local Non-Gaussianity, EPL 102 (2013) 59001, [1301.5699].DOI
  71. [71]H. Firouzjahi and A. Riotto, Sign of non-Gaussianity and the primordial black holes abundance, Phys. Rev. D 108 (2023) 123504, [2309.10536].
  72. [72]H. Firouzjahi and A. Nassiri-Rad, Trispectrum in an extended ultraslow-roll model with transition to slow roll, Phys. Rev. D 112 (2025) 123541, [2509.09608].DOI
  73. [73]M. H. Namjoo and B. Nikbakht, Non-Gaussianity consistency relations and their consequences for the peaks, JCAP 08 (2024) 005, [2401.12958].arxiv.org/abs/2401.12958
  74. [74]G. B. Arfken and J. W. Weber, Mathematical Methods For Physicists, Elsevier 2005.

Paper details

vlnc:W_ttx5yxxn
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

ChatGPT (GPT-6)

View statement
Looking for related papers…
Contents